Skip to content
WhittleOSWhittleOS
← Home

Changelog

What shipped

Real history, newest first. No roadmap promises — only things you can use today.

2026-09-08

Three keyword-roadmap pages, after the attractive numbers were removed

  • The roadmap suggested six posts. It became three site pages and four native cuts: one guide ranks e-commerce ideas from documented merchant pain, one explains why intent and page purpose beat a headline search-volume number, and one ranks low-cost developer ideas by how quickly they can produce evidence.
  • The e-commerce page starts with the awkward corpus result. A broad text match returned 24 strong records; reading every one left 14 genuinely about merchant operations, drawn from eight pages. None had been seen again. The page publishes that limit before it publishes the ideas, then kills half the editorial shortlist before asking you to test the rest.
  • The search-intent page does not repeat the private planning workbook's vendor estimates. It publishes the decisions those estimates forced instead: drop the wrong searcher, isolate an outlier before using an average, and keep a small commercial query only when the receiving page has a measurable conversion job.
  • The developer page uses the eight existing public idea cards rather than inventing another list. The test windows are editorial budgets, labelled as such; the verdicts and evidence gaps are imported from the cards that own them. The startup-ideas guide now makes its missing second step explicit too: noticing problems gives you material, while elimination gives you a decision.
  • Four external cuts are written and linked to their receiving pages. They remain explicitly awaiting a founder read; no social client, token or scheduler was added, and nothing has been posted from an account by this release.

2026-09-08

Tell us a rejection is wrong — and where we went wrong

  • A run drops most of what it finds and explains each one, and until today none of that ever came back the other way. If the checks were killing good ideas because our search read the wrong two dozen pages rather than because the ideas were weak, there was no way to find out: a run that is blind looks exactly as confident as a run that is right.
  • Every dropped idea on a run now carries a short form. You pick what should have happened, and — the part that matters — where we went wrong: searched the wrong corners, missed the page, read the page and missed the problem, had the evidence and dropped it anyway, or ranked it wrong. If you have the page we missed, there is a box for the link.
  • The tag is the whole point. “You are wrong” is a number that implies nothing. Nine people saying we missed the page means our search is too narrow; the same nine saying we had the evidence and dropped it anyway means a check is too strict. Those need opposite fixes and they look identical without the label.
  • It is on the published example run as well as your own, and it needs no account. The people most likely to know that a real-estate idea was dropped in error are readers of the example, not the person who paid for a run.
  • One thing it deliberately does not do: filing one never changes the run. Not the verdict, not the score, not the shortlist. A tool that moved its numbers when a stranger disagreed would be the flattery machine rebuilt from the other end, which is the thing this product exists not to be. Challenges go into a count we read when deciding what to fix, and that is all.

2026-09-07

Every guide now shows its next reading and its actual next step

  • All 26 guides already carried hand-picked related paths and a CTA key, but the shared page ignored both. The metadata looked complete while readers saw neither. The shared guide template now renders the selected links and exactly one tracked closing action on every article.
  • The article checker now refuses to call related paths valid unless the shared template still renders them. The customer-copy sweep also reads MDX guide prose, including accessible labels and nested components, while excluding imports and interpolation identifiers.
  • The remediation corrected the validator roundup's source claim to the measured linked share, removed unsupported category superlatives, replaced retired internal wording, exposed the official Y Combinator source, aligned the agency and coaching title promises with their evidence, and added concrete examples and figures where the earlier review found abstract or unillustrated explanations.

2026-09-07

Business ideas for coaches — the best-documented problem is the one software cannot touch

  • Of the 26 strongest problems we hold for coaches and consultants, one theme label repeats more than any other and is the only one appearing three times: renewals. That is an awkward thing to publish on a page about business ideas, because keeping clients is simultaneously the best-documented problem in this vertical and the one a product can do least about. In a coaching business, retention is the service.
  • So the page splits the problem instead of selling it. Four parts of keeping a client: the relationship and the awkward conversation, which software can take nothing from, and the remembering and the visibility, which it can take almost entirely. A lapsed renewal nobody chased is a pure mechanical loss with a date on it, and that is the half worth building. Two of four is the honest assessment of the vertical, and the two rows reading “nothing” are the ones a hopeful product description quietly claims anyway.
  • As on the agency page, the word list that decides what counts as a retention problem is printed next to the number, so anyone who disagrees can re-run it against the topic page instead of trusting ours.
  • The limits again, on the page rather than under it: every record was found by exactly one sweep, so “most-repeated” means three times in what we documented and nothing about how often coaches raise it; 26 records come from 16 pages; and three of the 26 are not coaches at all — one is a finance team wanting to reduce spend on consultants, which the bare word “consult” drags in.
  • How that last number was reached is worth recording, because the first attempt was wrong. A regex over the buyer field said six rows were off topic, and three of those were practitioners — course creators, photographers, freelance creators — caught by the word “creator”. The count only settled once all 26 buyer fields were read. That is twice in one day that an automated count agreed with a plausible conclusion: on the neighbouring topic the same class of metric returned zero of 25 and was a pure artifact of the search terms describing problems rather than people.

2026-09-07

Business ideas for agencies — and the three limits on the evidence behind it

  • We hold 25 strong records for agency problems. Filtering them by their own theme labels, 52% are about getting paid — invoicing, collections, revenue visibility, profitability, quote-to-cash — and almost none are about the client work itself. That inverts where an agency owner looks for a product: the craft is what clients pay for precisely because it is bespoke, and bespoke is the definition of a hire rather than a product.
  • The word list that decides “about money” is printed on the page next to the number. A different list gives a different count, so anyone who disagrees with mine can re-run it against the topic page rather than taking the figure on trust. The grouping is reproducible instead of editorial, which is not the same as being right.
  • Three limits are on the page rather than in a footnote, and each of them makes the page weaker. Every record in our corpus has been found by exactly one sweep, so nothing here supports a claim about what agencies complain about most — only about what we have documented. The 25 records come from 19 pages, because reading one page often yields two or three statements, so they are not 25 independent findings. And the topic match is a substring: reading all 25 by hand, two are independent coaches and one is a staffing-agency recruiter, which leaves 22 genuinely on topic. The page uses 22.
  • One observation we cannot support and say so: nothing about creative review or scope disputes appears in the documented set, and those are the loudest complaints in any agency. Our read is not that they are rare but that they are hard to write down publicly where a search can find them. That is a hypothesis, labelled as one.
  • This page exists because two others do not. The plan called for six vertical guides; counting the corpus first killed two of them outright — one topic has 36 records and not a single strong one, and the topic the plan called our thinnest turned out to be fourth from the bottom. Printing the rows before writing is the cheapest step in the process and it removed a third of the work.

2026-09-07

We told you a problem “comes up a lot” when we had seen it once

  • Every problem in the pain library carried a score and a phrase: “comes up sometimes” or “comes up a lot”. Both are claims about frequency. Measured on the corpus this morning: 1,433 of the 1,434 problems in it have been found by exactly one search sweep. Precisely one has been found twice. So there was no frequency to report, and 196 problems were telling you they come up a lot on the strength of a single sighting.
  • The score itself is severity multiplied by how many sweeps found the problem again. With that second number stuck at one everywhere, the score is severity and nothing else — which would have been fine to say, and is what it now says.
  • The worse half was the disclosure. Each row can show a “seen in N sweeps” chip, and it was set to appear only when N was greater than one — so on 1,433 rows out of 1,434 the frequency phrase stood with nothing beside it, and the one fact that would have corrected it was hidden in exactly the case that needed it. Every row now says how many sightings it has, including “seen once”.
  • The phrases are gone. A problem is now “described as mild” or “described as painful”, which is what severity actually measures. Only the top band still mentions recurrence, and it has earned it: the arithmetic puts it out of reach below three sightings. That entitlement is not written down anywhere as a claim — a test re-derives it from the scoring code, so changing the formula or the bands fails the build instead of quietly making a label false.
  • The explainer on the topic pages and on Trending now counts, from the rows in front of you, how many have been found more than once, and says so. On most topics today the honest sentence is that none of them have. It will change on its own when a sweep re-surfaces something rather than needing an edit.
  • This came out of research for a different piece of work — measuring whether six market topics had enough documented problems behind them to write about. Two of the six turned out not to, and the counting that established it is what surfaced this.

2026-09-07

Put our misses on the page a buyer lands on

  • The page for people searching “startup idea validator” listed four ways our check differs from the category and then asked for a signup. Everything on it was about us. Nothing on it was checkable in the thirty seconds someone spends before deciding whether to trust a tool — and measured 09-07, Google had fetched that URL on the 5th and declined to index it.
  • It now carries the two things the rival comparison pages on that query cannot produce. First, what our gate did against eight companies whose endings are already documented, fed in as anonymized one-liners describing them at founding: the solo-founder lens said stop on 4 of the 5 that failed, and on 1 of the 3 that succeeded. That last number is the one a vendor deletes. It is on the page, in the table, next to the others.
  • Underneath it, all six of our own caveats — not in a footer. N is eight, hand-picked because the endings are known, with hindsight in them, so no accuracy percentage follows and we do not quote one. The page’s FAQ already said there is no honest accuracy number; a table of hits without those caveats beside it would have put the page in contradiction with itself two screens later. There is now a test that fails if the caveats are moved away from the counts.
  • Second, what a tool built to encourage you says about the same idea. On 2026-06-10 we ran our own one-liner through three tools in one day and kept the wording: one called it promising at 72 out of 100 — “excellent potential, among the top ideas we’ve seen” — while our own two modes returned 15 and 0. One idea, one day, one run each; the page says exactly that and claims no rate. The date is there so it can be re-run.
  • One tidy-up came with it. Three surfaces quote the benchmark counts, and the counting lived in two places. It now lives in the module that owns the rows, and both surfaces read it — verified identical on all eight figures before and after. A summary that can drift from its own table is not a benchmark.

