AI knows your business. It just never brings it up.

There is a comforting test business owners run on themselves: they open ChatGPT, type their own business name, and read what comes back. It is usually flattering. "Yes — [Business] is a well-regarded med spa in Dallas, known for…" That test feels like proof of AI visibility. It is very nearly the opposite.

Every audit we run asks the engines about the same business two ways. The branded question is the one owners ask themselves: "Is [Business] a good choice for [category]? What do people say about them?" The unbranded questions are the ones a customer asks, and they never contain the business's name: "I'm looking for a [category] in [city]. Who do you recommend and why?" and "What are the best [category] in [city]? Give a short ranked list." Each business gets both, on both ChatGPT and Perplexity — six answers in total. Comparing the two halves of that grid across 168 businesses is the cleanest measurement we have of the difference between being known and being recommended.

Finding 1: named 98.8% of the time by name, 19.8% of the time by need

Only four of 336 branded answers failed to identify the business at all. Effectively every business in the corpus exists inside the models' reach — a page, a listing, a review profile, something. But when the same engine, in the same session-style prompt, is asked the question that actually precedes a purchase, four out of five of those businesses are absent from the answer. Recognition is not distribution.

Finding 2: 58% show up in none of their four buyer answers

The average hides the shape. Businesses do not sit evenly around 20% — most of them sit at zero. Of the 168 businesses, 97 (58%) were named in not one of their four unbranded answers, while 96 of those same 97 were named in both of their branded ones. Only 8 businesses (5%) appeared in all four.

That zero column is the population most likely to believe it is doing fine, because the self-test — typing your own name — returns a paragraph of praise. The 97 businesses in it average an AI-visibility score of 17/100; the 71 that appear in at least one buyer answer average 41/100. The single behaviour that separates the two groups is not how the AI talks about you. It is whether you get retrieved when your name was never mentioned.

Finding 3: it is not a reputation problem

The obvious hypothesis is that the engines leave these businesses out because they have something bad to say. They don't. Across the 332 branded answers where the business was named, sentiment was positive in 88%, neutral in 10%, and negative in just 2% (6 answers out of 332). The models are not omitting these businesses from shortlists because of poor reviews — they are omitting them because, at the moment the shortlist is being assembled from third-party sources, the business is not among the candidates. Which sources those are, we measured separately: where AI looks before it recommends a local business — the business's own website is cited in only 23% of answers.

Finding 4: the pattern holds in every vertical and on both engines

And it is not an artefact of one model. On branded questions ChatGPT named the business in 99.4% of answers and Perplexity in 98.2%. On unbranded buyer questions ChatGPT named it in 17.6% and Perplexity in 22.0% — Perplexity, which searches the live web for every answer, is slightly more generous, but nowhere near closing a 5× gap. Two independently-built systems, the same behaviour: they know you, and they do not recommend you.

Why the two questions get such different answers

A branded question is a lookup. You have supplied the entity; the model only has to find something written about that entity anywhere in its reach and summarise it. An unbranded buyer question is a selection. The model has to assemble a candidate set for a category and a city — and it builds that set out of the third-party corpus it trusts for local recommendations: review aggregators, directories, "best of [city]" roundups, forum threads. If you are not in that corpus, you are not a candidate, no matter how good your website is or how positively the model describes you once someone types your name.

This is also why the self-test is so misleading. It tests the lookup path, which almost everyone passes, and tells you nothing about the selection path, which most businesses fail.

What to do with this

If you run the business: stop testing your own name. Ask the engines your customer's question — category plus city, no brand name — and see whether you are in the answer. If you are not, the fix is not on your homepage; it is in the off-site sources the engines assemble shortlists from, and it's ordinary, doable work that almost none of your competitors have started.

If you sell marketing services: this is the most disarming opener we have measured. A prospect who has already "checked ChatGPT" believes they are visible, and a screenshot of the branded answer proves them right — which is why leading with the unbranded answer, where they are absent and a named competitor is not, lands so hard. It reframes the conversation from "do you have an AI problem?" (which they will deny) to "here is the exact answer your customer sees, and here is who is in it instead of you." That gap is the retainer.

Either way the first step is running both questions properly — every engine, every phrasing, with the competitors tallied — which is what an AskedAbout audit does. Reselling to clients? The white-label Agency 5-pack puts five of these audits under your logo for $249.

Methodology

168 completed AskedAbout AI-visibility audits of real local businesses across 41 US metros and four verticals (med spas, personal-injury law firms, dentists, home services), run June–July 2026 — every audit in our corpus that completed the full six-answer grid. Each audit issues three prompts to two engines (OpenAI's ChatGPT and Perplexity): one branded prompt naming the business ("Is [Business] a good choice for [category]? What do people say about them?") and two unbranded buyer-intent prompts naming only the category and city. That is 336 branded and 672 unbranded answers, 1,008 in total. Each response is parsed for whether the business is named, its rank, the sentiment of the mention, and which competitors are named instead. Sentiment figures cover the 332 branded answers in which the business was named. Synthetic smoke-test and QA rows are excluded by exact business name via our canonical `audit_cut` script — never by test flags, which our own prospect audits also carry. n is stated so you can weight the figures; we re-publish as the corpus grows.

More from the same corpus: how many competitors AI names instead of you, the med-spa benchmark, the personal-injury benchmark, and where AI looks before it recommends a local business.