OpenAI and Foursquare have a real partnership. Foursquare’s Places dataset — over 100 million points of interest — genuinely feeds ChatGPT’s local answers, and has since the partnership was announced. That’s not in question, and nothing below should be read as suggesting otherwise.
What’s in question is a specific number: that 60–70% of ChatGPT’s local results come from Foursquare’s city-guide listings. It’s repeated constantly in local SEO circles. As far as we could find, nobody had checked it against a large, current sample. So we built one — 2,880 prompts across 12 verticals and 12 US markets, deliberately weighted toward the categories and cities where Foursquare should be strongest — and we traced the number back to where it actually came from.
Both halves of that turned out to be more interesting than we expected.
Numbers
0.06% — Foursquare’s share of citations in our large-scale US measurement, essentially indistinguishable from zero (one real citation in 4,607 successful runs)
95.83% — the share of runs where Yelp appears in the same structured-grounding slot the original claim was about for Foursquare
87.5% — the share of downstream repetitions of the “60–70%” claim that dropped every caveat the moment they repeated it
60–70% — what a real, small study actually found, in Spain, in May 2025 — fifteen months and, it turns out, one major OpenAI data partnership ago
What everyone repeats
“60–70% of ChatGPT local results come from Foursquare city-guide listings.” You’ll find it in local SEO blogs, LinkedIn posts, and marketing guides throughout 2025 and 2026, usually stated as settled fact. It has a specific ring to it — precise enough to sound measured, round enough to be memorable.
We went looking for where it came from. We found it. We also found that the ChatGPT it describes doesn’t quite exist anymore.
Where we traced it
The earliest version we could find traces to a small, real study — not a fabrication, which is worth saying plainly before anything else. On 5 May 2025, Natzir Turrado, a Spanish SEO consultant, ran 50 prompts across five Spanish cities, inspecting ChatGPT’s raw JSON responses directly rather than just reading the rendered answer. The finding, in his own words: 60–70% of the businesses ChatGPT displayed in first place came from Foursquare.
That’s a real measurement. It’s also a narrow one — five Spanish cities, one country, and specifically the top-ranked result, not “local results” broadly — and none of that scope survived the trip into general circulation. Not even the original post flagged it; the finding was written up in general terms despite an all-Spain sample. A companion post by a second researcher, published the same day, ran a smaller check (three queries) and found no clean percentage at all, just a note about how quickly Foursquare’s data resolves compared to alternatives.
Nine days later, a digital marketing blog folded both posts into a single flat claim: “Between 60% and 70% of local results on ChatGPT come from Foursquare.” No geography, no “first result only,” no sample size. Two months after that, a LinkedIn Pulse post asserted “over 70%” from an unnamed vendor’s “analysis of millions of citation data” — no methodology disclosed, and functioning, on inspection, as promotional content for a citation-tracking product. We’re not naming that author here; per our own ground rules for this piece, repeating a plausible-sounding number isn’t the kind of thing that should follow someone around, even when we can trace exactly who wrote it.
BrightLocal handled this correctly, and it’s worth saying so plainly. Their 22 July 2025 piece on AI search cited both of the above sources, but hedged: “Reports suggest that a significant 60% and 70% of local results on ChatGPT come straight from Foursquare’s city guide listings.” They linked straight to their sources rather than presenting the number as self-verified. That’s exactly the right instinct, and among everything we read while tracing this claim, they’re the only ones who showed it.
They didn’t catch that the number traces to a five-city Spanish study, or that one of their two cited sources doesn’t actually contain the percentage it’s credited with. But nobody else did either — and “reports suggest,” followed by a live link, is a genuinely useful piece of process for anyone who wants to repeat an interesting number responsibly.
From there, the caveats didn’t just fail to travel — they were actively replaced with more confidence. We found eight further repetitions between May 2025 and June 2026: a blog that dropped all sourcing and rounded “60–70%” up to “70%+”; a guide that added an entirely new comparative claim (“more than Google Business Profile and Yelp combined”) that appears nowhere in anything upstream of it; a post built around a personal anecdote with no source at all — using a plumber as its example business, the exact kind of category our own data shows Foursquare has essentially no presence in. Of those eight downstream repetitions, exactly one carried any hedge language. That’s an 87.5% caveat-stripping rate, and it’s the most useful finding in this whole trace — more on that below.