2026-09-07

The micro-SaaS page argued about two numbers and never ran them on anything

  • It made a decent case — a one-person business is decided by how many customers your price requires and how many hours those customers cost you — and then never applied it to a named idea. Meanwhile eight real finalist cards from one of our own sweeps sat two directories away, unused. The page now runs the arithmetic on one of them.
  • The idea it runs on is the cheap-positioning one: a no-frills invoicing tool for freelancers on a tight budget. Our own engine did not clear it — HOLD at 51. Sweeping a price ladder against it at $3,000 a month shows why without needing an opinion: at $9 support alone is 19.2 hours a week, at $19 it is 9.1, and both are past the point where support is the job. The bottom two rungs are ruled out before a line of code exists. The first genuinely comfortable price is $99, where the whole business is 31 customers.
  • That is the difficulty with cheap-and-no-frills stated as arithmetic instead of advice: the thing that makes it attractive to the buyer is the thing that makes it unrunnable by one person. The card's own recorded reason says the same in words — competing on price compresses margins — which is two parts of the system agreeing, and worth exactly that much, because both parts are ours.
  • The second half is the whole library, sorted the way anyone would sort it: best number first. The verdict column does not descend with the score. Two ideas share a score of 60 and one is a go-and-test while the other is a kill. That is not a rounding artefact — a score says how promising an idea looks across seven areas, and the verdict is decided by checks that can fail outright. A tool that only shows you the number has handed you the half that cannot say no.
  • Also published, because it is the part pages like this leave out: every card that did not clear, with the engine's own leading reason. The two held ideas failed on the same thing — nobody had paid yet, and a price test is cheap. The two killed ones failed on the wedge, which is not a problem more evidence fixes.
  • One thing fixed underneath. The support bands — background noise, real tax, support-is-the-job — were three hand-typed copies of the same two numbers: in the calculator, on the calculator's explainer page, and in this guide's prose. They are now one exported constant that the calculator itself switches on, so a page cannot quote a threshold the tool does not use. The rendered pages are unchanged; the numbers just stopped being typed.

2026-09-07

Five pages, one per analysis — including the three that can refuse to answer

  • One guide for each of the five focused analyses this product runs, written so the page and the code cannot disagree: every figure, verdict ladder and field list is computed by importing the module that owns it. Where a page claims a guarantee, the claim was produced by breaking the output on purpose and reporting what the code actually refused.
  • How to price a micro SaaS. Three verdicts, and none of them is “too high” — that is not how a one-person product fails. The interesting part is a coincidence we did not arrange: the pricing floor written into the prompt, about $19 a month, falls on exactly the two rows where support hours alone exceed a solo founder's week. Both halves of that agreement are ours, so the page calls it internal consistency rather than proof, and publishes it anyway because neither half was tuned to the other.
  • Will your SaaS be support-heavy. The one analysis whose scale runs backwards — low is the good verdict — so the page opens on the polarity rather than burying it. At 100 customers, half an hour of support each per month is 11.5 hours a week: a part-time job hidden inside a number that sounds like nothing. It is also the second most common reason our gate kills a candidate, named in 187 of 706 drops, behind only a missing documented problem.
  • My startup idea is too broad. Everyone reaches that diagnosis alone and then gets told to niche down. The answerable questions are what a narrowing must contain — four fields, mandatory, established by blanking each one and watching the guard reject it — and whether a narrowing exists at all, which depends on you as much as the idea: two of the three things that trigger an honest refusal are read off your own profile. Three more fields are asked for and not enforced, and the page says which, because a guarantee is only as good as its stated edge.
  • What to build first. This one checks whether you have earned the question. Walking every state the check can be in gives seven, and four of them get a refusal instead of a scope — including a killed idea, an inconclusive one, and the state every idea is in before anything has evaluated it. The refusal is empty on purpose: a scope with a caveat attached is still a scope, and you will build from it.
  • Am I the right founder for this idea. Seven checks and not one is about credentials. It is the only place in the product where a fact about you can kill the idea outright, so the page publishes what each rung of the fit ladder actually does downstream, and one thing we had not written down anywhere: a plain recommendation to drop the idea lands on the same decision at every rung, including the top one. The score is not the decision.

2026-09-07

We were telling you an idea had no problem behind it. We only knew we had not found one.

  • The most common reason this product gives for dropping an idea was labelled “No real problem behind it”. That is a claim about the world, and we do not have it. What the check actually does is look for the idea among the documented problems that particular run collected — on the published example, 60 searches and 24 pages read — and drop it when there is no match. If the run found nothing at all, everything is dropped on that reason.
  • So the honest sentence is “we found no documented problem”, and that is what it now says, everywhere it appears: the label on each dropped idea, the tally at the top of a run, the front page, the pricing and market pages, and the FAQ.
  • This mattered more than a wording nit because of how often it fires: 21 of the 36 drops on the published run, and around 440 of the roughly 700 across every real run we have done. It was the single most-repeated sentence the product says, and it was the one place we asserted more than we had found — in a tool whose entire argument is that it does not do that.
  • Two smaller things came out of the same thread. Our own internal style rule had been recommending the wrong phrasing as the correct plain-English replacement, which is how it reached four pages: each author did exactly what the guide said. And there is now a build check that fails if the over-claiming version comes back.
  • What has not changed, and is the harder half: the run still cannot tell you the difference between “we looked and found nothing” and “we found evidence there is no demand”. Those should lead you to different next steps. Today only the first one is real, and the label finally says so. Raised by a reader of the published run, in better words than we had.

2026-09-07

AI startup ideas, and the objection that turns out not to be the killer

  • Every AI idea meets the same objection: a vendor will ship this as a feature and there will be nothing left. It is a fair question and, on the one Discovery run we publish in full, it is not what did the killing. The check that asks it was named on 6 of the 36 ideas that run dropped. The check asking whether there was a documented problem behind the idea at all was named on 21.
  • The detail underneath is the part worth having. Of those 6, exactly one died of the wrapper question by itself; the other five failed something else in the same pass — support load, a compliance barrier, or no documented problem. So the wrapper is usually the visible symptom of something that already has its own name, and the cheap checks are the ones to run first.
  • Both figures are computed from the run rather than written down, and the page fails to build if a regenerated fixture ever inverts the comparison it argues. A page whose whole point is a gap between two counts should not be able to keep rendering confidently once that gap closes.

2026-09-06

Three more, and the one where we put our own idea on the slab

  • Where to find real customer problems argues that the stock answer names a destination when the question is about a property: can you open the thing your claim rests on? It publishes the five tiers a source is graded in, the three ways a source can be reached at all, and the two facts that cut against us — sites that say no in robots are skipped, and the most candid places are usually the least reachable. On the run we publish, 32 of 56 problems carry a link; the other 24 are marked unverified rather than quietly promoted.
  • Why AI says your startup idea is great is the mechanism, not a grudge. Helpfulness and judgement are different objectives and one of them was trained for; a false yes costs the tool nothing today and costs you months. To make it concrete without asking anyone to trust our opinion of someone else, we used our own idea: one day, three tools, verdicts quoted verbatim. One called it promising at 72 out of 100. Our two modes returned 15 and 0. That is one run on one day and the page says so — it supports one sentence and no rate.
  • Is my startup idea already taken answers yes, and then explains why that cannot be acted on. The question that can: are the incumbent's own users complaining about one specific thing, and is it something they will not fix rather than have not fixed yet. Three possible decisions come out, and one of them is walk away — usually triggered by silence rather than by a strong competitor.

2026-09-06

Three pages that put our own numbers where they can be argued with

  • Platform risk, counted instead of warned about. Everyone writes essays about building on someone else's platform; nobody says how often it is actually fatal. Across the 706 candidates our gate has killed, platform dependency was named in about a fifth of them — third place, behind missing evidence of a problem by a factor of three. It is also the one check we deliberately made survivable, and the page carries both halves, because the count alone reads as fear and the softening alone reads as an excuse.
  • Do startup idea validators work. We ran eight companies with known endings through our own gate under two founder profiles, published every row including the ones that went against us, and still refuse to quote an accuracy figure — eight hand-picked cases with hindsight in them cannot support one. The page ends with the test to apply to any tool in this category, ours included: make it show you the rows where it was wrong.
  • Famous startup failures. Not the 2019 statistic everyone pastes. Five companies whose stories are finished, the one-liner each could have been described by on day one, and the property three of them shared: the money had to go out before the answer could come in. The fifth is the one our own gate did not kill, and it stays on the page.
  • One correction shipped with them. An internal note recorded those kill counts as coming from 91 production runs; re-measured, the counts reproduce exactly but 68 of those runs are stub-mode synthetics that kill nothing, so the real denominator is 23. The interesting numbers were right and the one nobody checked was wrong.

2026-09-06

Three guides that answer questions the category answers badly

  • How to score a business idea publishes our actual weights — seven categories, what each is worth, and the arithmetic. Search that question and you get a page of vendors explaining how to build a weighted model; not one of them says what they picked. Unpublished weights cannot be argued with, which is the point of not publishing them. The useful part falls straight out: the heaviest area carries four times the leverage of the lightest, so the same three rating points are worth eight points and a grade change spent on distribution, and two points spent on cash flow.
  • The validation checklist publishes both lists, because there are two and they are not the same list — ten checks for a candidate we sourced, twelve for an idea you bring. It argues for cost order over interest order: the cheapest question on the list is also the one that killed the most candidates on the run we publish, twenty-one of thirty-six.
  • How many customers do I need answers the arithmetic nine results explain around, and then asks the question none of them asks: at that customer count, how many hours a week is support? At five thousand a month and fifteen minutes per customer, the floor sits at forty-nine dollars — below it, support alone is more than a solo founder's whole week, and the customer count was never the binding constraint.
  • Every figure on all three is computed by the code that does the real work — the scoring module, the triage checklist, the two free calculators — so a published claim and an applied rule cannot come apart.

2026-09-06

The run now names the one assumption most likely to sink your best idea

  • A total out of 100 is an average of seven ratings, and an average hides the thing that actually kills an idea. The published example run tops out at 52 — four areas sitting at 6 or 7, and distribution at 3 out of 10, with the reason given as a crowded market and no evidenced way to reach solo agents repeatedly. Read as a single number, that idea is “middling”. Read properly, it has one specific problem and six that are fine.
  • So the recommended first move now names that area, its rating, and how much of the total it carried — next to the idea it belongs to. Where two areas rate equally badly, the one worth more of the score is named, because a 3 in an area worth a fifth of the total is not the same danger as a 3 in an area worth a twentieth.
  • The point is what it does to the test you run next. A generic demand test can pass — fifty signups off one forum post — while the weak area stays exactly as unproven as it was. Naming it turns a low score from an opinion you can argue with into something specific to go and disprove, against the deadline the plan already sets in advance.
  • It costs nothing and it is not a new judgement: the per-area ratings were already computed, already paid for, and already shown further down the page. This reads them and puts the worst one where the decision is made. Suggested by someone who read the published run and said the score should drive the test rather than just be a harsher number — they were right.

