The Foursquare claim we published in June does not hold up

The claim we ran in June

We published a number in June, and it has since been tested. It did not survive. The post said most of the local business names ChatGPT reads out come from a directory called Foursquare, and it told owners to go claim a listing there. We are taking that advice back.

If claiming a Foursquare listing is sitting on your marketing to-do list because you read it somewhere, cross it off. It costs about 20 minutes and a phone call, and nothing in this study, which tested ChatGPT in the US in August, supports the idea that it changes anything there. If you already did it, you have lost nothing. Leave it alone.

The number was real once, and it described something far smaller than what got repeated. Natzir Turrado, a search consultant, ran 50 prompts across five Spanish cities on 5 May 2025. He read the raw data ChatGPT sends back behind each answer, which is where the source of each business name shows up. In that sample, 60% to 70% of the businesses ChatGPT listed first came from Foursquare.

That finding arrived with three qualifiers, and all of them were gone within nine days. The study covered one country. It looked only at the business ChatGPT listed first, and it was collected in May 2025. By the time the figure reached an operator's to-do list, none of that was attached to it.

Turrado measured what he measured and reported it accurately for the sample he had. What failed was the way the number travelled after he published it.

Ben Fisher, who runs a local search company called Steady Demand, went looking for where the number started. He found ten write-ups. Two were the original pair from May 2025. Of the eight that repeated the figure afterwards, seven dropped every qualifier and one kept a hedge. Those ten are the ones his search turned up, so more of them very likely exist.

Our June post belongs on that list, and Fisher did not catch it. The post is still live this morning, with the claim sitting in the pull quote and no scope attached to it.

"Caveats are the first thing to go when a number is useful enough to repeat," Fisher wrote.

What the new test measured

In August 2026 Fisher tested the claim himself, and he built the sample to give Foursquare the best possible shot. He used 12 kinds of business across 12 US metros, and picked 7 of those kinds because Foursquare's data is strongest in them. Everything was collected inside one 72-hour window, so the study says nothing about which way things are moving.

That gave him 4,607 finished runs. In the 2,879 of them aimed at ChatGPT's main product, Foursquare turned up as a source ChatGPT cited or fetched 0.00% of the time. The count covers the visible links and also the pages ChatGPT pulled without ever showing them. It also covers the links hanging off the little business card ChatGPT draws when it names a company.

A second set of 1,728 runs, collected separately as a check, came back at 0.06%. That works out to a single citation, and the study says the true figure on both counts cannot be much above zero.

Fisher also logged which sources showed up on that business card across the same sample. Here is the whole set, measured on ChatGPT only, in the US, in one August window.

Read that table as a measurement and nothing more. Turning any single row of it into a to-do item would repeat the exact move this correction is about.

OpenAI announced a licensing deal for local answers with Yelp on 23 July 2026, four weeks before this data was collected. A licensing deal here means OpenAI pays for the right to use somebody's data, and this one does not lock anybody else out. The August picture may not be the picture six months from now.

Why zero citations is a smaller finding than it sounds

Fisher is careful about this. A count of 0.00% says Foursquare did not show up as a source ChatGPT cited or fetched in any of those runs. It falls well short of proving that Foursquare's data never reaches the model. The partnership between the two companies is real and nobody disputes it, and the part that failed testing is a single percentage somebody attached to it.

He built a separate test for the harder question of where the names come from. Across 137 matched city and category pairs, the businesses ChatGPT named overlapped Google's ranked list about 7.5% of the time, and Foursquare's about 5.7%. Both of those are small. Fisher reports the gap between them as real, and he still declines to say the door is shut.

He lists his own limits: one study, ChatGPT only, US only, English only, one 72-hour window. He published his predictions before he collected anything, scored them against his results afterwards, and marked 2 of the 5 inconclusive against himself.

What to do with the 20 minutes

Two things come out of this, and only one of them has anything to do with Foursquare.

The small one is the task itself. It has lost the number that justified it, so the 20 minutes goes somewhere else. A dental practice could spend it reading what an assistant says when a patient asks for a dentist who takes their insurance. An HVAC owner could spend it putting a price range in plain text on the page where customers go looking for one. Neither of those comes with a percentage attached to it, and we are not going to invent one.

The bigger one is a habit. When a marketer quotes you a precise-sounding number about AI search, ask what it was measured on, and when. The original write-up usually contains both answers, even if the version that reached you left them out.

Fisher adds a third question, and it is about age. "Assume a partnership number has a shelf life," he wrote. Deals get renewed and models get retrained, so a sample collected in the spring of last year describes a product that has moved on since.

Those questions would have stopped this number in nine days. It took more than a year, and it reached your to-do list through us.