The AI does not give your rating. It sets the bar at 4.5

Short answer: across 3,600 ChatGPT and Gemini answers in four countries, one gave the star rating of a business it named. But Gemini handed the customer a cut-off in 67 answers: filter the map at 4.5 stars, keep only businesses with at least 50 reviews. ChatGPT did it once. The cut-off is concentrated in France and Brazil, and in the trades where the AI names nobody: electricians, plumbers, heating engineers, locksmiths. The assistant does not score you. It tells your customer which score to reject.

It does not read your reviews for your customer. It tells them which ones to skip.

What was measured

Our four studies, re-read on September 30, 2026: the same buyer questions ("the best electricians in Belo Horizonte", "a good plumber in Lyon") put to ChatGPT (gpt-4o-mini) and Gemini through their APIs, without web search, in the United States (collected September 13, 2026), the United Kingdom (July 31), France (July 30) and Brazil (August 6). 3,600 answers, 1,800 per engine. In each one we looked for a star rating written as a figure ("4.8 stars", "4,5/5", "nota 4,8") or a number of reviews ("50+ reviews", "plus de 50 avis"), and read every passage found. The detectors were tested first on real phrasings in the three languages, including phrasings that must not count ("check the reviews", "open 9 to 5", "Level 5 Salon"). Each passage was then sorted: an official classification (a 4-star hotel, a Michelin star, 29 answers, set aside), the rating of a named business, or advice to the customer. A cut-off is advice that sets a floor: "4.5 or above", "at least 50 reviews".

The bar is 4.5 stars, and often 50 reviews

When the answer names a rating floor, it is 4.5 stars in 50 of 70 passages. A higher bar, 4.8, comes up 12 times, mostly in France and Brazil; a perfect 5.0 four times, all in Brazil. For the number of reviews, 50 is the figure given most often, in 21 passages; Brazilian answers set it lower, at 10 or 20 reviews, in 10 passages. And 20 Gemini answers, all French or Brazilian, spell out the same rule: a business at 4.5 on 200 reviews is more reliable than one at 5.0 on three. A perfect score on a handful of reviews is treated as a warning sign, not a strength.

Where it names nobody, it hands over a filter

The cut-off is not spread evenly. Electricians, plumbers, heating engineers, locksmiths and gas engineers account for 47 of the 68 answers that give one; Brazilian electricians alone for 13. These are the trades where the assistants most often decline to name a business at all (the trades where nobody is named) and send the customer to a platform instead (for a plumber, the answer is a platform). The filter replaces the name: "open Google Maps, search your area, keep those above 4.5 with dozens of reviews". In the United States the pattern barely exists: 3 answers out of 872, which matches what we found on September 19 (what AI tells customers to check).

What it means if you are the business

The customer who follows this advice meets your rating before your name. At 4.4 stars, or at 4.9 on twelve reviews, you are out of the list they build, however good the answer might have been about you. Three things follow from the measurement. The volume of reviews counts as much as the average: the answers explicitly prefer 200 reviews at 4.5 to a handful at 5.0, so asking every satisfied customer is worth more than protecting a perfect score. The bar sits at 4.5, so the useful question is how far you are from it, not whether you are "well rated". And some of these answers tell the customer to read the recent reviews and the one-star ones, and to see whether the business replied: an unanswered complaint is read by exactly the customer the AI sent. None of this gets you named; it keeps you from being filtered out after the answer.

What this does not prove

• These are the models' answers through their APIs, without web search: the way our audit asks. With search switched on, the consumer apps can show map cards with the actual ratings, which this measurement does not see.
• Four studies collected between July 30 and September 13, 2026, one pass per question. The models have moved since, and nothing here measures that drift.
• 68 answers giving a cut-off is a small base: 3.7% of Gemini's answers, and a handful per trade and city. The pattern is clear by country and by engine; the figure for any single trade is not.
• The sorting into cut-off, advice and classification was done by rules, then every passage was read. A borderline phrasing could still sit on the wrong side of a rule; the counts per country moved by a few units between our two passes, not the order.
• Nothing here shows how many customers actually apply the filter. The measurement says what the assistant tells them to do, not what they do.