AI reads the directories, then recommends the businesses

What we measured, and how

The corpus is 173 completed audits — 1,024 answers (64 med spas, 42 personal-injury law firms, 40 dentists, 27 home-services businesses, across US metros, 2026-06-12 to 2026-07-11). Synthetic QA rows are removed by an exact-name filter rather than a regex over the vertical field, a discipline we adopted after publishing a false finding in July and correcting it publicly. Every answer carries two things we can count separately: the businesses the engine named (extracted per answer as the competitor set) and the domains the engine linked to (extracted as raw URLs from the answer text). This post is the comparison between those two lists — what the engine reads against who the engine recommends — which, as far as we can tell, nobody publishes because it requires holding both in the same record.

Finding 1: directories are named as a recommendation 0.27% of the time

Across 4,015 business mentions spanning 2,230 distinct named entities, the complete list of directory and aggregator brands the engines put forward as an answer to "who should I hire" is this short:

Eleven mentions, in eleven months' worth of the kind of question the entire local-SEO industry is organised around. When someone asks an assistant "what are the best med spas in Dallas," the answer is a list of med spas — not a link to a list of med spas. The intermediary layer that local search has trained businesses to compete on does not survive into the answer.

Finding 2: the same directories are a quarter of everything the engine cites

Now the other list. Of the 511 ChatGPT answers in this corpus, every one carried at least one link (only 11 of 513 Perplexity answers did — a property of what that API returns, not of Perplexity's sourcing, which is why this cut is ChatGPT-only). Those answers reference 1,107 distinct domains across 2,815 domain-appearances, and the top of that list is almost entirely directories:

In total, 715 of 2,815 domain-appearances (25.4%) are directory or aggregator domains, spread across 41 distinct directory sites, and 73.6% of cited answers cite at least one of them. That 25.4% is a floor, not a ceiling: our classifier is a hand-curated token list, so directory-shaped domains it does not know about — `toprateddentist.com`, `trustanalytica.org` and `lawleaderboard.com` are all in the top 20 — are counted as non-directory. Social and forum domains, by contrast, are just 2.5% (71 appearances), which is worth stating plainly given how much AEO advice currently reduces to "post on Reddit."

Finding 3: the directories that matter are vertical-specific, and not the famous ones

The single most useful operational detail in this cut is that the directories an engine leans on are different in every vertical, and in three of our four they are not the household names:

A med-spa client's AI visibility runs through `medspascout.com` and `medicalspalocator.com`, sites most agencies have never opened, far more than through Yelp. Only one directory — Birdeye's review pages — is load-bearing in more than two of our four verticals. This is checkable per client rather than assumable: the answer text names its own sources.

What this means if you sell AI visibility to clients

• Listing work is not dead — it moved upstream. It no longer wins the click, because the answer never hands the buyer to the directory. It decides whether the engine has anything to read about the client before it writes a shortlist. Sell it as evidence supply, not as a destination.
• Audit the directories the engine cites for that vertical, not the ones on the standard checklist. In this corpus that means `medspascout.com` before Yelp for a med spa, and `expertise.com` before Angi for a home-services business. The per-answer source list is free to read — it is in the answer.
• Never promise a client they will "show up on Yelp in ChatGPT." Directory brands are 0.27% of recommendations. What the client can be present in is the source set the engine reads before it names someone.
• Report the two layers separately. Citations and recommendations move differently, and a tool that reports one of them and calls it "AI visibility" is reporting half the outcome. We publish both because on our data they disagree.

Limits — what this cut does not show

• Citations are ChatGPT-only. All 511 of its answers carried links; only 11 of 513 Perplexity answers did, because that API does not return citation URLs in the answer body. So Finding 2 and Finding 3 describe ChatGPT. Finding 1 (who gets named) covers both engines.
• Free-check engines only. This corpus is free checks, which run ChatGPT + Perplexity. Gemini and Claude are in the paid audit and are not in this cut.
• Our directory classifier is hand-curated. It is a token list over domains, checked against the top of the distribution. Errors run in one direction — an unrecognised directory is counted as a business site — so 25.4% understates the true share.
• URLs are extracted from the answer text. A source the model used but did not link is invisible to this method. It measures what the engine shows as its evidence, not everything in its retrieval set.
• Convenience sample. These are businesses we chose to audit, weighted toward med spas and personal-injury law, and single-sampled per question. Directionally strong, not nationally representative.
• We sell this measurement. AskedAbout sells AI-visibility audits, so discount the conclusions and check the method — which is why the method is stated in full.

You can run the free 60-second check on any business and see which businesses the engines name in its place. Agencies baselining a book of clients can run five at once with the $249 Agency 5-pack — 25 buyer questions × 4 engines × 3 samples per business, white-labeled, including the per-answer source map this post is a corpus-wide cut of.