2026-09-06

The guides read like articles now, not like documentation

  • A guide used to open with a title, one line of summary, a byline and a bordered box of anchors — and that box was the largest thing on the screen. It now opens on a masthead band: the summary is set at reading size because it is the sentence that decides whether you start, and underneath it a strip states what the article establishes in its own terms rather than repeating a byline. On this page that reads Method, Gradable metrics, What a pass proves.
  • The contents moved into the left margin on a wide screen, numbered, and they stay with you as you scroll. Each section carries the same number, so the two never disagree — the first build of this had the list skip 03 while the article still showed it, which is worse than not numbering at all. Below 1100px the list falls back above the article, where a narrow screen wants it anyway.
  • Reading time is counted from the article's own prose rather than typed into its metadata, so it cannot quietly go stale the first time a page is edited.
  • Two things this turned up that were not cosmetic. The roundup guide had a paragraph nested inside a paragraph, which is invalid HTML and was throwing a hydration error in every reader's browser since August. And the palette did not change at all — same tokens, same colours; every part of this was placement.

2026-09-06

The guides were set in the colour we reserve for captions

  • Every paragraph of every guide rendered in the muted text colour — the one this design system uses for figure captions, bylines and metadata. A whole article set one step down in contrast reads as small print no matter what it says, and that is most of why the guides looked like documentation rather than something worth starting. Body text is now full contrast, at 17px over a slightly narrower column, and the change lands on all nine at once.
  • The header was the other half. It gave a reader a title, one grey line and a boxed list of anchors — the largest object above the fold was navigation. The summary is now a real opening paragraph at reading size, the contents sit on a quiet rail instead of in a filled card, and there is a derived reading estimate up top so you can decide before you commit. Each section now opens under a hairline rule, which is what turns a wall of prose into an article you can skim.
  • Nothing new entered the palette and no page's words changed. This was placement and contrast, and both were ours.

2026-09-06

A guide for the question the whole category answers the same way

  • Search for how to validate an idea without talking to customers and every result tells you to talk to customers — two of them argue the premise is wrong. There is now a page at /guides/validate-startup-idea-without-talking-to-customers that answers it properly, and the answer is not that you skip the customer. It is that you replace what you collect from them: a conversation gives you an opinion you grade afterwards, a number checked against a bar you wrote down first gives you something you cannot re-grade.
  • The useful part is what it concedes. A landing-page test that goes well is routinely described as proof the idea works. Measured against our own scorecard, a clean pass covers 55% of what decides an idea and leaves support load, platform exposure and cash flow completely untouched — and traffic on its own evidences nothing at all, because a channel is proven when it delivers a person, not attention.
  • Both charts call the grader rather than quoting it. The percentages, the categories and the worked example's verdicts are produced by the same function the product runs when it decides what a reported outcome is allowed to affect, so the page cannot drift from the thing it describes. Reading the rendered page caught the one number that was not: a sentence saying two of three met, next to a chart the grader had scored one of three.
  • What it does not claim: that async testing beats interviews at understanding a market. We have run no comparison and will not publish one without it. The claim is narrower — a result you committed to in advance survives your own motivated reading of it.

2026-09-06

The Y Combinator guide now runs the current list through our own line

  • Y Combinator reissues its Requests for Startups every batch, and /guides/y-combinator-startup-ideas argued about accelerator-shaped ideas in general without ever naming one. It now sits all thirteen items of the Fall 2026 list against the single rule the page is about — whether the barrier in the way is a process money can clear or a domain money cannot enter.
  • The split is 4 prohibited, 2 process, 7 decided by something else entirely. That last column is the point: for over half the list the regulatory question is not what decides it, which throws the decision back onto the checks that never cared how well funded you are.
  • Who said what is kept separate, deliberately. The item titles are quoted verbatim and we characterise nothing about what YC wants. The one-line barrier beside each is ours and the page says so. The verdict in the third column is neither — it comes from the same function the kill gate calls, so the table cannot drift from the product, and extending that list moves both in one commit.
  • Building it found a real hole in that function: the weapons veto matched "munition" but not "munitions", so a weapons kill sitting next to a payments line-item could have been downgraded to an overridable one for a funded team. Fixed, with the guard written first and watched fail.

2026-09-06

What actually kills an idea, counted rather than asserted

  • There is a new guide at /guides/why-startup-ideas-fail, and it does one thing: it takes the run we publish in full and counts which check did the killing. 36 of 69 ideas were dropped, and the biggest single cause was not a crowded market, a competitor, or the price being wrong. It was that no documented complaint could be found behind the idea at all — 21 of them, more than twice the next reason.
  • The page volunteers the awkward part. Those reasons add up to 57 against 36 dead ideas, because 21 of them failed more than one check, so the rows look like they divide up the drops and they do not. Most weak ideas are weak in more than one direction at once, which is worth knowing if you were hoping that answering the one objection you have heard would save yours.
  • Both figures on the page are computed from the same function the run's own page uses, rather than drawn once and exported. Regenerate the run and the sentences, the bars and the funnel all move together. A chart carrying one of our numbers is still one of our numbers, and a picture of it goes stale exactly as fast.
  • What it does not claim: it is one run, the checks are ours rather than the market's verdict, and the ratings underneath sample fresh each time. All three limits are on the page, at the end, in those words.

2026-09-06

You can take a validation result back, but you cannot try again for a better one

  • Reporting what happened when you ran a validation plan was free and unlimited, the newest report was the one that counted, and the page named every criterion you missed and by how much. Read that back and the problem is obvious: it told you the answer and then invited you to submit again. The whole reason this product will accept numbers you report about your own test is that the bar was written down before you ran it — and a bar written in advance is worth nothing if the answer can be rewritten after you have seen how it graded.
  • So within one plan, the report that counts is now the weakest one you filed. Correcting a number downward takes effect immediately. Correcting upward does not — the earlier, worse report stays the graded one, and the page tells you that is what happened rather than quietly showing you a grade that does not match what you just typed.
  • Weakest rather than simply first, because first-only would trap you: mistype 500 where you meant 50 and you would have minted a result you could never withdraw. The direction that can only ever cost you evidence stays open. The direction that can only ever gain it is closed.
  • This is per plan, not per idea. A new plan and a new test start clean — an early failed test should not cap an idea forever, or the product would be punishing you for testing. The trade-off, said plainly: if you report an interim result and the same test later does better, you cannot upgrade it. That one needs a fresh plan, which is what running a test against a bar set in advance actually means.

2026-09-06

The published example run now says which checks did the killing, before you scroll

  • The example run at /sample-discovery has always listed every idea it threw away and the reason for each. But that list sits below the shortlist and the whole evidence base, so anyone reading only the top of the page got the arithmetic — 36 of 69 dropped — without the map. Somebody read the run this week and asked us to add the tally. It was already there, a couple of thousand words further down. That is a placement problem and it is ours, not theirs.
  • So the tally is now in the opening paragraph: no real problem behind it 21, too much hands-on support 9, needs heavy compliance 8, hard for one person to sell or run 6, just an AI wrapper 6, leans too hard on one platform 4, and three more with one apiece. Which assumptions keep failing is a more useful thing to know than how many failed, and you can have it without buying anything.
  • It also now admits something that had only ever been a code comment: those numbers add up to 57, not 36, because 21 of the ideas failed more than one check. The rows look like they divide up the drops and they do not. Better that we say so than that you add them up and catch us.
  • Both places on the page are printed by one counter, so they cannot drift apart, and nothing in that paragraph is a number we typed in by hand.

2026-09-06

Ran your validation plan? Now you can tell us how it went — and it counts

  • Until today the product could tell you “go test this” and then had no way to hear the answer. That was not an oversight so much as a dead end: the strongest verdict, “build it”, needs confirmed distribution and hard evidence of demand, and neither is something you can learn by reading pages about an idea. You learn them by running the test. In every run we have ever done, on every mode, “build it” has never once been issued — not because the tool is sulking, but because the door was on the wrong side of the wall.
  • So the no-call validation plan now closes its own loop. When a plan exists for an idea, its page has a short form: the dates the test actually ran, and what you counted — visitors, signups, paid conversions, money collected, replies. Numbers and dates only. There is no “how did it go?” box, and there is not going to be one.
  • What makes this something other than typing your way to a good verdict is the order things happen in. The plan writes down what success looks like — a threshold and a deadline — before you run anything. When you come back, code compares your numbers to that, and nothing else. You cannot re-run the plan to fetch an easier target afterwards: a plan created after your test started is refused, and it says so.
  • Meeting the bar supplies the evidence the “build it” verdict was missing, so it becomes reachable on your next check. Missing it, or half-meeting it, supplies nothing — a partial test is an unfinished one, not a smaller proof. And your report carries a plain label everywhere it appears: you reported these numbers, we did not measure them, and we did not open the page you shipped. A verdict standing on your own account of your own test should say so.

2026-08-29

Leaning on a marketplace is no longer, by itself, a reason to bin an idea

  • Discovery had two of its own settings quietly fighting each other. If your profile says you sell without calls, a marketplace or an app store is often the only channel that puts a buyer in front of you without a conversation — and then the sweep would throw those ideas out for depending on a platform. The check that kills an idea outright never knew it was penalising the price of a channel our own profile question had pushed you toward.
  • So platform dependency is now the one check that does not kill on its own. An idea that fails it and nothing else is kept, scored in full, and ranked below everything that cleared every check — it takes a shortlist slot only when there is room. An idea that fails it plus anything else still dies, as before. This is how the single-idea check has always treated it; the sweep was simply stricter than the mode that owns the rule. Looking back over every run we have done, this changes the fate of 15 ideas out of 706 rejections — small, and every one of them was a Shopify app, an Upwork tool, a Discord product or an Instagram scheduler.
  • Your result page now has a section called “Carried forward with a flag” listing exactly which ideas got through this way and the reason given. We would rather show you a relaxed check than have you find an idea on the list and not know one was relaxed.
  • And the founder profile has a ninth question: how you feel about building on someone else's platform. It decides the above — set it to low and the sweep goes back to dropping those ideas outright. That value already existed and was already being fed into the analysis, but nobody was ever asked for it: it was set to “medium” for every account, and the sweep had taken to quoting it back in its own rejection notes as though it were your answer. Now it is.

