One question won, one question lost, same day, same two apps

A result that reports a win beside a loss, on the same domain, on the same day, on the same apps, is the only shape of AI visibility evidence a careful buyer has reason to believe. The scan published on our proof page is that shape. The domain under test is clawlaw.in. It sells legal search, and our founder is behind it, which is a commercial interest worth stating before any of the figures. The date was 2026-10-16, the apps were ChatGPT and Perplexity, and the five questions behind it read as a buyer would write them, with nothing in any of them identifying the company being looked for. 6 of the 10 answers named it. Two of the five questions are printed in full, and they are the two that disagree with each other.

The pair, as it came back

Question asked blind	ChatGPT	Perplexity
Pricing options for a solo practitioner advocate in India using legal case management software	Named first, alongside CaseNiti, LawSathi, LegalPrime and Lexentia, 5 sites read	Named first, alongside Nowlez, Legodesk and Peshi, 5 sites read
Which platform offers the most comprehensive court record search across all Indian courts and tribunals	Not named, and no business named at all, 6 sites read	Not named. Manupatra and SCC Online named, 8 sites read

Same company, same day, same two apps, opposite outcomes. Neither result is noise and neither cancels the other. They are answers to two different questions, and the gap between them is the finding.

Why the pair is more persuasive than the win alone

Publish the first row on its own and a reader has to take the method on trust, because there is no way to tell a real reading from a question written to produce that answer. Publish both rows and the method becomes visible: whatever the questions were, they were not all written to win, because one of them did not. A loss in the same table is the cheapest available proof that the set was not curated.

It also stops the one claim a single row invites. A reader who sees a company named first on one question will reasonably ask whether that holds across the category. The second row answers that directly, which is a better outcome than leaving the question open or implying a breadth nobody measured.

There is a practical test for whether a published scan was curated, and it works on somebody else's as well as ours. Count the rows and count the losses. A set with no losing row in it is either a set of questions chosen after the answers were seen, or a set so small that it cannot have covered the category. Both of those are worth knowing before reading the figures, and both are visible from the table without asking anybody a question.

What each row is actually evidence of

• The first row is evidence that on a pricing question, on that date, both apps reached for sources that named this company before any other. It is evidence about one question.
• The second row is evidence that on a coverage question, on the same date, neither app named it. That is a different kind of gap, because a comprehensiveness claim is something a product either publishes or does not.
• The pair together is evidence about the method: blind questions, both apps, one date, every answer recorded whichever way it went.

What none of it is evidence for is a rate. Five questions give five answers, and a count of 6 out of 10 travels with its denominator or it means nothing. We report it as a count and never as a percentage for that reason.

Keep the dataset separate from our other readings

This scan is ChatGPT and Perplexity, on 2026-10-16. Our wider audits are ChatGPT and Gemini, 597 usable readings on 15 October 2026 and 1,260 readings on 16 October 2026. Those are different apps, different question sets and different sizes, and adding any of them together would produce a figure describing nothing. Our product reads five surfaces. Every dataset named on this page covers two, and which two changes between them.

How to produce the same pair on your own category

Write five questions a buyer would type, with your company name in none of them, and make sure at least two of them are ones you expect to lose. Run each in a fresh conversation on two apps, record every answer including the names that came back instead of yours, and keep the sources each answer leaned on. The answers that did not include you are the useful half, because the companies named instead are a shortlist somebody else's buyers are being handed. Keeping that as a dated log rather than a one off verdict is the discipline, and keeping the log rather than the verdict sets out the five fields that make one worth maintaining.

We sell this measurement and the page writing that follows it, so the scan behind our own published pair is the product doing its own job in public rather than a demonstration built for the occasion.

Questions about a blind scan that reports a win and a loss

Why not publish only the question you won?

Because a single winning row is indistinguishable from a question written to win. The loss is what makes the win readable, and it costs one table row.

Does losing one question mean the product is weak there?

It means no source either app reached for named it on that question on that date. That points at what is published about the category rather than at the product, and it is a different problem from being absent everywhere.

Is 6 of 10 a good result?

It is a count from five questions on two apps on one date, which is a probe rather than a rate. The useful comparison is the same five questions on the same two apps next month, not somebody else's percentage.

Why two apps rather than one?

Because one app can be unrepresentative of the others, and agreement between two independently built products is a stronger signal than either on its own. It is still only a signal about the questions asked.

Can I see the other three questions?

The scan page for that domain holds all five with the answers behind them, which is where the remaining three sit. The two in the table are printed because they are the pair that disagree.

The row most suppliers would have deleted

Almost any version of this scan could have been published as a single strong row, and almost every version of it in this trade is. The decision to print the second row alongside the first was taken before the questions were run, which is the only point at which that decision is honest. Deciding afterwards which rows to publish is a selection, however good the figures turn out to be, and a reader who cannot see the losing question has no way of knowing a selection took place.