Does ChatGPT link the brands it names? In 520 local answers, a quarter of brand mentions carried no link at all

The claim being made about unlinked mentions

The load-bearing number in this argument is older and harder than the new study. Ahrefs, 2025-05-26, across 75,000 brands: branded web mentions correlate with AI Overview brand visibility at 0.664, against 0.218 for number of backlinks and 0.295 for referring domains. Mentions outrank links by roughly 3×. Ahrefs states the caveat plainly in its own post — “correlation ≠ causation” — and it matters here, because established brands accumulate mentions and visibility together and brand strength is an obvious common cause.

The July 2026 BrandMentions study (author: Cornelia Cozmiuc) adds a fresh composition read on top of it: of the mention footprint around repeatedly visible brands, “roughly 70 to 80%” came from third-party sources, while rarely-visible brands leaned on their own domains at “30 to 40%” of their total footprint — and the third-party mentions were majority text-only, with hyperlinked mentions accounting for only “about 25 to 35%.” The study carries its own honest disclaimer: “Every relationship reported here is an observed correlation within a single sample frame, not a demonstrated causal link.”

Two things are worth flagging before anyone builds a strategy on it. BrandMentions sells a brand-mention monitoring tool, so the finding is congenial to the seller — as is ours, and we will say so again at the bottom. And its figures are reported as ranges rather than point estimates, which is a real limit on how precisely you can act on them. What neither study measures is the other end of the pipe: whether the engine, when it finally writes an answer, hands the brand a link.

What ChatGPT actually does with 25 brands it knows

We check 25 review, directory and reputation brands — hand-picked because their names are unambiguous in English prose — against every ChatGPT answer in our audit corpus. For each answer we ask two independent questions: does the brand's name appear in the prose, and is its domain linked anywhere in that answer? Unit is answers, not links.

The brands buyers know get named. The domains they've never heard of get linked.

Read the table top-down and the top of it looks like a link machine: Birdeye named 143 times and linked 141, Three Best Rated 46 and 46, Expertise.com 41 and 39, Zocdoc 27 and 26. Near-perfect conversion from mention to link. Now read the rows where the two columns come apart: Yelp, named in 57 of 520 answers and linked zero times. Super Lawyers, named 26, linked 3. Avvo, named 22, linked 9. Martindale-Hubbell 19 and 8. Angi 19 and 10. RealSelf 13 and 6. Facebook, named 6, linked 0.

Those two groups are not interchangeable. The reliably-linked set is largely SEO aggregator directories most buyers could not name unprompted. The reliably-unlinked set is the consumer-facing reputation layer — Yelp, Avvo, Super Lawyers, Facebook — the names a person recognises and weighs. ChatGPT is using the second group as evidence it repeats and the first group as pages it points at. The Yelp row has a documented mechanism behind it: Yelp's robots.txt blocks OAI-SearchBot from everything except /article, so its ratings reach the answer secondhand, through pages that quote them.

Three implications if you care about AI visibility

• Any tool that counts citations is undercounting brand presence by roughly a quarter. 158 of 639 brand-answer pairs here exist only as prose. A link-based tracker — ours included, when it reads the link list — records those as absence. If your AI-visibility report is a list of domains cited, it is measuring the engine's footnotes, not its sentences, and those are different populations.
• Getting linked and getting named are different games, and only one of them is worth winning. The buyer reads the answer; almost none of them click the citation. A mention with no link still moves the sale. That is the honest case for the Ahrefs correlation — not that unlinked mentions are a ranking trick, but that they are how the engine's actual prose gets written.
• For your own business, the target is being the named thing, not the linked thing. The pages ChatGPT links here are aggregators republishing third-party reputation data. You do not get onto that list by publishing more; you get onto it by having a correct, well-covered record on the pages that already feed it. We measured what that surface looks like: 1,092 distinct domains across 511 answers, 96.8% of them cited in only one vertical, and 89.7% appearing in exactly one metro. There is no master list.

Method

• Corpus: the production `promptRuns` table, cut by `ops/audit_cut.py` (the canonical smoke-row-free cut — 24 synthetic QA rows dropped). 176 audits, 1,042 answers, of which 520 are ChatGPT/OpenAI answers, run 2026-06-12 to 2026-07-27, across 41 US metros and four verticals (dentists, med spas, home services, personal-injury law).
• Naming: case-insensitive, word-boundary regex per brand, run against the answer text after URLs are stripped out, so a domain string inside a link URL is never counted as prose.
• Linking: every URL in the answer collapsed to a registrable domain and compared to the brand's domain. An answer counts once per brand either way.
• Brand list: hand-curated and deliberately conservative — only names unambiguous in English prose, each pinned to one domain. Brands whose names are common words were excluded rather than guessed at. All 158 unlinked matches for the four riskiest patterns (Martindale, Angi, Super Lawyers, Facebook) were spot-checked in context before publication; all were genuine brand references.
• Script: `ops/unlinked_mentions.py` in our repo, built on the same cut as the 1,092-domain citation analysis.

Limits

• "Named" is a superset of "linked" by construction, so the only meaningful column is the last one. ChatGPT writes markdown links whose anchor text is the brand name, so a linked domain is essentially always also named in prose — which is exactly why linked-but-not-named is 0 for all 25 brands. The 24.7% figure is the defensible one: it requires no link to that domain anywhere in the answer.
• ChatGPT only, and that is not a stylistic choice. We capture Perplexity's text through the API rather than its interface citation panel, so its link counts are an instrument artifact of ours. We can report Perplexity mentions; we cannot report Perplexity links, and a mentioned-but-not-linked claim needs both.
• 25 brands, not a census. This is a fixed panel, not every brand ChatGPT names. A different panel gives a different percentage. The point is the pattern across the panel — consumer-known brands unlinked, aggregators linked — not the precision of 24.7%.
• Convenience sample. 176 businesses we chose, four verticals, 41 metros, single-sampled per question. No restaurants — the vertical where Yelp is strongest.
• We sell this measurement. AskedAbout sells AI-visibility audits, and a post arguing that link-counting undercounts visibility is congenial to a company that reads answer text. The method is written out above precisely so you can check it rather than trust us.

The free 60-second check runs a business's own buyer questions and returns the answers the engines gave — the prose, not just the link list, so you can see whether you are being named without being linked. Agencies mapping this across a book of clients can run five businesses at once with the $249 Agency 5-pack.