Avouch research: what AI recommends, and why.

The studies

1. What actually decides who AI recommends

   In August 2026 we judged 6,517 AI answers from two service markets, engine by engine. On the first reading, comparison and ranking sites decided who got picked more often than the press or YouTube on all seven engines. They also beat Reddit on the five engines where Reddit was cited enough to compare. A second AI model then tried to overturn each verdict that said a source decided the pick, and comparison sites still led on four of the seven, two of them (ChatGPT and Claude) by four upheld verdicts or fewer. On Google AI Overviews, YouTube was cited in 1,216 answers and decided none of the picks.

   Published 27 July 2026, corrected 30 July and 4 August 2026 · 6,517 answers · 7 engines · a second AI model checked every verdict that said a source decided the pick

Avouch is a UK AI SEO agency. Our studies shape what we do for the companies we work with. To see how, read about our AI SEO agency work.

Studies being re-checked

We are re-checking these against the raw data. Where a figure turns out wrong, we correct it on the study and say so, as we did on the solicitors and accountants studies on 5 October 2026.

1. Which AI engines actually recommend anyone, 34,411 answers classified

   Whether each of seven AI engines names one company as its pick, or only lists options, read engine by engine.

   Published 6 August 2026 · 34,411 answers · 7 engines · one AI model classified the answers, with no second reading yet

2. What ChatGPT and Google AI actually say about solicitors, our own sweep data

   Who three AI engines named, and who they recommended, in 130 answers to 50 buyer questions about Southampton solicitors.

   Published 5 August 2026, corrected 5 October 2026 · 50 buyer questions, 130 answers · ChatGPT, Gemini and Claude, each read on its own · one AI model sorted named from recommended

   Also: What ChatGPT, Gemini and Claude actually say about accountants, our own sweep data

3. Your buyers are already asking AI instead of asking you

   Published 20 September 2026 · 389 public posts and replies · a keyword and tag count, made 18 August 2026

4. How buyers actually choose a professional, measured

   Published 20 September 2026 · 389 public posts and replies · a keyword and tag count, made 18 August 2026

How we measure

• We start from what buyers ask. We read what buyers post on Reddit, review sites and forums, and what people search for on Google. We use both, with questions we write ourselves or with AI, to choose the questions.

• We read each engine on its own. Up to 9: ChatGPT, Claude, Gemini, Perplexity, Grok, Google AI Overviews, Google AI Mode, Brave and Copilot. Each one can give a different answer to the same question.

• We keep the answers. Word for word, with the engine and the date, so the figures in our AI answer studies can be traced back to the answers behind them.

• We count named and recommended apart. An engine can name a company and still send the buyer somewhere else. Your Recommendation Rate™ is how often AI picks you as the answer.

• We say how we counted. A keyword count is called a keyword count, and each study on this page says how it was counted.

How we correct ourselves

On 18 August 2026 we checked the numbers we use as proof against the raw data. Four were overstated and two understated, and we corrected all six.

Whose data this is

Our two largest studies use the AI answers we collect for the companies we track. We never name a company we track without its permission.

Compiled by Steen Stones, founder of Avouch. Last updated 5 October 2026.