2026-08-24

Two doors instead of twenty things to click

  • The front page and the signed-in home have been rebuilt around one question: do you have an idea, or not? Two doors, the same two on both pages, in the same words. Before this, the home screen offered about twenty places to click — a headline button, a six-tile feature block, a row of twelve capability names, a second grid of tiles, and a callout — and the front page ran to nine sections before it got round to asking what you wanted to do.
  • The twelve capability names are gone from both pages. They were the product's internal vocabulary — pricing power, support burden, platform risk — and reading them was a job of work before you could start yours. They still all exist, and you still get to all of them: they open from an idea you have saved, which is the only place they can run anyway. Our own written UX rules have said not to put them on the home screen since before the product launched, and the implementation had drifted away from that.
  • What we deleted and what we kept: gone are the feature-card grids, the three-step explainer (the how-it-works page says it properly and is linked), and the counting-up numbers strip. Kept are the parts you can check — the published run's own figures, the source-linked problem library, and the day a rival validator scored our own idea 72 out of 100 while our tools scored it a kill.
  • One thing now asks a question before it runs. Discovery — the sweep that goes and finds ideas — ranks everything it finds against your founder settings, and it was willing to run on the defaults we had filled in for you, which means it was scoring a shortlist for somebody else. It now asks for those settings first, and the doors say so up front rather than springing it on you after you have bought credits. Checking a single idea still never asks: that gate was tried in front of the free first check and nobody got past it, so it is staying off there.

2026-08-16

A place for the questions that are not about us

  • There is now an answers library at /answers. The FAQ answers questions about this product; these answer the question you actually have — whether an idea is worth building, and how you would know. The first one is up: whether you can just use ChatGPT for this. Our answer starts by agreeing with you for two paragraphs, because for stress-testing an idea you already have, a chat is genuinely fine and free. The part it cannot do is go and look.
  • Every one of these pages has to name what in our own system backs it before it can be written — a run, a measurement, or a check the product really applies. If a question does not have that yet, it does not get a page. That rules out several questions we would otherwise rank well for, and we would rather have a short shelf than pad it out with advice anyone could have written.
  • No numbers are typed into these pages. Every figure is read from the thing that produces it, so a page cannot quietly keep claiming a number after the number changes — a test fails the build if one is typed by hand. The same goes for the share images: they are generated from the same source as the page rather than exported once and forgotten.

2026-08-13

The share cards were missing on five pages, and one date was telling search engines a page existed before it did

  • Five pages — About, the FAQ, How it works, the benchmark and the tools guide — were sharing with no preview image at all. Not because nobody made one: they each set their own social title and description, and doing that silently replaces the site-wide card instead of adding to it. Anyone pasting those links into Slack, X or LinkedIn got a bare grey box. Fixed, and a test now fails the build if a page sets its own social details without an image.
  • Our idea-library pages were telling search engines they were published on the day the run happened — but the pages themselves went live over a week later, so we were dating them before they existed. The run date is now recorded as when the analysis was made, which is what it is, and we no longer claim a publication date we cannot support. The same correction applies to the published Discovery run.
  • The guide's page said it had never been edited since publication, which was not true. It now shows both dates, and the visible one and the machine-readable one are the same value — if it ever goes stale you can see it, which is the only thing that keeps a hand-maintained date honest.
  • Smaller fixes in the same pass: the idea pages now lead with the verdict in search results rather than burying it after the description; the pricing page names its actual price range, taken from the price list itself so it cannot go stale; the tools index finally declares the same structure its three children already pointed at; and the guide's breadcrumb no longer shows a middle level that was a dead link.

2026-08-13

You can now see how much we actually have on a topic, and when we last saw it

  • A topic page showed up to 80 problems and said nothing about the rest, so a topic with 36 problems and one with 486 looked identical. Each page now says how many it is showing out of how many we have sourced, and the topic index shows the number beside every topic. Both numbers come from the query, so they cannot drift from what is on the page.
  • Every problem now also shows the month a sweep last turned it up. That was in our data and never shown, which meant you could not tell something found last week from something found a year ago — on pages whose whole argument is that you can check the evidence yourself. It is a month rather than a date on purpose: it records when our search last saw the problem, not when the person wrote it.
  • And an admission. The idea-library cards had a section headed "the evidence behind this", described as "what a real sweep leaves you with". We went to add it to the six cards that were missing it — and found that the two that had it were written by hand when that feature shipped, not produced by the runs those cards came from. Their texts appear in none of our run records, and their links open a search rather than the statement itself.
  • So we did not add six more. The two that exist now say what they are — illustrations of the problem, each linking to a search where you can read people describing it — and the six say plainly that no per-problem source list was recorded for that card, with a link to the topic page where every row does carry a source and a label. A test now fails the build if the old framing comes back. Filling that section out would have been easy and would have made the pages look better; it would also have been the exact thing we criticise other tools for.

2026-08-13

The real-estate topic page was showing Shopify sellers writing product copy

  • Every one of the ten problems at the top of the real-estate page was about e-commerce: Etsy and Shopify merchants writing SEO product copy. All ten had landed there because the page matched on the word "listing", and a product listing is not a property listing. It matches on real-estate wording now, including terms for the two segments that were being missed entirely — MLS feed integration for a solo agent's own site, and short-term rental coordination.
  • The client page had the mirror-image problem: it matched the bare word "client", which appears in nearly every service business, so 96% of its problems also appeared on some other topic page. It matched "delivery" too, which was mostly Shopify shipping dates. Narrowed to onboarding, hand-offs, client portals, status updates and customer success, and renamed to match what it actually holds. Same address — the page was fixed, not removed.
  • The SaaS-metrics page was matching the bare word "retention", which pulled in client retention for pet sitters, trainers and photographers. That one is now customer retention specifically.
  • What we did NOT do is worth more than what we did. Two more pages were going to be narrowed the same way — "lead" also matches "leads to", "content" matches almost any sentence — and before shipping it we read the problems each change would have REMOVED. On both, most of what would have gone was squarely on topic: real lead pains for photographers and pet-boarding operators, real content pains for tutors and solo creators. Narrowing them would have thrown away around forty genuine problems to remove a handful of irrelevant ones. Both words stay.
  • What fixes those two instead: every problem on a topic page now shows its theme. A problem can be on a page because its theme matched, even when the sentence you read does not contain the topic word — on one page that was 35% of what you saw, with nothing on screen explaining why. That works whatever the words are. Tightening words does not: on one of the pages we tested, tightening made the unexplained proportion worse rather than better.
  • Every page was counted before and after, and all thirteen still hold well above the minimum we require before publishing a topic at all.

2026-08-13

No page we ask Google to index is reachable only through the footer

  • Two pages were in our sitemap — so we were telling search engines they mattered — while the only link to either was in the footer at the bottom of every page: the list of ways this product could fail, and this changelog. The About page now links both, in the section about the parts we would rather not say, which is where they belong.
  • A test now checks every page in the sitemap for that condition and fails the build if one falls back into it. Getting it right took two passes, and the first was wrong in a way worth admitting: it counted links on signed-in pages, which no search engine can load, and so reported three legal pages as well-linked when the footer was in fact the only way to them. Those three now carry a written exemption — for a terms page, the footer is where people look — and the test only counts links a visitor without an account could follow.
  • The same fix uncovered that our idea library page could have been trimmed to a top-few list without anything noticing that the other cards had become unreachable. There is now a test that renders both library pages and checks every single entry is linked.

2026-08-13

Every pain topic and idea card now points at its neighbours

  • The topic pages and the idea cards were near dead ends. Each had a couple of links out — the index, a sample report — but nothing sideways: no way to get from booking pains to the scheduling side of contracting work, no way from an idea card to the other ideas from the same market, and no way from either to the problem underneath.
  • Each topic page now ends with three or four related topics, hand-picked rather than computed from keyword overlap — overlap is what put loosely-related rows on the same page in the first place, so we are not using it to decide what is related to what either.
  • Each idea card now lists three other ideas from the same market, with their verdicts, and links back to the pain topic that market sits in.
  • The link back to a topic says "sits in", and that wording is doing real work. Nothing was ever scored against a topic: the eight published cards come from two runs over one market, and the corpus the topic pages read did not even have a way to be written to until a week after those runs — so those cards' problems were never in it. A first draft of this change said the cards were "scored on this problem", which would have been a made-up connection on the pages whose entire pitch is that you can check the source. Eleven of the thirteen topics have no card at all, and they say so and point at the library rather than being handed one from an unrelated market. A test now fails the build if a card is relabelled into a market it has nothing to do with.
  • Both templates also link the pricing page, which sounds trivial and was not: the only commercial link on those pages went to signup, which is closed to search engines, so a crawler could see no way from a topic page to what any of this costs.

2026-08-13

The 404 page sent you somewhere you could not go, and four links were named after URLs

  • The way out of our "this page doesn't exist" page was a link to your ideas list — which, for the person most likely to be reading a 404, is a page that bounces you to a sign-in screen. It now offers the pain library, which is public and needs no account.
  • Four links on the homepage were labelled with their own address: "See /trending", "Browse /library". They now say what is on the other side — the feed of trending problems, the scored idea library, the whole Discovery run.
  • And several pages that should have pointed at each other did not. If you were on the pricing page wondering how we differ from another tool, we offered you nothing; the comparisons were reachable only from the footer. The pricing page, the FAQ and the benchmark page now link them in the text, the benchmark page links all four rather than one, and the three free no-signup tools are now linked from the pages where you would want them instead of only from the tools index.
  • Nothing was removed and no price, number or verdict changed — this is signposting.