Why the sample favors Foursquare
If you sample plumbers and dentists, you’ll find roughly 0% Foursquare presence and the test will be unfair — nobody would trust it. So we didn’t build the sample that way. Seven of our twelve verticals (restaurants, bars and nightlife, coffee shops, hotels, museums and attractions, live music venues, bakeries) are exactly the categories Foursquare’s dataset is strongest in. Our twelve markets are weighted toward dense urban cores — New York, San Francisco, Chicago, Los Angeles — where Foursquare’s city-guide data should have the most to work with. Two categories (plumbers, auto repair) were kept in specifically as unfavorable controls, to show the gradient rather than to drag the average down.
If Foursquare’s share is still near zero on its best ground, in its favorite cities, the claim doesn’t have anywhere left to hide.
Finding 1 — Foursquare’s citation share is not measurably different from zero
A quick definition, since it matters: what we’re calling “citation share” here is how often Foursquare shows up as a source ChatGPT actually links to or retrieves — a citation-frequency number, not a measure of how much Foursquare’s data shapes the answer underneath. We built a second test for that gap; it’s Finding 3.
On the number itself: across 2,879 successful runs on ChatGPT’s primary product surface, Foursquare-controlled domains appeared in exactly 0.00% of citations — not in the links ChatGPT shows, not in the pages it retrieves but doesn’t show, and not in the URLs attached to its structured business cards. That held in every one of the twelve verticals we tested and every one of the three market tiers, including the food-and-nightlife categories and the dense-urban markets the sample was built to favor.
We ran a second, independent measurement on a completely different API surface as a cross-check — 1,728 more prompts, collected separately. There, Foursquare’s share was 0.06%: one real citation, to a Foursquare place-redirect page for a gym in Los Angeles, out of 1,728 runs. One citation is not zero. It’s also not meaningfully different from zero at this sample size — the 95% confidence interval on that number tops out at 0.33%.
| Layer | Share | 95% CI |
|---|---|---|
| Cited (primary surface) | 0.00% | [0.00%, 0.13%] |
| Retrieved but not cited (primary surface) | 0.00% | [0.00%, 0.13%] |
| Business-card grounding (primary surface) | 0.00% | [0.00%, 0.13%] |
| Cited (independent cross-check surface) | 0.06% | [0.01%, 0.33%] |
What this means: if you’re a local business owner who’s been told to prioritize your Foursquare listing because “60–70% of ChatGPT comes from there,” that advice doesn’t hold up against a large, current, US sample. It might have held up in Spain in May 2025. It doesn’t now.
Finding 2 — Yelp has the slot the original claim was describing
The zero above raises an obvious question: if Foursquare isn’t grounding these answers, what is? We checked. On 23 July 2026 — four weeks before we collected this data — OpenAI announced a non-exclusive data-licensing deal with Yelp for ChatGPT’s local answers. We checked whether that shows up in our numbers. It does, decisively.
| Provider | Cited | Retrieved | Business-card grounding |
|---|---|---|---|
| Yelp | 1.04% | 2.08% | 95.83% |
| TripAdvisor | 13.55% | 20.70% | 28.90% |
| OpenTable | 1.70% | 3.58% | 16.88% |
| 0.07% | 0.17% | 4.62% | |
| Apple Maps | 0.14% | 0.97% | 0.00% |
| Foursquare | 0.00% | 0.00% | 0.00% |
95.83% of our runs had at least one Yelp link attached to ChatGPT’s structured business cards — the same integration slot the original claim described for Foursquare. And that rate barely moves by category: it’s 93.75% for hair salons and 97.92% for plumbers, the exact “unfavorable” category we included as a control. Yelp reads as a universal default now, not a category specialist.
Foursquare, meanwhile, is the only major local-data provider we checked — out of ten, including names as small as Bing — that measured exactly zero everywhere we looked.
