Checked 10 November 2026.
On 10 November 2026 every headline figure in the 12 studies below was recomputed from the published data files. All of them came out the same. No figure on the site was changed.
The strongest check is on the UK news series. A second robots.txt parser, written separately from the one the study used, classified all 150 cells from the raw files saved on 2 November 2026. It agreed with the published file on 150 of 150.
What was checked
| Study | Figures recomputed | Recomputed from | Result |
|---|---|---|---|
| UK news robots.txt, November 2026 | All 150 cells. 9 of 10 sites block at least one. Training 67 of 90, retrieval 32 of 60. Per-site and per-crawler totals. 0 cells changed since October. 191 other tokens, 320 pairs | The ten raw robots.txt files | Match |
| UK news sites and AI crawlers | Every per-crawler count, the ten site totals, the four sites that block ClaudeBot but not Claude-SearchBot | October CSV against November raw files | Match |
| AI companies’ robots.txt | 3 of 17 hosts block at least one crawler. 255 cells. 3 of 6 product sites, 0 of 11 corporate. Blocked and named counts for chatgpt.com, grok.com and claude.ai | The published CSV | Match |
| AI search statistics, traced | 32 statistics, 28 traced, 0 pass five checks, 8 pass four, 1 with raw data, market split 10, 9 and 9, 11 exact, 16 mixed, 1 changed, 4 superseded, median age 15 months | The published CSV, with each score recounted from its five check columns | Match |
| Undated “current” figures | 40 counted answers, 7 undated, 4 from Google AI Mode and 3 from Gemini | The coded answers CSV | Match |
| P37, Bank Rate catch-up | 0 of 8 stale at 24 hours, 6 named the decision, 2 unreferenced. Google AI Mode: 5 of 6 answers from T+3h unreferenced | The coded answers CSV | Match |
| AI visibility trackers | 13 tools. Counts for all nine questions. Best score 5 of 9, reached by 6 tools | The published CSV, with each score recounted | Match |
| Backlinks and AI citations | 16 studies, none with raw data, each score recounted | The published CSV | Match |
| llms.txt census | 5 of 33 readable sites, 1 of 10 news, 4 of 23 universities, 1 gap | The published CSV | Match |
| Personas | 12 counted answers. 3 of 4 finance-director answers share no product with the no-persona answers. ChatGPT asked a question in 6 of 6 | The coded answers CSV | Match |
| Purple pound | Both engines that answered gave £274bn with no year | The coded answers CSV | Match |
| Prompt runs | All 31 rows of the lookup table, from the formulas on the page | The formulas | Match |
What this does and doesn’t show
- It shows the published numbers follow from the published files, and that the robots.txt classification holds up under a second parser.
- It doesn’t re-check how an AI answer was coded. Where a study rests on reading answers by hand, such as whether an answer gave a date, the check starts from the coded file, not from the original screenshots.
- It doesn’t re-collect anything. Sites and AI engines change, so a fresh run today could differ. That’s what the dated snapshots are for.
Run it yourself
The scripts are plain Python with no dependencies beyond the standard library: reproduce-2026-11-10.zip. Download the CSV files from the data page into one folder and point the scripts at it. If you get a different number from the one on a study page, email contact@jacobwhitmore.com and I’ll log the correction.