A reading without a date cannot be checked against today

A date on a reading is not provenance, it is an instruction: run this again and see whether it still holds. The blind scan published on our proof page carries one for that reason. It was run on 2026-10-16 on ChatGPT and Perplexity, against clawlaw.in, and 6 of the 10 answers named the company. Anybody can put those questions to the same two apps this afternoon and compare. Strip the date out and the same figure becomes unfalsifiable: there is no moment it describes, so there is nothing to disagree with and nothing to check.

What the date makes possible

A dated reading does three things an undated one cannot. It fixes a comparison point, so a second reading has something to be a second reading of. It lets a reader decide whether the result is current enough for their purpose. And it puts the author at risk, because somebody can re-run the questions and publish a different answer.

That last one is the point. A figure nobody can contradict is not evidence, it is a decoration. All 102 build spec conversations run on ChatGPT and Gemini on 15 October 2026 asked for a date, a last verified line and a refresh frequency. It is the only requirement in that specification that every single conversation raised, and it is also the cheapest to supply.

The scan that shows why the date matters

Two of the ten questions from the 2026-10-16 scan are published in full, and they are published as a pair because separating them would misrepresent the day.

Question asked blind on 2026-10-16	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, with Manupatra and SCC Online named instead, 8 sites read

One question won and one question lost, on the same domain, on the same day, on both apps. The questions never contain the company name, which is what makes the result a fact about the company rather than about the question.

Now notice what the date is carrying. Those two rows are a statement about 16 October 2026. They are not a statement about the following month, and the page does not make one. Whether the lost question is still lost is a question somebody answers by asking it again, which is the only honest way that question gets answered.

How fast a reading can go stale, measured

This is not a theoretical caution, and we have paid for the lesson. On 16 October 2026 we published four findings about which servers refuse which AI crawlers. Two of them did not survive a second reading taken hours later the same day: what looked like a block turned out to be a rate limiter, and one host returned an explicit HTTP 429 to make the point. Both findings are off the table with the reason on the page.

Hours, not months. That is the interval over which one of our own readings reversed, and the only reason it was caught is that the first reading had a date and a method attached to it, so a second reading was a thing that could be run.

The same applies to anything restated from somebody else's documentation. On 16 October 2026 we read OpenAI's own crawler list and it names four identities, GPTBot, OAI-SearchBot, ChatGPT-User and OAI-AdsBot. Two of our live pages had said three, and one of them asserted it as a count, which had quietly become false. A restated list without a read date becomes a wrong claim without anybody editing it.

The three fields that make a reading re-runnable

• The date, as a date. Not this month, not recently. A reader has to be able to work out how old the figure is without knowing when the page was written.
• The apps, by name. 98 of 102 build spec conversations asked for the engines enumerated rather than the phrase AI. The blind scan above is ChatGPT and Perplexity, and that pairing is part of the result rather than a detail about it.
• The questions, in the form they were asked. A reading is only reproducible if the wording travels with it, and the wording has to be free of the brand name. Naming your own brand in the question returns a flattering answer and measures the question rather than the company.

Setting a re-read interval you will actually keep

A refresh frequency is a commitment and it should be sized to be kept rather than to impress. Monthly is enough for most businesses and it is enough to tell a real change from the ordinary movement between two readings. What matters more than the interval is that the previous reading stays on the page when the new one arrives, with both dates, because a figure that is quietly replaced takes its own history with it.

And resist the urge to turn two readings into a trend. Two points with two dates are two points, and the honest sentence names both. For the arithmetic of reporting counts over stated denominators rather than a single rate, what a good mention rate is without a fake benchmark works through the version that survives being checked.

We sell this reading as a service, taken on a stated schedule, and the date is on every number we publish about ourselves for the same reason it is on the ones we publish for a customer.

Questions about dating an AI visibility reading

How often do these answers actually change?

Enough that two of our own four access findings reversed within hours on 16 October 2026. For named company answers the useful assumption is that a reading describes its own day, and anything longer is something you establish with a second reading rather than assume from the first.

Is a reading from last month still useful?

It is useful as a comparison point and it is not useful as a current claim. Keep it, date it, and run the questions again before putting the figure in front of somebody making a decision.

Is a ten question scan too small to publish?

Not if the ten is printed beside the result, which is the whole requirement. 80 of 102 build spec conversations asked for the sample size N to be stated, and stating a small N is a different act from claiming a large one. What would be too small is a count the page declines to give.

Why publish the question that lost?

Because the won question is only believable beside it. 68 of 102 build spec conversations asked for failures and negative results, and a pair showing one win and one loss on the same day is more informative than either row alone. On this page they are never quoted apart.

Does a date protect a reading from being wrong?

It protects it from being unanswerable. A dated reading that later turns out wrong gets corrected with the correction dated too, which is what happened to the two findings we withdrew. An undated one just sits there.