AI Assistants Cited Pages With the Number. Traced Back, About Four in Ten Reached a Study Saying the Same

Ask an AI assistant what share of salespeople give up after a single follow-up and you are likely to hear 44 per cent. On 25 October all five assistants in a new study gave that figure, and the cited pages back it up, in the sense that they contain it. Followed backwards, the trail goes to a sales firm’s statistics page from 2020 whose “source” link opens a freelance writer’s sample article, online by 2017. That article lists two slide decks, a blog list and two news stories; according to the study’s authors, none of them is a study of follow-ups, and they found no study behind the number.

The study, “Zombie stats”, was published on 28 October by Mack Grenfell for Trakkr Research. Trakkr is not a neutral observer: it sells tools and services to help brands get cited by AI assistants, and the study ends on the pages that assistants choose to cite. With that in mind, the method is laid out in detail. The authors asked each assistant 200 questions that call for a number, word for word, through its company’s API with web search on and a US location: 160 everyday questions across 16 topics, plus 40 famous statistics they already suspected, reported separately. All the answers were collected on one day. The answers to the everyday questions contained 3,167 statistics, 2,419 of them with a citation; the authors followed a random sample of 1,285 cited statistics up to five pages deep.

The verdicts were largely written by AI. A model judged every trail, a stronger model re-judged 300 of them and agreed 82 per cent of the time, and people re-checked 150 trails at random, with a third check wherever the two disagreed; in all, 180 trails were checked by hand. People found originals the model had missed. By the model alone, 23 per cent of the statistics were sourced; the rates the authors report are corrected with the hand checks. Corrected, 40 per cent reached a study that says the same thing (95 per cent range 34 to 46), 8 per cent reached a study that says something different, 40 per cent could not be checked because the page was locked, gone or silent on the number, 11 per cent led nowhere at all, and 1 per cent of the pages would not load. “Could not be checked” does not mean false.

The finding that matters most is not about invention. The page an assistant cited contained the figure 82 per cent of the time. But only 21 per cent of the pages cited for a number were the original research. Company blogs were 54 per cent of first citations, and their trails reached a study 37 per cent of the time, against 62 per cent for research firms and surveys and 61 per cent for government and official bodies. The authors put it in one line: a citation tells you where the AI found a number, not where the number came from.

Reaching the study did not settle it either. Of the statistics that got there, 17 per cent had changed on the way, and the digits themselves moved only about one time in four. More often what changed was what had been measured (20 per cent of changes), the year (17), the source (16), who was asked (14) or the place (4). The best example is “94% of first impressions are design-related”, which Google AI Overviews stated as a fact about every website. It goes back to a paper presented at the CHI conference in 2004 by Sillence, Briggs, Fishwick and Harris: 15 women aged 41 to 60 looked for information on the menopause in four weekly two-hour sessions at an internet café in Newcastle, and 94 per cent of their comments about rejecting or mistrusting health sites concerned design. A 2018 press release made it what people notice first; a 2024 magazine column made it 94 per cent of all first impressions. Another number is older still. “80% of sales need five follow-ups” is usually credited to a sales association; in 2021 its successor, Sales & Marketing Executives International, wrote that the figures come from a 1942 survey of its Long Island chapter, which asked members about calls made and sales closed, with a sample of fewer than 40, and that the original claim was 80 per cent of sales on the fifth to twelfth contact.

Agreement between assistants was no guarantee. Comparing headline figures on 71 questions, the authors found four or five different answers on 28 and a spread of 20 points or more on 31: asked what percentage of marketers use marketing automation, the assistants said everything from 49 to 96 per cent. On the 10 questions where all five gave the same figure, 2 of the 5 that were traced were sourced; of the 29 statistics quoted by three or more assistants, 11 were. ChatGPT quoted the fewest numbers, 1.3 per answer, and sourced the most, 57 per cent; Claude quoted 6.8 per answer. The other four landed between 38 and 46 per cent, which the authors call too close to rank.

This is one company’s study, not peer reviewed, run on one day through APIs rather than the apps people use, from a US location, with trails stopped at five pages; the per-answer data is not published, only aggregates and worked examples, and the trails have not been re-traced here. Nor does it show that repeated numbers are wrong: on the famous claim that 93 per cent of communication is non-verbal, all five assistants flagged the limits of Albert Mehrabian’s 1967 experiments, and AI Overviews called it a myth. What it describes is quieter. For the most part the assistants did not make the numbers up. They inherited them from pages that had copied them from other pages, and the citation, real and checkable, made the copy look like a source. The number survived; the room where it was measured, with its 15 women or its fewer than 40 members of one chapter, did not.