2026-08-13

We stopped repeating a number about a competitor that we could not show you

  • Our comparison pages said, in several places, that independent reviewers report IdeaProof's scores skew high — an average around 78 out of 100 — and that its reports list only about five sources. Both were hedged as "reviewers report" and both were dated. Neither had a source we could point you to. Going looking for one, we found something worse in our own notes: the same file records that IdeaProof has essentially no independent reviews at all — none on Trustpilot, G2 or Capterra. There cannot be reviewers numerous enough to average a score that almost nobody reviewed. So the claim was not just unsourced, it was unlikely.
  • Both are gone, from every page and from the FAQ answers that Google reads as structured data. What replaced them is something we actually did and can date: on 2026-06-10 we put our own idea through IdeaProof and it came back 72/100, "PROMISING", "Excellent potential! You're among the top ideas we've seen", while our own gate killed it. One run on one idea, and we say so rather than turning it into a claim about anyone's average.
  • Where a competitor advertises something — "50+ sources, every claim cited" — we now say that they advertise it, which you can check on their own site, and we leave the counting to them. Two entries in the list of other tools also described how a rival works inside; that was never something we could verify, so they now describe what each tool sets out to do instead.
  • New rule, written into the code so it outlives this decision: a negative claim about a named competitor has to be first-party — something we ran, saw and dated. A test fails the build if that hedge comes back.

2026-08-13

Four pain pages claimed we had read pages we had only searched

  • Four of the pain topic pages opened by telling you the problems below came from pages we had opened and read — one of them said every signal was "sourced from a real page our discovery sweeps fetched". For most of those rows that was not true. Across the thirteen topic pages, 78% of the problems come from a search result where we saw the summary and never opened the page, and the label under every single row said exactly that the whole time. So the page disagreed with itself, and the sentence at the top was the part that was wrong.
  • That sentence is also the description Google shows under the result, which means the claim we could not back was the first thing a searcher read and the labels correcting it were the second. The four sentences now say where the problems were found and leave the question of whether we read the page to the label on each row, which is where it was always answered honestly.
  • The idea library had a smaller version of the same problem: the reasons under each idea used our internal rating words — "MEDIUM", "UNKNOWN", a "BUILD" verdict we never actually show you — in seventeen places, and those reasons feed the Google description too. They now read in plain words. The ratings themselves are unchanged; only the words are.
  • And on the pages listing other tools, two entries described how a rival works inside — what its model reasons from — which is not something we can check from outside. Those now describe what each tool sets out to do, which you can confirm on its own site. One entry also repeated a vendor's own count of what it has mined as though we had counted it; it now says that they advertise it.
  • What has not changed: the per-row labels were accurate before this and are accurate now, and where a row has a link it still points at the page the problem was found on.

2026-08-11

First-customer leads can now run on a schedule, and remember who you already saw

  • Finding your first customers was a one-off: you pressed the button, read the list, and next time you pressed it you got the same names mixed in with new ones and no way to tell them apart. Leads now live in a feed on the idea. Each one is a person, not a page — several people arguing in the same thread are several leads, each with their own quote and their own draft — and you can save the ones worth chasing or dismiss the ones that are not.
  • You can also let it run on a cadence you choose, from daily to monthly. It is off until you turn it on, and turning it on means agreeing to the price in plain words: one paid credit per completed scan including an honest no-result one, a monthly cap you set, no automatic top-ups, and pause whenever you like. Pausing refunds work that has not started. If a search provider fails, the scan is retried and the credit comes back rather than being spent on an outage.
  • Editing the idea or your founder profile pauses the schedule instead of quietly carrying on, because the approval was for the idea as it was written.
  • What has not changed: we find and draft, you send. Every lead still gives you Open and Copy, and a public post is still not treated as a verified person, an address, or permission to contact anyone.

2026-08-11

A first-customer lead could be thrown away over a www

  • Every first-customer lead has to point at a page the run really found — that check is what stops the tool handing you a link it made up. It was comparing web addresses too literally. If the page we found was written www.indiehackers.com/post/… and the lead came back as indiehackers.com/post/…, the same page read as two different ones, and the lead was discarded as invented. Same page, same person, same quote, gone over four characters. The check now ignores the parts of an address that never change which page it is — a www, capital letters in the host, the tracking tags a search engine bolts on — and stays strict about the parts that do: the path, and any id inside the address.
  • The other half of this is that we could not have told you whether it ever cost you a lead: a refused lead left no record anywhere. It leaves one now. What we can check says nobody was hit — every lead we have delivered came through in the same spelling we found it in, and no run has ever had all of its leads refused. So this is a fix for a way we could have lost your leads without either of us finding out, not an apology for leads you lost.

2026-08-10

First-customer leads: we now search the way your customers talk

  • A lead search used to take the buyer and problem you typed and paste them into a search box as one long sentence. Nobody writes that sentence. Your customers write "our AWS bill is getting out of control" — you wrote "log storage grows into hundreds of dollars a month", and the search found the second, which almost nothing matches. Every run now starts by rephrasing your idea several ways in the words the people you're looking for actually use, across different angles: the symptom, the cost, the manual workaround, the hunt for a tool, the venting.
  • We also stopped spending your run on pages that can't contain a person. Roughly half of every previous search came back as software directory listings and unrelated job ads — no human describing anything. Those are gone, and the freed-up room goes to forums and communities where people post in the first person.
  • One busy thread often holds several different people with the same problem, and we only ever gave you the thread once. Each of them is now their own lead, with their own quote and their own draft message.
  • None of this changes the honest part: every lead still has to be grounded in a real page we found, and a thin result still comes back thin instead of padded out.

2026-08-08

The privacy page now names what the lead search sends

  • First-Customer Leads searches from the buyer and pain saved on an idea. The privacy page used to say idea text was never sent to a source connector — false for this mode. It now names the query phrases and every supported public-source destination, while keeping the boundary clear: they do not receive your founder profile, email or billing data.

2026-08-08

First-customer leads: one click to the thread, message ready to send

  • The First-Customer Leads panel already found public, source-linked signals describing your problem and drafted a short message for each. It stopped there — you still had to hunt down the thread, copy the message, and paste it. Every lead now has an Open button that takes you straight to the source for a reply or assisted DM, or to your mail composer with the message already in the body. A Copy button puts the message on your clipboard in one click.
  • We still don't send anything for you, and there's no auto-blast: you open the message, read it, and send it yourself. That's deliberate — the platforms where these people posted ban automated messaging, and a wall of identical bot DMs is exactly what we're not. The tool does the finding and the drafting; the send stays yours.
  • We never put an email address on a lead we scraped it from — for the email option the composer opens with the message ready and you add the recipient.

2026-08-08

A second QA pass, and the fixes it turned up

  • The cleanup that removed our internal name for the problem corpus from the proof page left the sentences ungrammatical — "No our record of problems entry covers…" appeared a dozen times in the kill receipts, the one section the whole page rests on. They read as English again.
  • A Discovery run you paid for now opens "Why we rejected the rest" by default, the way the public sample already did. The section that shows what the gate threw away is the one that answers "what did you do for my money", and it was the page for the person who paid that kept it folded shut.
  • You can now delete a Discovery run or a whole sweep. It was the most expensive thing you could run and the only one you couldn't tidy away — a typo'd market lived on your account forever while a free idea deletes in two clicks. The credit isn't refunded, because it paid for a sweep that really ran.
  • Running "Mine pains" with nothing to mine from — no link, no pasted text — used to take a credit and hand back nothing, on a product whose promise is that you see the price before you spend it. The button now waits until you've given it a source.
  • Running the deliberately brutal Kill My Idea teardown on a promising idea used to drop that idea into your graveyard, the page titled "every idea you've killed". The graveyard now reflects your real decision, not the lens built to argue the other side. Its verdict still shows, labelled, on the idea.
  • On a phone, the landing page's first screen had nothing to tap — the one action was collapsed into the menu and the next was a scroll away. There's now a button on the first screen.
  • The evidence list stopped claiming the linked pages are forum posts and reviews when a run's sources are product listings — it now describes what a link actually is. A pricing card no longer implied the cheaper pack can't cover a full run when it can. The sign-up page and the landing now say plainly that Discovery sweeps need paid credits, the way the pricing and billing pages already did. And a benchmark page stopped calling its numbers "reproducible" — the ratings are sampled and move between runs, which is the honest thing to say and the thing the rest of the product says.
  • The waiting screen for a multi-market sweep stopped promising the markets "run one after another" — several can run at once, so a batch usually lands sooner than adding them up. And the landing page finally has a real title in your browser tab and in a shared link, instead of just the brand name.

2026-08-07

The answer comes first, and the score can be checked

  • The page we publish as proof — a real Discovery run, unedited — opened with its evidence and buried the shortlist. Measured: on a phone you scrolled past twenty-one screens of problems, kill receipts and sources before the first idea. All of it is still there and still open; it now sits after the five ideas instead of in front of them. On your own result page the shortlist and the recommended first move moved up to where you land.
  • Every idea shows a score and a letter grade, and neither could be checked. The scoring breakdown now prints the weight beside each of the seven areas, so the ratings and the total reconcile in front of you, and it says what the letters mean: A 80+, B 65–79, C 50–64, D under 50. Most ideas land in C or D — the bar is set for what one person can build, sell and support alone, and you can now see that rather than take it from us.
  • A run scores far more ideas than it shows you. "Why we rejected the rest" only ever covered the ones a deal-breaker check killed — on the run we publish, 36 of 69 — while the 21 that passed every check, got a full score, and then lost on rank were never named at all. They now have their own list, with their scores, under "Also scored, didn't make the shortlist".
  • The funnel had a rung reading "Backed by a real problem: 69 of 69" on the same page where the receipts said 21 ideas died for having no real problem behind them. Both numbers were right and the labels made them contradict: that rung is a loose text match, and the strict check runs later. It now says what it measures, and says how many the strict version dropped.
  • Two pages said a Discovery sweep of about thirty markets was what one run produces. One run screens up to 150 ideas; the sweep was thirty separate runs. Fixed, and the four numbers behind it now come from one dated record instead of being typed into three files.
  • The example result card on the home page and the landing page was invented: a grade-A 88 with "Build it". The run we actually publish tops out at 52, grade C, "Test it first". Both cards now show the real one.
  • Where a run can come back with nothing, or come back different the second time, we now say so before you spend rather than only on the pages you read beforehand: on the launch form, on the proof page, and beside every "run it again" button. Running the same market twice is now offered as an honest option, priced, rather than left out.
  • The pages said "you can check every claim" while the run we publish links 32 of its 56 problems. It now says which, and how many, read from the run itself. The evidence list also stopped promising that every link goes to someone describing the problem — some are product listings where the gap is visible rather than stated, and saying so is the version that survives a click.
  • A market that failed inside a sweep was a dead label: no link, no explanation, and nothing about the money, on the surface where the most credits are at stake. It now links to what happened and says whether the credits are back — checked against your ledger, not assumed. The email about a sweep does the same; it used to state the refund as a fact without ever looking.
  • Small things that were dead ends: a failed run told you to email us and gave you nobody to write to; your runs list showed "failed" with no word about the charge; the billing page asserted the refund policy without linking it. All three now link out. There is also a skip link on every page, and the waiting screen says how long your run has been going instead of looking identical at minute 2 and minute 38.
  • An idea whose only analysis was a Kill My Idea teardown was shown a card offering to "tell you whether to build it, test it, or drop it" — about an idea we had already told you to drop, at 15 out of 100. Each collapsed panel now carries its own answer in the header, so a fully-analysed idea reads as a list of verdicts rather than a dozen identical grey bars.
  • Our own words that were still on customer pages: "provenance", "candidates sourced", "sweep" where "market" was meant, "STRONG"/"MEDIUM" as raw labels, "Pain Box" seventeen times inside the reasons on the proof page, and a "BUILD / VALIDATE / KILL" verdict the product never actually prints. Replaced. The privacy page also now names the search provider that receives the market you type, and separates what our analytics do before you accept cookies from what they do on the server either way.
  • The two free pain pages were the worst of that: /trending and /pains labelled every problem with the raw "STRONG" or "MEDIUM" from our database — eighty of them across the two pages, on the first screen a visitor who has never heard of us sees. They now read "strong", "some" and "weak", the same words the paid result uses. Their opening sentences also stopped describing themselves in our funnel vocabulary and now say where the problems come from.

