AI Visibility Study 2026: who AI assistants recommend

What we found

Every number below comes from one file: the dataset linked at the bottom of this page. We asked ChatGPT (GPT-5 mini with web search) and Perplexity (Sonar) 5 local-intent questions about each of 10 businesses, 10 answers per business, 99 usable answers in total after 1 engine errors. The sample is small and stated plainly: 10 businesses, 2 cohorts, 2 cities.

This is a pilot. It covers 1 category in 2 cities, and 2 of the 9 cohorts the run was designed to cover, because field work stopped early. Read it as a first measurement with a stated sample, not as a market report. It also covers two engines: Gemini (2.5 Flash with Google Search) was not included, for the reason set out in the method box above.

The single most useful way to read this is as a distribution, not an average. Most businesses sit at zero.

Why so few businesses are named

The arithmetic is unforgiving. Answering one of these questions, an engine returns a short ranked list, about 5.7 names on average and that figure is an upper bound because automated extraction occasionally catches a heading instead of a name. A city holds hundreds of businesses in the category. Being named is not a score you earn a little of at a time, it is a slot in a list of a few.

So 10 of 10 businesses scored zero. The engines were not refusing to recommend anyone: in 86 percent of answers an engine named at least one business, just not the one being checked, and only 14 percent of answers named nobody at all.

These were not obscure businesses. Every one of them was visible enough to be listed on request. Being listable and being recommended turned out to be different things.

The measurement is not simply missing them. None of the five questions mentions the business by name, so every business in a cohort was asked the identical five questions, minutes apart. That gives a control. 1 of the 10 businesses (10 percent) turned up by name inside the answers collected for a different business in the same city, found by the same matching rule, and 1 of those scored zero in their own run of the same question. The measurement finds these businesses when an engine names them. The engine does not name them consistently.

The two engines name different leaders

For 10 of the 10 businesses where both engines named at least one rival, the rival named most often was a different business on each engine (100 percent). Treat "the AI answer" as a fiction. There are several answers, and they disagree about who the leader is.

Where the answers came from

The engines cited 643 sources across 99 answers, and 0 of them pointed at the website of the business being checked.

The pattern in this list is the finding. Engines answering a local recommendation question lean on aggregators, directories and roundup pages, and on the websites of the businesses those pages already favour. A business absent from that layer is absent from the answer, however good its own site is.

What the websites were missing

Each home page was audited against the same 13 readiness checks the free score uses. Median readiness across the sample was 48 out of 100.

The check missed most often was "Publish visible FAQ answers", failed by 100 percent of the sites, followed by "Lead with an entity-first H1" at 90 percent and "Show name, address, and phone in visible text" at 80 percent.

Readiness and citations are not the same thing, and this data does not claim otherwise. Fixing a home page does not by itself put a business into an answer, because the answer is assembled from what the rest of the web says. What readiness does is make the business quotable once an engine reaches it.

Category and city

What this means if you run one of these businesses

Limitations

State these before quoting anything above.

• Sample size. 10 businesses, 99 answers, 2 cohorts out of the 9 planned. This is a pilot. Cohort-level numbers rest on a handful of businesses each and should be read as texture, not as a ranking of cities. A result this uniform on a sample this size is a strong signal to test further, not a settled fact.
• One run per business. Each business was asked each question once. Because no question names the business, each cohort question was in effect sampled once per business in that cohort at a different moment, but no single business was re-tested. Engines resample their own answers: five identical runs of one unchanged business on 2026-09-20 gave scores of 39, 46, 45, 44, 48 (standard deviation 3.0) and 9 to 12 named answers out of 20, with a different top competitor between runs. A single run carries that much movement.
• US only. Columbus and Tucson. Nothing here transfers to other countries without testing.
• No Google AI Overviews. Overviews have no stable queryable interface, so they are not measured. They are a large share of real AI answer exposure and their absence is a real gap.
• Engines personalise. Our queries carry no account history, no browsing history and no precise location beyond the city named in the question. A logged-in customer standing in the city may see different names.
• How the sample was drawn. The model that found these businesses is named in the method box, along with whether it also answered the measured questions. Either way this is not a random sample of every business in a city, it is a sample of businesses a live search model could enumerate on request.
• Live sites only. Businesses whose home page did not answer a request were excluded before any questions were asked.
• Extraction noise. Turning an answer into a list of business names is automated, and it sometimes captures a fragment of review text instead of a name. That inflates the names-per-answer figure and the count of distinct rivals, so read both as an upper bound. It does not affect whether the business being checked was named, which is decided by matching against that business's own name.
• Categories measured. dentist. Nothing here should be generalised to categories with different search behaviour, such as restaurants, law firms or trades.
• Engines measured. ChatGPT (GPT-5 mini with web search) and Perplexity (Sonar). Gemini (2.5 Flash with Google Search) was not measured. Two engines are not the AI answer layer.

The dataset