Every number on this page comes from measurements we ran and published ourselves, with the method public and the source linked. Nothing here is aggregated from other people's surveys. Reuse any of it under CC BY 4.0 with a link to the source page. Dates matter: each statistic carries the date it was measured.
Answer stability
- 19% of same-question, same-engine repeated runs change whether a business gets recommended (1,211 classified answers, July 2026). Source: the Answer Volatility Reading, CSV sample included.
- 29% of questions get a different majority verdict on a different engine. Same source and dataset.
- The instrument behind these numbers disagrees with itself on identical text about 1% of the time, and gave the same verdict on the same answer 24 hours apart in 195 of 196 cases (99.5%). Source: the Instrument Reading, which notes that identical-text testing is a lower bound on the instrument's total contribution; even so, 1% does not explain 19%.
Concentration by trade
- Across 5 UK trades and 502 measured answers (Q3 2026), each engine's answers return to a short list of names while many businesses in each trade never appear in the sampled answers at all. Source: the Aeova AEO Index, per-trade boards included.
- In boiler replacement, one company (BOXT) led on all three engines: named in 44% of ChatGPT Search answers, 60% on Google AI Overview and 50% on Perplexity (102 measured answers, July 2026).
- In 2 of 5 measured trades the three engines crowned the same leader; in the other three they disagreed about who leads.
What the engines cite
- In one measured UK service market, review platforms were 56 of 4,649 cited sources, about 1% (July 2026), despite reviews being central to how buyers actually choose.
Method, in one paragraph
Real customer questions, asked repeatedly across ChatGPT search, Google AI Overviews and Perplexity, every answer captured and classified (recommended, mentioned, cited only, negative, absent), with the classifier's own error rate measured and published. The full method, worked through on real data: how we measure. New readings add to these numbers as they publish; this page updates with them, and the corrections policy applies.
Common questions
How reliable are AI search answers?
Measurably unstable: in our published dataset of 1,211 classified answers, 19% of identical same-engine repeats changed whether a business was recommended, and the engines disagreed with each other on 29% of questions. Any single answer is one draw from a distribution, which is why serious measurement repeats every question. Source: the Aeova Answer Volatility Reading, aeovapartners.com/answer-volatility/.
What share of businesses appear in AI answers at all?
In the five UK trades we have measured (502 answers, Q3 2026), each engine's answers keep returning to a short list of names while many businesses in each trade never appear in the sampled answers at all. Concentration varies by trade: in boiler replacement one company led on all three engines; in heat pumps the three engines could not agree on a leader. Source: the Aeova AEO Index, aeovapartners.com/aeo-index/.
Can I reuse these statistics?
Yes. Everything on this page is our own measurement, published under CC BY 4.0: reuse any number with a link to the source page. Each statistic links to the reading it comes from, with the method public.
Updated July 2026