2026-08-05

A broken run now tells you about your money

  • When a Discovery run failed, the page said "something went wrong, starting a new run usually clears it" — and nothing about the credits you had just spent, while the cheaper single-credit run has always said it. It now says whether the refund has landed, checked against the ledger rather than assumed: if it's there we say so plainly, and if it isn't we say a failed run is never charged and to email us. It also warns you that the Try again button costs credits, which it always did.
  • The same check now covers runs inside a sweep. A sweep is charged and refunded as one unit, so a failed market inside one never carried a refund record of its own — meaning the honest answer for the runs with the most money at stake was the one we couldn't give.
  • Before you fill in the Discovery form, we now check whether your credits can actually cover a run and say so at the top, with what a run returns and a link to a real one. You used to find out after filling in four sections and pressing the button.
  • The form itself now leads with what the credits buy — the funnel from ideas found down to the shortlist, every idea we killed at the checks and what killed it, and up to five ranked picks — instead of four numbers about pricing and nothing about the product. And the section asking for your own links stopped calling them "sources to ground on": it asks for anything you've already found, and explains in plain words that if a page tries to give our AI instructions we ignore them.
  • The FAQ now has a jump list, so you can go straight to your objection and link a single answer to someone else.
  • Several of our own words were still on the page: "stop-loss" for the point where you should walk away, "survivors" for the ideas that passed, "pain-gated" as a heading. All replaced — and a test now sweeps the pages, the emails and the downloadable report for them, plus for six claims we've had to correct before, so the next one gets caught by the build instead of by a reader.

2026-08-02

The ten checks we kill ideas with are now published

  • We've always said we throw almost everything away against a fixed list of deal-breaker checks. The list itself only existed inside our code, so "36 of 69 were dropped" was something you had to take on faith. All ten are now written out on the How it works page, each with the exact words you'll see on an idea it dropped. A test reads the real list out of the pipeline and fails the build if the published one drifts from it.
  • The result page used to end on the last idea. It now closes by telling you what the run actually covered — how many sub-markets, how many candidates — and what it didn't, plus the one pricing fact worth knowing there: adding a market to a run costs 1 more credit, not another 2. It only appears when the run found something; a run that honestly found nothing gets no sales pitch.
  • The Sources section was a crawler's log: "Swept 93 hits from 60 queries", "skipped (robots.txt)", "bot-blocked hosts". Same facts, in English — including why some evidence is a search result rather than a page. The evidence header stopped saying "pain signals" and "provenance", and each finalist's money line stopped producing sentences like "reliance on one platform medium".
  • After paying you landed on a page whose only button said "Back to home", and the flagship wasn't in any navigation at all. Both fixed: buying credits now points at the thing you bought them for, and Discovery is in the header. Signed-out visitors get a "See a real run" link there too — previously the header offered only "pay" or "sign up".
  • Every link in the product promises a "founder read"; the page it opened called itself "Founder profile first-pass". Now it says what the link said.

2026-08-01

The email about your run now contains the run

  • When a search finished, we emailed you its shape: "3 finalists — each scored against your profile, with a verdict and a plan." A product that did nothing could have sent that. The email now carries the actual work: how many searches we ran, how many pages we read, how many candidates we put through the checks and how many didn't survive, then the top idea by name with its verdict, its score, and the documented problem it's anchored to — with the link to that problem, when we have one.
  • If nothing cleared the checks, the old email said "0 finalists — each scored against your profile, with a verdict and a plan", which is a sentence about nothing. It now says what we searched, that none of it cleared, that this is a result rather than a broken run, and points you at the part that's actually useful: every candidate we dropped and why.
  • If your run broke, we sent you nothing at all — after the page had told you to leave and wait for the email. That was the worst silence in the product. You now get an email saying we couldn't finish, that the credits are already back in your account, and that it's on us rather than on your market.
  • A sweep of several markets sends one email at the end. Each market's line used to read "failed" — our internal database word — with no explanation and no mention that we'd already refunded it. It now says "couldn't finish — credit refunded", tells honest zeroes apart from breakages, and states the refund plainly. The header also undercounted: a market that finished with some unknowns was counted as neither done nor failed, so a five-market sweep where everything finished could announce itself as "3/5". Same bug on the sweep page, fixed in both.
  • One page was promising an email it could never send: an individual market inside a sweep. That page now says the email comes when the whole sweep finishes, and the sweep page — the longest wait in the product, and the place the promise is most true — says it at all for the first time.

2026-08-01

Discovery no longer starts with a form about you

  • Opening Discovery without a founder profile showed you eight dropdowns instead of the page. That's the same wall we removed from the idea check in July, and it was worse here: if you came from the pricing page, signed up and bought credits, you arrived with paid credits in your balance and got asked about your support tolerance before you could spend them. You now get the form, on sensible defaults, with a line saying they're defaults and a one-minute link to set your real ones.
  • The home page told new accounts that Discovery starts "from 2 credits" while their balance chip read 2 credits — an amount that can't buy a run. It now says, in the same breath as the price, that Discovery needs purchased credits and free ones cover the single-idea checks.
  • When you don't have enough credits, the message now says how many the run needs and how many you can actually spend on it, tells you which pack covers it, and links to a real completed run so you can read the output before deciding. It used to say "you're out of credits" and point at a price grid.
  • While a run works, the page now counts up what it's found — markets planned, searches run, pages read, ideas found, ideas dropped — and says which step it's on in words rather than one sentence that never changed for 25 minutes. The step names are plainer too: "Discovering" and "Searching" read as the same thing, so progress looked stuck. And the page used to quietly stop refreshing after 24 minutes, which is roughly our median run — it now keeps going to 45 and tells you when it has stopped instead of leaving a spinner that looks alive.

2026-08-01

We stopped hiding the work

  • Our own sample page — the one whose job is to prove we show real output — opened with all three evidence sections collapsed, and its copy asked you to click them open. They're open now. It also only told you how much got thrown away (69 → 5) and never what that took: 60 searches, 93 results, 24 full pages read, 41 links you can open. That line is there now, and every number is read from the run itself.
  • Two things that page was quietly not saying: 33 ideas passed every check but only 26 got scored, and the 7 that vanished were invisible between two rows of the funnel — the page now names them and says our scoring stage, not our judgement, is why they're missing. And the 36 kills were 36 paragraphs with no summary, so what actually does the killing was invisible: no documented problem 21, support load 9, regulation or trust 8, and so on.
  • The landing page never said what you receive. It opened with process and posture; the words shortlist, source link and next move appeared nowhere above the fold. It now leads with the thing itself. The hand-written 88/BUILD card in the corner is labelled Example, which it always was — our sample page said "nothing here was written by hand" while the landing quietly implied the opposite about an illustration.
  • The pricing page's only outbound link was Refunds, on a screen that tells you to check every claim. It now opens with three links to real output, explains what one run actually does before it charges you for one, says out loud that free credits don't cover a Discovery run, and answers the chatbot question where the money leaves instead of only on the landing page. We also dropped the "Most founders start here" badge: we've had six pack purchases, so it wasn't ours to say.

2026-08-01

How long a sweep takes, and reasons written in English

  • We told you a Discovery run takes "a few minutes". We went and measured our last eleven real runs: one market takes 10 to 40 minutes, about 25 in the middle. So that's what the page says now — on the form before you spend a credit, and on the progress page while you wait, where a quiet ten minutes previously looked like something had broken. The single idea check, meanwhile, we were underselling: it finishes in about nine seconds.
  • The reasons under each finalist were written half in our own internal shorthand. You'd read "Phase 6 rates demand as MEDIUM", or that a score "sits in the weak band", or see a raw id like b3_c14 where an idea's name belonged. Nothing in the product is labelled Phase 6 and we never show you a band, so those sentences pointed at things you couldn't look up. They now name the check that spoke — "the evidence check rates demand as medium" — and refer to other ideas by name.

2026-08-01

Sweeping many markets now costs what it costs

  • A Discovery run was capped at 15 credits no matter how many markets you picked, so a 29-market sweep cost the same as a 14-market one. That cap is gone: it's 2 credits for the first market and 1 for each additional, with no ceiling. Nothing changes for one market, or two, or ten — the price you already knew is the price. Only sweeps past fourteen markets go up.
  • We removed it because it was wrong, not to charge more. Each market you add is a separate search of its own — its own sources, its own kill-gate, its own shortlist — and we measured what one costs us: $2.68. Under the cap, our biggest sweep cost us about $78 to run and, on the largest pack, sold for $74.50. Below cost on that pack and thin on the others, which is not a discount — it's a thing that quietly stops being offered. Now the largest pack buys exactly one sweep of every market in our catalog, and that sentence is true rather than aspirational.

