Study plan, registered on 14 October 2026. This study has not been run yet. I’m publishing the method before any results exist so it can’t be adjusted to fit them. Results will be added to this page, with the date they were collected.
Most citation audits have to trust someone else’s source. This one doesn’t need to, because every finding it checks came from data I collected and published with a CSV. If an AI engine states one of these findings, I know what the right answer is, when it was measured and where it was first published. The study records whether engines repeat the findings, whether they get the numbers right, whether they say when the finding was measured, and whether they credit it.
It’s a different study from do the sources AI engines cite contain the figure?, which checks figures from other people’s sources.
The findings (frozen)
These are nine findings from studies on this site, each stated exactly as published. A finding published after this plan goes into a separate, dated list and is never added to this one.
| ID | Finding, as published | Study | Measured |
|---|---|---|---|
| F1 | Nine of the UK’s ten biggest newsbrands block at least one AI crawler | Crawler study | 2 Oct 2026 |
| F2 | The Telegraph blocks 14 of the 15 AI crawlers I checked but explicitly allows Google-Extended | Crawler study | 2 Oct 2026 |
| F3 | Only 5 of 10 block Google-Extended | Crawler study | 2 Oct 2026 |
| F4 | A site that says nothing about AI has voted for whatever its platform shipped (“the default vote”) | The default vote | 7 Oct 2026 |
| F5 | Both AI engines that answered gave £274bn, and neither said it was a 2020 figure | Purple pound | 9 Oct 2026 |
| F6 | The House of Commons Library briefing that Nationwide footnotes for £446bn doesn’t contain that figure, or any purple pound total | Purple pound | 9 Oct 2026 |
| F7 | Three of the four finance-director answers recommended none of the products the no-persona answers named | Personas | 9 Oct 2026 |
| F8 | ChatGPT ended all six of its answers with a question about my business | Personas | 9 Oct 2026 |
| F9 | 7 of 40 AI answers to “current” UK figure questions didn’t say when the figure was from. All 7 came from Google’s two engines | Undated figures | 10 Oct 2026 |
Questions (frozen)
Each question is one the finding would answer. None names me, this site or the study.
- How many of the UK’s biggest news sites block AI crawlers?
- Does the Telegraph block AI crawlers in its robots.txt?
- How many major UK news sites block Google-Extended?
- What happens if a small business website’s robots.txt doesn’t mention AI crawlers?
- How much is the UK purple pound worth?
- Is Nationwide’s £446bn purple pound figure supported by the House of Commons Library?
- Does ChatGPT recommend different accounting software to a finance director than to someone who doesn’t say who they are?
- Does ChatGPT ask follow-up questions when you ask it to recommend business software?
- Do AI search engines say when a “current” UK statistic is from?
Engines and sessions
The same four engines and settings as the 10 October run: ChatGPT (default model, signed in, temporary chat, unpersonalised), Google AI Mode (signed in), Claude (web search on, incognito chat) and Gemini (signed in, temporary chat). Each question gets a new chat, pasted exactly, with no follow-up, run from the UK. Every question runs twice on every engine: 9 × 4 × 2 = 72 answers per audit. Perplexity isn’t run, for the same reason as on 10 October: it asked for a sign-up.
How each answer is scored
- Finding present: full (states the finding with its number or named term), partial (states the gist without the number, or with a different one) or absent.
- Number matches: yes or no, against the finding as published. “n/a” if absent.
- Date stated: does the answer say when the finding was measured (any of the day, month or year)?
- Credit: linked (a clickable link to jacobwhitmore.com), named (Jacob Whitmore or jacobwhitmore.com named without a link) or none.
- Other source credited: any other domain the answer links or names for that finding.
From those columns each answer gets one outcome: credited and correct, credited and wrong, uncredited and correct, uncredited and wrong, or absent.
Exact-phrase search (same day)
Each finding is searched in quotation marks on Google and Bing, signed out, and every result outside jacobwhitmore.com is logged with its URL and whether it credits the source. Before trusting a zero result, two control searches check that the engine is honouring quote marks: one phrase known to be on this site, which should return it, and one made-up string, which should return nothing. If either control fails, that day’s results for that search engine are marked unreliable instead of zero.
What I will not claim
- That an absent finding means an engine hasn’t seen my pages. Answers vary by day and wording, and one audit is one snapshot.
- That an uncredited, correct answer took the finding from me. Someone could measure the same thing independently. It’s flagged only when the wording matches or no other source for the figure can be found.
- Anything about traffic. Being credited in an answer isn’t a click.
- Anything beyond these nine findings, four engines and stated dates.
Schedule
- Registered: 14 October 2026.
- First audit: 2 January 2027, three months after the first findings were published.
- Then every quarter: 2 April, 2 July and 2 October.
Calls resting on this study
P21 and P22 on the predictions page.
Data
Two CSVs per audit: findings-recall-audit-YYYY-MM-DD.csv (72 answer rows) and findings-phrase-search-YYYY-MM-DD.csv (18 search rows plus 4 control rows).