What this means: the original 60–70% figure wasn’t measuring a permanent property of Foursquare’s data. It was measuring a partnership snapshot from May 2025, in one market, on a product that OpenAI has since materially changed.
Finding 3 — the “invisible grounding” rebuttal doesn’t hold up either
Here’s the obvious objection to Finding 1: maybe Foursquare’s data reaches ChatGPT as structured place data rather than as something it cites — in which case a citation-share number close to zero wouldn’t tell you anything, because Foursquare could still be shaping the answer without ever showing up as a link. We took that seriously enough to build a dedicated test for it, and we said in advance what would count as a real answer either way.
The test: does the set of businesses ChatGPT actually names look more like Foursquare’s ranked results, or more like Google’s? If Foursquare is grounding answers invisibly, ChatGPT’s picks should resemble Foursquare’s more.
They don’t. Across 137 comparable market-and-category combinations, ChatGPT’s named businesses resembled Google’s set more than Foursquare’s (mean overlap 7.5% vs. 5.7%), and the gap is statistically real (p = 0.008), not noise. If anything, the invisible-influence story points toward Google, not Foursquare.
What this means: we don’t think this closes the door on every possible version of “maybe it’s hidden” — a single similarity metric never fully can — but it’s the strongest test we could design in advance, and it came back in the opposite direction from what would rescue the original claim.
The real story: caveats get stripped, and that’s the reusable finding
If this were only “we checked a number and it was wrong,” it wouldn’t be worth much — most repeated statistics eventually get one detail wrong. What made this trace worth doing is what happened between the original measurement and today.
Someone really did run a study and find 60–70%. It just described five Spanish cities, in May 2025, looking only at the first-ranked result. Every one of those three qualifiers disappeared within nine days, and none of the eight further repetitions we found ever brought any of them back. Instead the number got more confident with each retelling — rounded up, given a new unsupported comparison to Google and Yelp that appears nowhere in the original research, eventually presented as a personal discovery with no source at all.
Only one writer in the entire chain — BrightLocal — kept the hedge and linked the sources. Seven of the other eight dropped it entirely.
That’s not a story about any one author being careless. It’s a pattern: caveats are the first thing to go when a number is useful enough to repeat, and there’s no cost to being the one who repeats it wrong. If there’s a practical takeaway from this whole exercise, it’s not really about Foursquare — it’s that the next time you see a specific, round, widely-repeated statistic in this industry, it’s worth one search to see whether the trail leads back to a real measurement or just to another repost.
Predictions, scored
We wrote down what we expected before collecting anything, including our confidence in each guess. Here’s how they landed:
| Prediction | Confidence | Result |
|---|---|---|
| Foursquare share under 6% | 70% | Confirmed, far more decisively than expected — actual was 0.00–0.06% |
| Highest in food/nightlife, near-zero in home services | 80% | Inconclusive — every category measured zero; there was no gradient to observe |
| Higher in dense-urban markets | 65% | Inconclusive — same floor effect, every market tier measured zero |
| “Foursquare” named in text under 2% of responses | 75% | Confirmed — actual was 0.00% |
| ChatGPT’s picks won’t resemble Foursquare’s more than Google’s | 55% (genuinely uncertain) | Confirmed, and the effect ran the other direction — resemblance to Google was significantly higher |
We got two of five predictions half-right in an interesting way: we predicted a gradient between favorable and unfavorable categories, and instead found a floor — Foursquare’s presence was so uniformly low that there was no gradient left to detect. That’s a stronger result than what we predicted, not a weaker one, but it’s worth being honest that it wasn’t exactly what we wrote down in advance.
Scope note
This measures ChatGPT only, in the United States, in English, across a single 72-hour collection window in August 2026. It does not tell you anything about Gemini, Perplexity, Google’s AI Mode, or ChatGPT outside the US. It also can’t tell you what ChatGPT looked like before this window, or what it will look like after — which is, in a real sense, the entire point of this piece. The original 60–70% figure was accurate to a moment in time and a specific place. So is this one.
See the stats/study in context: Steady Demand Research Index
