"ChatGPT gets 70% of its local businesses from Foursquare." We tried to source that number, then measured 526 ChatGPT answers.

The claim, and where it actually comes from

The advice is everywhere and it is unusually specific, which is what makes it persuasive: not "Foursquare matters" but 60 to 70 percent. A number that precise implies somebody counted. We went looking for who, reading each page for its sourcing sentence rather than its headline. Here is the chain, in the order the pages were published:

So the fifteen months of confident percentages resolve to one measurement, reported second-hand, that we could not open. That is not an accusation that it is wrong — the described method (reading ChatGPT's internal JSON responses on local prompts) is a reasonable one, and it may well have been careful work. It is an observation that the entire industry is quoting a figure nobody can currently check, about an engine that has shipped several search-stack revisions since it was taken.

What the two companies say themselves: nothing

The claim is usually paired with the assertion that OpenAI and Foursquare announced a places-data partnership. We checked both companies' own surfaces on 2026-08-02. OpenAI's "Introducing ChatGPT search" — the canonical post describing how ChatGPT answers with web sources, carrying dated updates through 2025-02-05 — contains the string "Foursquare" zero times; it describes the value of search in terms of "sports scores, news, stock quotes, and more" and names no places provider. Foursquare's own resources and blog index and its press page carry no OpenAI or ChatGPT announcement in their listings. The most-shared artefact for the alleged partnership is a third-party social post, which is not a durable primary source and is not something we will report as one.

None of that proves no commercial relationship exists — companies license data quietly all the time, and a places API is exactly the kind of deal that never gets a blog post. It does mean that the foundation under a very widely-followed piece of advice is a number with no reachable origin and a partnership with no announcement. That is worth knowing before you spend a morning on it.

What changed between 2026-08-02 and 2026-10-07

On 2026-10-07 we re-read the three pages that carry the claim and both companies' own surfaces, then recounted the corpus with the same scripts. One page rewrote itself, one new primary source appeared, and the rest held.

The Yelp filing is the first primary-source statement we have found from either side of any ChatGPT local-data relationship, and it names Yelp, not Foursquare. The letter, signed by Yelp's CEO and CFO and dated 2026-08-06, reads: "In the second quarter, we saw robust demand for our data licensing products, including from our partnership with OpenAI, contributing to strong growth in other revenue. Yelp ratings and reviews recently began powering ChatGPT's local experience in relevant categories." The same-day results press release has the CEO saying "our trusted content is powering local discovery for ChatGPT and other AI partners." The letter also carries a chart footnote describing an analysis of "the number of times Yelp was cited by ChatGPT, Google Gemini, Perplexity and Google AI Mode in the fourth quarter of 2025" against five home-services platforms (Better Business Bureau, Angi, Thumbtack, HomeAdvisor and Nextdoor), United States only, links counted whether clicked or not, ads excluded. The letter does not print the counts and we have not seen the underlying analysis, so we quote the method and not a number.

That filing matters for this post in one specific way. The honest-limit section below says a licensed backend would be invisible to a citation count by construction. Yelp's statement is a licensed backend, documented by the licensor, in categories our corpus covers, and yelp.com is cited in 0 of 541 ChatGPT answers on the 2026-10-07 recount. A zero and a working data feed now sit side by side on the same engine, which is the case the caveat was written for. It does not revive the Foursquare number: no equivalent statement about Foursquare exists from OpenAI, Foursquare or any filing we can find, and the one guide that used to carry the percentage has replaced it with the Yelp citation.

What we can measure: 526 ChatGPT answers, 1,115 domains, zero Foursquare

Every AI-visibility audit we run asks ChatGPT and Perplexity the questions a customer actually asks — "I'm looking for a [category] in [city]. Who do you recommend and why?" — and we keep every URL the engine attaches to its answer. As of 2026-08-02 the production corpus is 178 completed audits, 1,054 answers, 526 of them from ChatGPT, across dentists, med spas, personal-injury law and home services in 28 US metros. 537 answers carried at least one URL, citing 1,115 distinct registrable domains. We looked up the directories that local-AEO advice most often tells you to go claim:

The six platforms most commonly listed as "what feeds AI local recommendations" — Google Business Profile, Yelp, Foursquare, Angi, HomeAdvisor, Nextdoor — account for 18 of 526 ChatGPT answers between them, and the Better Business Bureau, which is on almost nobody's list, accounts for 92 on its own.

