An AI-search statistic is safe to repeat only with its population, its window and its unit attached. Seven figures that circulate in GEO, SEO and PR content fail that test when copied without them. Each traces to a study that says something narrower than its retellings. For one, "2.8x more likely to be cited in ChatGPT on four or more platforms", we found no primary study.
We checked them because we had repeated them. Most appeared on pages we publish across AuthorityTech, Machine Relations Research and our founders' sites, and several went beyond their sources. On 5 and 6 November 2026 we re-read each source and corrected those pages. This register records what each primary source says so the next writer does not have to rediscover it. Its companion, the source register for earned-media shares, does the same for the 25%, 84% and 85.5% figures.
The register #
| Figure as repeated | What the primary study measured | What it does not say |
|---|---|---|
| "AI traffic is up 527%" | Previsible's 2025 AI Traffic Report, as reported by Search Engine Land: AI-referred sessions across 19 GA4 properties went from 17,076 to 107,100 between January and May 2025. | Not 2026, not search queries, not AI search at large, and a small base of 19 sites. The article states the window two ways, as a rise "between January and May 2025" and as a year-over-year comparison of January to May 2025 with the same months of 2024, so quote the window in its words. |
| "87.4% of AI referral traffic is ChatGPT" | Conductor's 2026 AEO / GEO Benchmarks Report and Search Engine Land's summary: an average across 10 industries, from 1,215 Conductor enterprise customer domains, U.S. traffic, monthly averages May to September 2025. The report's wider citation analysis covers 13,770 domains; 1,215 is the stated base for its traffic findings. | Not global, not current, not a share of AI usage, and not the same thing as the channel's size: in that report AI referrals were about 1% of web visits. |
| "Product-style pages earn 76% of citations, blogs 24%" | Deepak Gupta's GEO Measurement Study: 240 pages on his own four sites, 200 prompts, six engines, 13 weeks from February 2026, 50,431 citations. Blogs were about 40% of the tracked corpus. He calls it first-party observation on one corpus and reads his lifts as "direction, not effect sizes". | Not a cross-site benchmark. A retelling that adds "clean H1-H2-H3 pages are 2.8x more likely to be cited" is not in the study. |
| "2.8x more likely to be cited in ChatGPT on four or more platforms" | No primary study found. A vendor report states it twice without attributing it to any study or giving a method for it, inside a report that claims a 680-million-citation dataset overall; a PR agency report states it against single-platform brands with no dataset we could find; another page credits "Clearscope research cited by Evertune" and we could not find a Clearscope study. | Not a threshold, not a finding of any named study. Say that vendors repeat it and no method is published. |
| "37% of consumers start searches with AI" | An Eight Oh Two survey of 500 consumers who already use AI tools, collected November 2025, as reported by Search Engine Land. | Not consumers in general, and not 37% of searches. It is a share of AI users in one agency survey. |
| "Trust in AI search fell from 82% to 54%" | Fractl's AI statistics survey, Q2 2026, 1,008 U.S. consumers and 150 marketers: the share who found AI more helpful than traditional search fell from 82% to 54%. Fractl and Search Engine Land headline this as falling trust. | The question asked was about helpfulness, not trust. A separate trust question in the same survey: the share saying heavy AI use by a favorite brand would lower their trust rose from 20% to about 40% (39% in Fractl's Search Engine Land article). |
| "94% of B2B buyers use generative AI" | Forrester's Buyers' Journey Survey 2025, reported in its blog on zero-click buying: 94% of business buyers use AI in the buying process, up five points from 89%. | A January press release carries a different 94% about buyers in groups of six or more reporting benefits. The two should never be swapped. |
The Forrester figure was already worded correctly on most of our pages.
Four ways a number loses its denominator #
It loses its window. Previsible's 527% covers January to May 2025 across 19 sites, and its own article describes it both as a five-month rise and as a year-over-year comparison with early 2024. Retold as growth "in 2026", it describes nothing that was measured. When a source states its window two ways, quote the source's words and no others.
It loses its population. Conductor's 87.4% describes 1,215 enterprise domains in the United States. Eight Oh Two's 37% describes 500 people who already use AI tools. Gupta's 76% describes four sites one author runs. Each is a fair answer to a narrow question and a wrong answer to a broad one.
It loses its measure. Fractl asked whether AI search was more helpful. Its own headline and Search Engine Land's call the result falling trust, and pages repeating it say "trust" without saying what was asked. The measure is perceived helpfulness. A reader who needs a trust figure has to use the separate brand-trust question.
It is borrowed, or it never had a source. A write-up that summarizes a study can add a figure the study does not contain, and the next page credits the study for it. The 2.8x figure shows the end of that path: repeated by several vendors, each unsourced, with no study at the bottom.
How to cite a number so it holds #
Name five things in the sentence that carries it: who measured it, who or what was measured, over what window, what the unit is, and how large the base was. If you cannot name them, link the primary and describe the figure as reported by the page you read. If the primary cannot be found, say so in those words.
It is the rule we hold our own pages to. A related instrument, Answer-Source Fidelity, tests the other end of the chain: whether the page an AI answer cites supports the claim the answer makes.
What we could not establish #
We did not obtain Eight Oh Two's report, which requires registration; its landing page says only that it surveyed 500 AI users. Fractl says its 2026 survey repeats its 2025 questions; we did not find the 2025 questionnaire or sample, so we could not check the wording or base. Search Engine Land's reports are our route to the Previsible and Eight Oh Two studies. A study may carry caveats its summary omits, and the primary pages are linked so a reader can check. This register records what the sources say. It makes no claim about whether any of these figures, or this page, affects what AI engines cite.