2026-07-31

One email if you signed up and never got a result

  • If you make an account and don't run anything within a day, you now get a single email pointing at the two things your free credits actually cover — a founder first-pass and a single idea check — plus a link to a real Discovery run you can read without spending anything. Sent once, ever, and it doesn't ask you to buy: you haven't seen a result yet, so asking would be backwards. Every other email we send needs you to have already run or bought something, which meant anyone who stopped at the signup page heard nothing at all.

2026-07-31

Pain Receipts finds what you meant, and says when it didn't

  • Searching the free pain tool used to need every word you typed to appear in one place, so "hvac contractors" found nothing even though 59 sourced pains discuss contractors, and "AI tools for real estate agents" lost most of its results to the words "tools" and "for". Now filler words are ignored, and when nothing matches exactly you get the closest coverage we actually have — clearly labelled as that, never dressed up as an exact hit. The examples in the search box are ones we verified return real results.
  • Two honesty corrections while we were in there: the funnel numbers on the home page said 93 ideas "passed the checks" with the bar drawn nearly half-width, when 93 were the finalists out of 3,430 — a 2.7% survival rate, which we now show and state. And after a free founder read, the next step offered was a paid Discovery sweep that free credits can't buy; the free single-idea check is offered first now, and the paid one says it's paid.

2026-07-31

Checking your first idea no longer starts with a form

  • You used to have to answer eight dropdowns about yourself before you could check a single idea. Now the idea check opens straight to the idea. We start you on sensible defaults — solo, no sales calls, ~$10K/mo target, low support tolerance — and we say so on the page instead of quietly scoring you against assumptions you never made. Setting your real profile takes about a minute, you can do it after you see the first verdict, and a re-run picks up the new answers.

2026-07-30

You can read a whole Discovery run before paying for one

  • There's now a published Discovery run at /sample-discovery — a real sweep on real-estate & property software, dated, with nothing rewritten to look better. The full funnel it went through, every finalist card with its rubric breakdown, the evidence links, and the recommended first move with its stop-loss. It renders through the exact same components your own run does, so it can't quietly become a prettier version of the product. No account, no credits.
  • Two smaller honesty fixes shipped with it: a run that can't be scored now fails and refunds instead of handing you an empty shortlist, and when a provider is out of quota you get a plain 'try again shortly' instead of our vendor's internal error.

2026-07-19

The pain library grew — seven more topics

  • The browsable pain library now covers thirteen topics instead of six — new pages for lead & sales follow-up, coaching & consulting, small-agency operations, SaaS metrics & analytics, contractors & trades, fitness & wellness, and real-estate & property. Each is a page of real, source-linked customer pains, and the whole library keeps getting denser as more Discovery sweeps run.

2026-07-18

Purchase receipts and two useful reminders

  • Buying credits now sends a proper receipt email — how many credits were added and your new balance, with a link to your billing history. And two emails that only show up when they're actually useful: if you run out of credits while actively working, we'll point you at a top-up once (not daily); and if your last Discovery sweep is more than three weeks old, we'll suggest a fresh one — the shared pain corpus keeps growing, so a new sweep sees more than your last one did. Every one of these respects the email toggle in Settings and carries a one-click unsubscribe.

2026-07-18

A welcome email, and control over the ones you get

  • New accounts now get a short welcome email that points you at your first founder first-pass. And there's a real control for it: Settings has an email-preferences toggle, and every email carries a one-click unsubscribe — turn off run-completion and product emails whenever you like, and we'll respect it everywhere.

2026-07-18

Pricing that names the offer and reverses your risk

  • The credit packs now have names, not just prices — the recommended one is the Founder Discovery Pass (a Discovery sweep plus deep-dives on the finalists it surfaces). And the guarantees that used to be scattered in the fine print are gathered into one plain block: you see the exact price before you spend it, a run that fails on our side refunds its credit automatically, credits never expire, unused packs are refundable within 14 days, and a weak niche honestly returns zero finalists — you keep the rationale, not a build you'd regret.

2026-07-18

A browsable pain library — real problems, by topic

  • There's now a Pain Library (in the footer, under Product) that groups the real, source-linked customer pains our discovery sweeps surface into topics — booking & scheduling, invoicing & payments, freelancer back-office, client management, content & creators, and e-commerce. Every pain links to the page it came from, and shows how often it's re-surfaced across sweeps. It's the same evidence behind Trending, organized so you can browse the problems in a space before you commit to one.

2026-07-18

A proper FAQ — the honest answers in one place

  • There's now a Frequently Asked Questions page (linked in the footer) that answers the questions people actually ask before trusting a verdict: why the engine kills so many ideas, where the ideas and evidence come from, whether it'll make you get on customer calls (it won't), how pricing works, and — plainly — that we don't claim accuracy, only honesty and reproducibility.

2026-07-18

Compare is now priced by how many ideas you rank

  • Comparing ideas costs 1 credit per 3 ideas, rounded up — 2–3 ideas is still 1 credit, 4–6 is 2, up to 10 is 4. Each idea gets a full check; those aren't charged separately on top. The exact cost shows on the Compare page and the home tile before you run it. Single-idea checks (Kill My Idea, Founder Fit, Pricing Power, and the rest) are unchanged at 1 credit.

2026-07-17

Final pre-launch audit — 18 fixes, mostly the sharp edges you'd hit first

  • The sign-up verification code cells were nearly invisible in dark mode — now they look like the rest of the form. The avatar menu's “Manage account” window matches the app theme too, instead of opening a white island.
  • Downloaded and copied reports (.md / .csv) now read like the on-screen report — internal check codes like “KF-04” come out as the plain question (“Can you reach buyers?”), including on reports you exported before this fix.
  • A multi-market sweep's “Start here” list now puts every green-lit idea above the “Not yet” ones — a sweep with two actionable ideas used to show only one at the top, because a raw score sort buried the second.
  • If a deeper-mode run fails on our side, its card now says so right where you clicked (credit already refunded, try again) — it used to quietly reset and look like nothing happened.
  • Compare is now linked from your ideas list when you have 2+ ideas, your founder first-pass is findable again (ideas list, founder-profile page, and the home tile now name it), and a declined Discovery launch shows its explanation next to the button, not three screens up.
  • Smaller: one recommended pack on pricing (“most founders start here”) with the never-expires promise right under the cards, the public benchmark table speaks the same verdict words as the product, and support@whittleos.com is tappable on the contact page.

2026-07-15

Credits and copy, cleaned up further

  • Free signup credits are for idea checks, not Discovery sweeps — pricing, sign-up, and the Discovery form now say so upfront, instead of only after a sweep gets declined.
  • A couple of report sections (“What would change the answer”, “What we couldn't tell”) still showed a stray leftover list marker in the text — fixed.

2026-07-15

A day of fixes from a full production audit

  • A failed run no longer replaces the whole page with “Something went wrong” — you get an inline note where you clicked: the run failed on our side, any spent credit was already refunded, try again.
  • Reports speak your language now. Internal codes that leaked into verdicts and Discovery finalists (KF-08, “Phase 6”, risk_platform_regulation, b3_c1) render as the plain names the rest of the product uses — including on reports you've already run.
  • Pain Receipts actually finds your niche: “wedding photographers” now returns the wedding-photographer pains (search matches every word against the pain, its theme, AND who feels it — it used to need the exact phrase).
  • Billing history rows link to their results; runs of deleted ideas say so instead of dead-ending; the idea library no longer scrolls sideways on phones; demand badges read “21/100 · low” instead of the cryptic “21 · cool”.
  • Honest-copy pass: the privacy policy names the AI provider we actually run on (Anthropic), stale numbers got corrected, and “free First-Pass” now says what makes it free — the 2 credits every new account starts with.

2026-07-15

Receipts, one click from your billing history

  • Every credit-pack purchase in Billing now has a “Receipt →” link that opens your Paddle invoice PDF — before this, getting a receipt meant emailing support. Fresh purchases can take a few minutes to have their invoice ready; the link says so instead of erroring.

2026-07-15

Edit and delete your ideas

  • “Edit idea” on every idea page — the same four questions, prefilled. The verdict told you to fix things and re-run; now there's somewhere to fix them. Existing results keep showing the old text's verdicts until you re-run a check.
  • “Delete idea” with an explicit confirm. Deleting removes the idea from your lists and graveyard; completed runs and reports stay in Runs and your billing history — credits aren't refunded, they paid for runs that really executed.

2026-07-15

You don't get stranded anymore

  • After buying credits, the “Payment received” banner now offers “Continue where you left off →” — straight back to the idea or sweep where you ran out. No context? It offers home instead.
  • If a run's button ever sticks on “Running…” while the work actually finished, the page now checks for the result by itself after a couple of minutes and offers a one-click refresh — and reminds you a failed run refunds its credit automatically. No more wondering whether to click again (don't — it never double-charges).

2026-07-14

One place for every run you've made

  • New /runs page: every analysis you've ever run, newest first, each linking straight to its result — no more remembering which idea a report lived on. Linked from Billing next to your transaction history.
  • After a run, the page address now reads like English (?ran=founder-fit) instead of an internal code (?ran=Mode_11) — old links keep working.
  • Idea lists now say “Dropped” on a dropped idea's chip instead of the imperative “Drop it” — a status you read, not a button you fear clicking. The verdict page itself still tells it to you straight.
  • “Runs” now lives in the header next to Ideas — one click from anywhere, desktop and mobile.
  • The flagship has one name everywhere now: Discovery. “Full pipeline” badges and “Start a full sweep” buttons were three names for the same thing — they all read Discovery / “Start a Discovery sweep” now, matching billing and pricing.

2026-07-14

Big results you can actually navigate

  • An idea with many analyses is no longer a 12-panel wall: completed results collapse into one row each, and a sticky chip bar jumps straight to any of them (the freshly-run one opens and highlights itself, as before).
  • A finished multi-market sweep now opens with “Start here” — the top 5 ideas ranked across ALL markets, each linking to its market — and the market list sorts best-first with a “best market” badge, instead of 29 identical rows.
  • The comparison result ends with “Do this next”: open the winner, close its unknowns, or — when nothing clears the bar — run a Discovery sweep or pivot the closest idea. No more dead end after the ranking.

