Every local market we audited has an AI shortlist: a mean of 4.7 rival businesses named in every single audit, no matter which business we asked about

What we measured, and how

Our production corpus is 173 completed audits (synthetic QA rows excluded by a name-level filter, not a vertical regex — a discipline we adopted after publishing a false finding in July and correcting it). Each audit asks buyer-intent questions like “What are the best med spa in Dallas, TX?” and records, per answer, which businesses the engine named. For this cut we grouped audits by vertical × metro and kept only groups with three or more distinct audited businesses: 15 markets across med spas, cosmetic dentists, personal-injury law, and electricians, covering 59 businesses and 343 answers from ChatGPT and Perplexity between June 12 and July 10, 2026. Then we asked a question a single audit can never answer: do the rival names change when the subject changes?

Finding 1: every market has a fixed incumbent shortlist

In each of the 15 markets we counted the rival names that appeared in every audit of that market — different subject business, different audit, different day, different run. Not one market came back empty. The mean was 4.7 universal names (minimum 2, in cosmetic dentistry in Tampa; maximum 8, in med spas in Newport Beach). These are the businesses the engines return no matter who is asking or who is being asked about — the market's de facto AI shortlist.

Finding 2: the long tail is mostly noise

Across the 15 markets the engines named 598 distinct rival businesses in 1,363 naming slots. 372 of those names — 62% — appeared in only one of the market's audits. The top three names in a market took 25.8% of its rival mentions on their own, and the universal shortlist took 34.6%. So a competitive picture built from one audit is mostly transient names: roughly two thirds of what you see is a rival that will not appear the next time the same market is queried about a different business.

This is the market-level version of a run-to-run result we published earlier: only 31.4% of recommended brands survived every repeat of the identical question. It also matches what the first critical academic survey of generative engine optimization, published on arXiv on July 15, 2026, concluded from 45 studies — visibility is a distribution, and a point estimate is not a stable indicator.

Finding 3: most businesses are absent from their own market's answers

For each audited business we checked whether it was named in the answers generated for its direct competitors in the same market. Only 25 of 59 — 42% — appeared at all. The other 58% were entirely absent from their own market's competitive answers: an engine asked about the best providers in their city, in their category, never returned them, no matter which neighbouring business prompted the question. Mean mention rate for their own branded audits was 41.8%, so these are not dead businesses — they are businesses the engines will name when asked about them directly and not name when asked about the market.

Why this matters

• The competitive set is smaller than the competitor list. A client's real AI competition in a metro is a handful of names — 2 to 8 in our data — not the 30–56 distinct rivals a single audit's list implies. Strategy should be scoped to the shortlist; the tail is sampling noise.
• “We appear when people search our name” is not visibility. 58% of these businesses never surfaced in an unbranded, market-level answer. Branded recall and category recall are different measurements, and only the second one wins a customer who has not heard of you yet.
• One audit cannot find the shortlist. The shortlist only becomes visible when you audit several businesses in the same market — the intersection is the signal. That is a structural argument for auditing a book of businesses rather than one, and it is why we sell a multi-business pack at all.
• Track the shortlist, not the rank. Engines do not publish a ranking, and the ordering within an answer moved between runs in our corpus. The stable, trackable object is membership: are you in the set of names this market's answers keep returning?

Limits of this cut, stated up front

• Two engines, not four. These are ChatGPT and Perplexity answers (172 and 171 respectively). Gemini and Claude are in the paid audit but not in this cut, and engines diverge sharply — our per-engine data shows the same business can be a fixture on one engine and absent from another.
• Markets of 3–8 businesses. A market with 3 audited businesses gives a weaker intersection than one with 8. The Charlotte med-spa market (8 businesses) is the strongest single case and it still produced 3 universal names.
• A one-month window. Audits ran June 12 – July 10, 2026. Shortlist membership over a longer horizon is exactly the thing we do not yet have and are not going to assert.
• We sell audits. Every finding here argues, conveniently, for buying measurement. Read the numbers and the method; discount the conclusion accordingly.

The takeaway

AI answers about a local market are not a fresh draw each time. They have an incumbent core — a mean of 4.7 businesses per market in our data — surrounded by a large, mostly transient tail. Being in that core is the thing worth measuring and the thing worth moving; being named once in one answer is not evidence of either.

The free 60-second check gives a business its mention rate across ChatGPT and Perplexity for its own category and city. Agencies who want the market-level view — the intersection that reveals the shortlist — need several businesses in the same market: the $249 Agency 5-pack runs the full repeated-sample audit (25 questions, 4 engines, sampled 3× each) across five businesses, white-labeled, which is how this analysis was possible in the first place. It is the SKU built for agencies and GEO consultants.