Last week we ran a bigger version of the test we published in September. Then it was 15 firms across four industries. This time it was 45 real UK companies in two sectors that sell on trust and reputation: 26 in executive coaching and leadership development, 19 in accountancy and financial services. All of them are established businesses with a website, a named leader, and clients.
We asked three AI engines two kinds of question. First, the question a buyer asks: who are the best executive coaching firms in the UK, which accountants are recommended for startups and small businesses in London, and so on. Then, for each of the 45 firms, the question a buyer asks once they have a shortlist: what does this company do, who runs it, where is it based, who do they work with.
The engines were ChatGPT with web search on, Google's Gemini, and Perplexity. Every query was run from a UK location on 24 September 2026 and every answer was kept word for word.
What we found
Four of the 45 were recommended by at least one engine. Forty-one were never mentioned. Not ranked low. Not mentioned. When a buyer asks an AI engine who to consider, 41 of these firms do not exist.
Asked about a firm by name, the engines got 19 of the 45 materially wrong or incomplete. The most common problems were a wrong or missing description of what the company actually sells, an inability to say who leads it, and confusing the company with another business of a similar name. In five cases at least one engine simply could not say who runs the company, even though the answer is on the firm's own About page.
The engines were honest about their uncertainty, which is its own problem. Many answers carried hedges: "I am reasonably confident about the broad picture", "the public search results I found do not clearly name a founder". A buyer reading that does not phone the firm to check. They move to the next name on the list, which is one of the four.
Who the engines recommend instead
The recommendation lists were dominated by two kinds of name. In accountancy: the Big Four, the national mid-tier firms, and the cloud-first online accountants that publish a lot of comparison content. In coaching: a handful of firms that have been written about elsewhere, and the professional bodies. In both sectors the same few names came back across all three engines, which tells you the engines are drawing on the same small pool of third-party mentions.
That is the mechanism. AI engines do not recommend firms they know from a website. They recommend firms other websites talk about.
What this means if you run one of these businesses
Three things, in order of cost. First, your own site has to answer the four questions plainly: what you do, who leads it, where you are, who you work with, in text an engine can quote, not in a hero image. Half of the description errors we saw came from sites that make an engine guess.
Second, someone else has to say it too. A directory listing, a trade press mention, a client's case study, a professional body's member page. The engines weighted these heavily and so should you. This is the slow part and there is no shortcut we would recommend.
Third, check what the engines say now, before a buyer does. It takes ten minutes and it is free. We built a tool for exactly this, the AI Visibility Score, which runs the same questions against your company and shows you the verbatim answers. If you would rather do it by hand, ask each engine the two questions above and read the answer as a stranger would.
Method, so you can repeat it
45 UK firms, chosen from public directories: 26 in coaching and leadership development, 19 in accountancy and financial services. Three engines: ChatGPT with web search enabled, Gemini, Perplexity, queried through DataForSEO's live endpoints from a UK location on 24 September 2026. Two recommendation questions per sector, then one direct question per firm covering services, leadership, location and clients.
"Recommended" means the firm was named in an engine's answer to a recommendation question. "Materially wrong or incomplete" means at least one engine misdescribed the services, could not name the leader, confused the firm with another, or gave a wrong location. Firms are not named here; each one has received its own answers privately. The smaller September test is here.
We publish this because we sell the fix, and we would rather you check the numbers than take our word for it. The raw answers for any firm in the sample are available on request.