2026-07-14

Honest buttons, visible prices, and a re-run that finally exists

  • The first-pass page now shows its credit price right at the run button — it said nothing before, which broke our own “you see the exact cost before every run” promise.
  • The new-idea form's button now says what it does: “Save idea →” (saving is free). It used to say “Run analysis” while only creating the idea — the actual analysis always started from the idea page.
  • You can re-run the Single Idea Check straight from the verdict card after fixing a red flag or updating your profile — the card told you to re-run but offered no button. The new verdict replaces the old one.
  • One label language everywhere: the comparison summary, the pricing page, and the home funnel no longer leak internal codes like KILL/VALIDATE or “Survivors” — you'll see the same “Build it / Test it first / Drop it” words the rest of the product uses.
  • The landing-page generator card no longer vanishes on ideas without a validation plan — it now says what unlocks it. Multi-market sweeps in your billing history are labeled as sweeps and link to their results. Failed first-passes say “credit refunded” right in the list.
  • Fixed: the Discovery page no longer scrolls sideways on phones when a recent run has a long niche name, and pure keyboard-mash ideas are rejected before they can spend a credit.
  • Smaller words, clearer choices: each credit pack now says what it's for (“Two idea checks — or your first Discovery niche”), home tiles say what the click does (“Check this idea →”, not “Open →”), the sweep price cap reads plainly (“capped at 15 credits total”), the graveyard chip says “Dropped”, and the home callout counts all 12 deeper modes.

2026-07-13

The site works properly on your phone now

  • The header no longer breaks on narrow screens — it collapses into a menu instead of clipping the logo and pushing your avatar off the edge. Reported from a Samsung A73; fixed for every phone.
  • An idea's verdict badge no longer disappears from your lists after you run many deeper analyses on it — /ideas, /compare, and the home list now always show the latest real verdict, the same one the graveyard shows.
  • Smaller fixes: wide comparison tables tell you they scroll sideways, the cookie bar stops covering the last line of a page, a typo'd URL shows a proper 404 page instead of bouncing you to sign-in, and source links only render when there's a real page behind them.

2026-07-13

A clearer next step after every verdict, and a tidier idea page

  • Each Single Idea Check verdict now surfaces a one-click next step — Pivot this idea, Plan a no-call validation, or Scope the MVP — matched to the decision, so you’re not left scrolling to figure out what to do.
  • The deeper analyses (founder fit, pricing power, platform risk, and the rest) are grouped under one “Go deeper” section in a two-column layout instead of a long single stack.

2026-06-22

Honest benchmark — we ran 8 ideas with known outcomes through our own gate

  • New public /benchmark page: 8 startup ideas whose real-world fate is known (5 documented flops, 3 successes), fed anonymized to our gate, with the verdict checked against history.
  • On a solo-founder profile we hard-killed 4 of the 5 historical flops the flattering category green-lights, and flagged the 5th; on a funded-team profile all 3 successes pass. Both profiles are shown, with every miss and limit stated.
  • No accuracy claim, no rocket emoji — a dated, reproducible snapshot of what the gate actually does. Each rival tool's own column on the same 8 ideas is being collected and will slot in.

2026-06-22

Sharper verdicts — HOLD when it's genuinely unclear, and resource-aware on regulation

  • When too many kill-filters come back UNKNOWN to actually judge an idea, we now HOLD with the open questions instead of defaulting to a premature KILL — absence of evidence isn't evidence of a bad idea.
  • For a funded, full-time team, a clearable cost-of-entry like PCI-DSS, SOC2, or money-transmission licensing is now treated as a costed risk that goes to scoring, not an instant kill — the verdict you'd get building it solo on nights and weekends is unchanged.
  • Safety-critical and prohibited domains (medical, regulated financial/legal advice, gambling, controlled goods, and the like) stay hard kills regardless of how well-funded you are.

2026-06-21

Pain Receipts — a free, no-signup lookup for real customer pains

  • Type a niche at /tools/pain-receipts and see the real customer pains our discovery sweeps have already mined from public discussions — each with a clickable source you can read yourself.
  • Every result is ranked by how often it re-surfaces and how severe it is, and carries an honest provenance label. Don’t trust the score — click the source.
  • If we haven’t sourced your niche yet, you get an honest “nothing yet” — we never invent a pain to fill the gap. No signup, no email gate.

2026-06-19

Turn your no-call validation plan into a real landing page

  • After you plan a no-call validation, generate a clean, shareable landing page from it — hero, problem/outcome, and your chosen CTA (waitlist, payment, pre-order, or paid pilot).
  • It’s grounded in your own validation promise — no fabricated metrics or fake social proof — hosted at a public link you can point traffic at, or download the HTML to self-host.
  • Point the CTA at your own waitlist/payment link to start collecting real signal.

2026-06-19

Find where your first customers are — real source-linked posts + ready async messages

  • On any idea, we scan public threads (Hacker News, forums, job boards) for real public posts where people describe the problem you solve.
  • Each lead carries a clickable source, a short grounded quote from that source, and a drafted async message you can review — a DM, a community reply, or a waitlist invite. No calls, ever.
  • We only surface leads we can ground in a real source; if the signal is thin, we say so instead of inventing names.

2026-06-19

Trending problems now ranked by demand

  • The public Trending problems page shows a demand score on each pain and ranks them by it — a blend of how often the pain re-surfaces across sweeps and the severity it was tagged at.
  • It's an honest relative-intensity signal, not an upvote count or a market percentage — every entry still links to its source.

2026-06-17

See the receipts — every verdict now shows its work

  • Discovery results show why candidates were rejected, the sources we scanned, a per-finalist scoring breakdown, and a recommended first move — not just a count.
  • Single Idea Check shows the full kill-filter table (every filter, pass or fail) and why a high score was downgraded; Compare shows a per-category matrix across all your ideas.
  • New public Trending problems page: real, source-linked customer pains ranked by how often they re-surface. The graveyard now suggests a pivot for ideas you killed.

2026-06-17

Verifiable sources on every pain signal

  • Discovery results now show an Evidence Base — the pain signals that shaped which candidates cleared the gate. Sourced entries carry a clickable link to where the pain was documented.
  • Pain Mining and Competitor Teardown panels show a 'source ↗' link on each grounded pain — the real fetched page or snippet, not a summary.
  • Each new Discovery run seeds a shared corpus that makes future sweeps smarter. The gate compounds across users and verticals.

2026-06-17

A darker, clearer redesign

  • The whole app is dark by default now — one consistent look across the landing, the home, and every page. Toggle to light anytime.
  • The home leads with Discover, the full-pipeline flagship: what you get, that it ranks candidates against your founder profile, and that every other mode runs on its finalists.
  • Credit costs now show up front in amber on every mode, so you always know what a run costs before you start it.

2026-06-15

Credits: pay per decision, no subscription

  • Buy prepaid credit packs on /pricing — $19 for 2, $49 for 6, $99 for 15, $149 for 30 credits. Credits don't expire.
  • 1 credit = one verdict with its reasons and your next move — Single Idea Check, Kill My Idea, Compare, Pain Mining, and the rest. Discovery is 2 credits for the first niche, +1 for each additional (capped at 15).
  • New accounts start with 2 free credits — enough to try it before paying. Your balance shows in the header, and you always see what a run costs before it runs.

2026-06-10

Public growth surface: sample report, free tools, idea library

  • A real, unedited Single Idea Check is now public at /sample-report — “Try a paid pilot”, 73/100, honest 5/10 evidence score included.
  • Free no-signup calculators at /tools: customers-needed (the ceil(target/price) reality check) and support-burden (hours/week with solo sustainability bands).
  • The idea library at /library: eight real finalist cards from one autonomous sweep — four worth testing, two on hold, two dropped.

2026-06-10

Discovery sources: scraper + official APIs join the search sweep

  • Named-revenue comps from TrustMRR, alternatives maps from SaaSHub, launch listings from BetaList — scraped politely (robots.txt respected) by our own fetcher.
  • Hacker News discussion threads (with points/comments) and Remotive job listings join as structured evidence via official public APIs.
  • All three source mechanisms compose; every discovery run now draws on more sources than a human curates by hand.

2026-06-10

Cross-run pain memory

  • Every discovery run now contributes its grounded public pain evidence to a shared corpus — and benefits from all prior runs' corpus at triage.
  • The kill-gate gets stricter and better-informed as the corpus grows. Nothing user-private is ever stored.

2026-06-09

Billing complete: plans, upgrades, cancellations

  • Founder $19 / Builder $49 / Pro $99 with per-tier mode access and quotas.
  • In-app plan switching (prorated), scheduled cancellation reflected on the billing page, tax-inclusive flat pricing.

2026-06-07

Discovery batches: sweep the whole catalog

  • Pick multiple seed verticals (or the whole catalog) and run them as one batch with live progress and a completion email.
  • Batches are reachable from /discover with per-batch finalists counts.

2026-06-04

Discovery quality: pain gate, deterministic scoring, first moves

  • Candidates without a real, sourced pain are killed by default — the gate that separates a decision engine from an idea generator.
  • Scores are computed deterministically in code from the 7-category breakdown (20/20/20/15/10/10/5 weights) — the model never emits a naked number.
  • Every discovery report opens with ONE recommended first move plus a stop-loss; the rest of the finalists are benched, not cheered.

2026-06-02

Full mode catalog: 15 analysis modes + the discovery flagship

  • Single Idea Check, Founder Fit, Kill My Idea, No-Call Validation, MVP Scope, Pricing Power, Support Burden, Platform Risk, Pivot Wedge, Pain Mining, Competitor Teardown, Compare, First-Pass, Idea Graveyard — all live.
  • Full Pipeline Discovery sweeps a niche end-to-end in ~6 minutes: sources → pains → kill-gate → scored finalists.

Try it: see a real report or sign up and use your 2 free credits on a real check.

Rather meet these in Google than in your inbox? You can add WhittleOS as a preferred source in Google. It asks Google to favour our pages where they are already relevant to something you searched for — it does not put them in front of you where they otherwise would not be. One click, undone in the same place.