Recount, 2026-10-07. Same cut, same scripts, that day's export: 183 completed audits, 1,084 answers, 541 from ChatGPT (543 from Perplexity), 552 answers with at least one URL, 1,133 distinct domains. The newest completed audit is dated 2026-09-09, so the corpus added 5 audits and 15 ChatGPT answers since 2026-08-02 and has been still for four weeks. Every directory that read zero in August reads zero now: foursquare.com, yelp.com, google.com and business.google.com, facebook.com and instagram.com, yellowpages.com, superpages.com, manta.com and citysearch.com. The counted domains moved by at most eight answers:

The six platforms named in the advice (Google Business Profile, Yelp, Foursquare, Angi, HomeAdvisor, Nextdoor) account for 21 of 541 ChatGPT answers on the recount, against 18 of 526 in August; bbb.org alone is 93. The recount artifact is content/news/foursquare-recount-2026-10-07.json (5 of 5 self-checks: every August zero still zero, every corpus count grew or held, no audit after 2026-09-09, the two engine keys only, every directory count ChatGPT-only).

The honest limit of this measurement

A zero in a citation count cannot refute the Foursquare claim, and we are not going to pretend otherwise. The claim is about a retrieval backend: if ChatGPT resolves a local prompt by calling a licensed Places API, the response is structured data, not a web page, and there is no URL to attach. It would be invisible to our corpus by construction — and equally invisible to every agency blog repeating the number, none of which has published a measurement either. The two things a citation count can see are these:

• What the answer points the reader at. When ChatGPT recommends a dentist and links somewhere, it links to birdeye.com, bbb.org or a niche vertical directory — not to Foursquare, Yelp or Google. That is the surface a prospective customer clicks through to, and it is the surface your reputation is read from.
• Where the engine goes when it does read the web. 537 of 1,054 answers carried a URL. Whatever fills the other half, the read half is a ranked, checkable list — and it is not the list in the advice.
• What the claim would need in order to be actionable. "Foursquare supplies 70% of names" only becomes advice if being in Foursquare changes whether you get named. Nobody, including us, has published that experiment. Until someone does, the claim is a description of plumbing, not a lever.

What we would actually do with the next hour

1. 1Find out which domains carry your category in your city, rather than assuming the national list applies. The concentration is extremely vertical-specific — medspascout.com is cited in 63 answers and every single one is a med spa; it is worth nothing to a law firm. Our earlier read of this corpus found no universal directory: 1,092 domains, cited vertical by vertical.
2. 2Claim the two or three that actually appear, and get the details right there — hours, services, address, and above all recent reviews. A profile on a cited directory is the mechanism by which the engine learns you exist.
3. 3Do not skip Foursquare on our account. It is a free listing that takes ten minutes and there is a plausible mechanism for it mattering that our data cannot see. Ten minutes is a fine bet on an unfalsified claim. A morning, or a paid "Foursquare optimisation" engagement, is not — that is buying certainty nobody has.
4. 4Measure the thing you actually care about, which is whether the engines name you on a buyer's question. That is a direct observation, and it does not require resolving any of the above.

Methodology

Corpus cut of 2026-08-02 from the production promptRuns table via ops/audit_cut.py (the canonical cut; 27 synthetic smoke rows dropped), with domains extracted by ops/cited_domains.py. Every URL in every completed answer in a published vertical, collapsed to a registrable domain. The unit is answers-citing-domain, not links — an answer counts once for a domain no matter how many times it links it, which is why birdeye.com reads 143 answers and 277 links. Verticals: dentists, med spas, personal-injury law, home services. Engines: ChatGPT (526 answers) and Perplexity (528). The directory lookups above are exact registrable-domain matches against the full 1,115-domain set, so "0" means the domain is absent from the entire corpus, not merely outside the top ranks. Recount of 2026-10-07: the same two scripts on that day's export (72 synthetic rows dropped; the filter is by business name, and the dropped count grew with the growth agent's test audits, none of which enter the cut): 183 audits, 1,084 answers, ChatGPT 541, Perplexity 543, 552 with a URL, 1,133 domains, newest completed audit 2026-09-09. Script and output: content/news/foursquare_recount_2026-10-07.py and foursquare-recount-2026-10-07.json.

Sourcing checks were performed on 2026-08-02 and repeated on 2026-10-07 (the table under "What changed") by reading each cited page and its sourcing sentence directly; the Yelp filing was read on EDGAR. "No link to the original" means the page names the research but provides no hyperlink to it and we could not locate a published version by search.

The takeaway

The most-repeated number in local AEO is one unlinked 2025 measurement wearing fifteen months of confident restatement. It might be right. But the advice built on it — go claim Foursquare — is being sold with a certainty the evidence does not carry, while the domains that demonstrably carry your category's answers go unclaimed. Prefer the measurement you can open over the number everybody quotes. As of 2026-10-07 the one such measurement from a data supplier is Yelp's own filing, and it says Yelp.

The free 60-second check asks ChatGPT one of your buyers' real questions and tells you whether it names you. The $79 audit runs 25 questions across ChatGPT, Perplexity, Gemini and Claude, three times each, and returns the citation source map — the exact third-party domains the engines pulled from for your category in your city, which is the only version of this list that can tell you where to spend the hour. Agencies auditing a book of clients can run five at once with the $249 Agency 5-pack.