Keep the superseded number in print, with the reason it changed
Our 15 October and 16 October 2026 runs are both published in full, and so are the two corrections that moved figures on them, because a research page that overwrites itself has destroyed its own evidence.
Published by AI Knows Us · Updated 19 October 2026
Both of our citation runs stay published. The run of 15 October 2026, with 102 questions per market, 597 usable readings of 612 attempted and 10,652 citations across 2,708 domains, sits beside the run of 16 October 2026, with 210 per market, 1,260 readings all usable and 20,143 citations across 4,947 domains. The older one has not been retired, folded in or quietly replaced. Keeping it there is what allows a reader to see whether a finding held, and a page that silently restates its own numbers has thrown away the only evidence it ever had of being reliable.
This is a rule about publishing rather than about measurement, and it is the cheaper half of being trustworthy.
The rule, in four lines
- Never overwrite a published figure. Add a dated row beside it.
- When a figure changes, print the cause. The cause is more useful to a reader than either version of the number.
- Keep the older run reachable. Not in an archive nobody links to, but beside the newer one, with its own date and its own denominator.
- Say which run a sentence is quoting, every time. Two runs in one paragraph without labels is how a reader ends up quoting a figure against the wrong denominator.
Two corrections of our own, both still on the page
The rule is worth nothing unless it survives a figure you would rather not print. Ours has been tested twice.
| What was published | What it became | The cause, printed with it |
|---|---|---|
| The vendor own site share as 91.3% | 85.6% | The classifier that produced it had a pattern requiring a literal string, so Google's and OpenAI's own documentation was being counted as vendor marketing. Source classification now runs through one tested file that refuses to load if any of 24 named examples sorts wrongly |
| Two proxy hostnames listed among the most cited domains | Both rows removed and resolved back to the pages they serve | Opening the cited addresses showed an iframe proxy and a reverse proxy carrying real documentation. They are an artefact of how the data was fetched, not a finding about any engine |
The second one moved the headline share by four tenths of a point, from 85.6% to 85.3% on the first run and from 82.4% to 82.0% on the second. The published headline stays as measured in both cases and the difference is recorded on the page, which is a smaller and more honest move than restating either figure.
Why the cause matters more than the correction
A reader who sees a number change learns almost nothing. A reader who sees why it changed learns something they can apply to the next number anybody shows them. The first correction above is a case in point: the defect was a pattern that could not match a whole class of hostname, and the general lesson is that a classifier which cannot fail on a known example will quietly sort a category wrongly for as long as nobody checks. That lesson transfers to anybody building their own source table.
It is also the reason a withdrawal can be worth more than the finding it replaces. Two hosts in our access run looked like per agent blocks and turned out to be rate limiters, one of them with an explicit HTTP 429 behind it. The page that carried those rows now carries the withdrawal and the reason instead, and of the 102 build spec conversations we ran on 15 October 2026, 68 asked for failures and negative results to be shown.
How to version a research page in practice
- One row per reading, never one number per subject. Date, apps, size, and the question that reading can settle.
- A correction note at the same prominence as the claim. Not a footnote if the original was a headline.
- A last verified line, and a stated interval for the next reading. On that one requirement the tally across our 102 build spec conversations was unanimous: a date, a line saying when it was last verified, and how often it gets re-checked.
- A sweep after every publication. New work makes old work wrong, and the sentence that goes stale is the unqualified present tense. A sentence scoped with a date or with the words in this dataset survives on its own.
The last point is the one that catches people out, including us. Publishing something new is the moment to ask what it just contradicted, and to fix the contradiction the same day.
The edges of both runs
Each read two apps, ChatGPT and Gemini, on one date, with the market written into the question wording. Our product covers five surfaces and three of them are absent from both. Neither bank holds a per industry rate, because no question named an industry. Anything at city level was read by a different scan covering ten named places. And both runs count a domain appearing in a source list rather than a business named in an answer's text.
We sell this measurement and the writing that follows it, which is exactly why the corrections are printed rather than absorbed: our own citation count went from 3 of 597 to 6 of 1,260, and both numbers stay on the site. The Google Search Console side of the same discipline, with both windows and the export date, is our proof page.
Questions about keeping an older study in print
Are both runs still published?
Yes. The 15 October 2026 run and the 16 October 2026 run are both on the site in full, each with its own date, size and denominators.
Why not merge them into one bigger dataset?
Because the banks are different. Merging would produce a total nobody could check against either original, and it would destroy the comparison that makes both worth reading.
What happens when one of your figures turns out to be wrong?
The corrected figure goes up, the old one stays visible, and the cause is printed beside it. Two such corrections are described above with their causes.
Does a correction not damage the rest of the page?
Somebody who already suspects the publisher of having an interest goes looking for the rows that hurt. Corrections printed in place are what that search finds, and a page with no history of ever having been wrong offers nothing to find.
Which engines and dates do the two runs cover?
ChatGPT and Gemini on 15 and 16 October 2026, across the United States, the United Kingdom and India. The earlier of the two sits at 597 readings across ChatGPT and Gemini in three markets.
The archive is the argument
Anybody can publish a figure. What distinguishes a research page from a brochure is whether the earlier figure is still there to be checked against, with the reason it moved written next to it. That costs a paragraph and it is the only part of a published measurement that gets more valuable the longer the programme runs.