Why don't ChatGPT, Gemini and Perplexity recommend the same local businesses?

Tuesday morning, a water heater gives out. The tenant doesn't type "plumber near me" into Google - he asks his AI assistant, a program that answers in plain language, like ChatGPT. Then, to be safe, he asks a second one. The two answers have nothing in common.

The short answer: the three assistants don't read the same sources. One answers from memory, one leans on reviews, one reads directories live. Being recommended by "the AIs" isn't a thing - you are visible engine by engine, source by source. That is what AEO is about (Answer Engine Optimization: existing in assistants' answers, not in a list of links).

What did we measure this morning?

On 6 August 2026 we asked all three assistants the questions a homeowner actually asks - "Who are the best plumbers in Bordeaux?", "List the most recommended plumbers in Lille" - twelve answers, collected that same morning, using the public methodology of our July study. The cities are French; the pattern is not. Our UK study found the same split at scale.

Across four questions, ChatGPT declined to name anyone three times - then put forward a name we could not verify.

ChatGPT (queried via API, without web search - our methodology says so upfront) answers three questions out of four with a polite refusal: "I can't provide specific recommendations", followed by generic advice. On the fourth, it names "Plomberie Lille Services", complete with confident-sounding reasons - "known for fast response times". We could not match that name to an identifiable business, so we flag it as unverified. It is the most troubling finding of the morning: a confident answer tells you nothing about its reliability.

Gemini names businesses on all four questions, and announces its method: "based on Google reviews, local reputation and pricing transparency". A list built from memory, on the strength of aggregated reviews.

Perplexity - the only one of the three that searches the web while answering - thinks in directories: ThreeBestRated, PagesJaunes, review platforms. The names it surfaces are the ones those directories rank. In Bordeaux, one plumber comes out on top because ThreeBestRated puts him in its top three.

Why are the three answers so different?

Because there is no such thing as "the AI": there are three regimes of sources. The model's memory, frozen at training time. Aggregated reviews, which reward a well-kept Google profile. Directories, read at the moment the question is asked. Our July 2026 UK study showed it at scale: across 1,464 answers, the assistants agreed with each other on only 3% of the names they cited.

The practical consequence: a business can rank first on Google, appear in Gemini, and be invisible to Perplexity - simply because no directory that engine reads mentions it. AI visibility isn't won "inside the AI"; it is won inside each engine's sources.

What should you do tomorrow morning?

• Keep your Google profile current, with recent reviews: that is the raw material Gemini itself says it uses.
• Be present in the directories assistants actually cite - this morning, for plumbers: ThreeBestRated and PagesJaunes in France; check your market's equivalents. A listing takes ten minutes to verify.
• Measure the three engines separately: a score computed on a single assistant lies by construction. Our free audit asks ChatGPT and Gemini in 60 seconds, no account - a subscription adds Perplexity and its sources.

One honest note to close: this measurement is a dated snapshot, and assistants' answers move from day to day. Which is exactly why a single number is worth less than a measured trend - ours and yours alike.