Zombie stats: Where the numbers AI quotes come from
A · Strong evidence A study by Mack Grenfell of Trakkr, which sells AI-citation monitoring, but with a stated sampling method, model judging corrected by hand checks with uncertainty ranges, named trails and released aggregate data; vendor-run, yet checkable.
Summary based on the full text·Original: about 12 minutes to read
In short
A research page (28 October 2026) by Mack Grenfell at Trakkr. The team asked ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews 200 questions that call for a number, then followed each cited statistic back to its origin. Fewer than half reached a study that says the same thing. The assistants rarely invented numbers; the cited page held the figure 82% of the time. The problem is upstream: most cited pages are blogs and roundups repeating figures with no traceable source, and agreement between assistants was no sign of accuracy.
Key numbers
- 40%traced AI-quoted statistics that reached a study saying the same thing (95% range 34% to 46%)
- 1,285cited statistics followed up to five pages deep, from 1,000 answers by five assistants
- 21%share of pages cited for a number that were the original research
- 11 of 29statistics quoted by three or more assistants that traced to a matching study
- 23%sourced share found by the AI judge alone, before hand checks corrected it
As reported by the source.
English·Web page·opens trakkr.ai
What it reports
- Of 1,285 traced statistics: 40% sourced, 8% reached the study but the number or context had changed, 40% could not be checked (locked, gone or silent), 11% led nowhere.
- ChatGPT sourced 57% and quoted 1.3 statistics per answer; Claude 38% at 6.8 per answer. Perplexity, AI Overviews, Gemini and Claude were too close to rank.
- Company blogs made up 54% of first citations and reached a study 37% of the time; research firm and government pages about 61-62%.
- Of statistics that reached their study, 17% had changed on the way; only about a quarter of changes were the digits, the rest were context such as year, population or what was measured.
- On 28 of 71 questions where all five gave a percentage, there were four or five different figures; one question drew answers from 49% to 96%.
- Four statistics quoted by all five assistants, including "44% of salespeople give up after one follow-up", traced to no study at all.
How they got it
Questions were asked once, on 25 October 2026, through each vendor's API with web search on and a US location: 160 everyday questions in 16 topics plus 40 suspected famous statistics. Random samples of cited statistics per assistant were traced up to five pages deep. An AI model judged every trail and a stronger model re-judged 300; 150 random trails were checked by hand, and all rates are corrected using those checks. A cross-check on 300 statistics from AI Overviews answers to real searches also found 40% sourced.
Limits
- API answers can differ from the consumer apps; one day of answers.
- Paywalled originals count as uncheckable, not missing.
- The methods section gives both 150 and 180 as the number of hand-checked trails.
- The vendor sells a product tied to the finding.
Why it matters for companies
Staff increasingly lift statistics from AI answers into decks, business cases and published content. This study shows that a citation tells you where the assistant found a number, not where the number came from, and that the most repeated figures are no better sourced. Any number headed for a decision or a public document needs a trace to the original study, with sample, date and population checked. Firms that publish research can raise their own citation quality by keeping methods and headline findings on open pages.
Summary based on the full text. Read the original for the full argument and data.Read the original on trakkr.ai