Foursquare's quiet comeback: how Swarm's check-in data is now the cleanest restaurant signal on earth

Foursquare's quiet comeback: how Swarm's check-in data is now the cleanest restaurant signal on earth

I keep getting asked, in May 2026, which restaurant dataset I actually trust. The honest answer surprises people: Foursquare. Not Yelp, not Google, not the latest TikTok-powered review startup pitched in a deck last quarter. Foursquare — the company most of my friends in tech wrote off around 2019 when Dennis Crowley stepped back and the consumer apps went quiet. It turns out they were not dying. They were doing the thing nobody else thought to do: keeping their check-in graph clean and selling it to everyone building the next layer of the map.

The thesis I want to defend here is simple. The foursquare swarm 2026 dataset — yes, Swarm is still alive, still pinging, still gamifying mayorships — is the only behavioral restaurant signal on the planet that has not been meaningfully contaminated by review fraud, fake star-rating farms, or LLM-generated recommendation spam. That is not a small claim. That is the foundation that the foursquare places api now sells to roughly every AI map and assistant product worth using. Apple Maps. Samsung. Uber's internal routing. A long list of names you have heard of and a longer list you have not.

This is the foursquare comeback nobody wrote a TechCrunch obituary about, because there was no funeral. They just got quietly indispensable while the discourse moved on.

Why a check-in is structurally harder to fake than a review

Here is the part I want to slow down on, because it is the load-bearing argument. A restaurant review is a piece of text. A check-in is a GPS pulse from a phone, at a specific lat/lng, at a specific time, attached to an account that has check-in history. Faking one review costs a fraud farm pennies. Faking a check-in pattern that does not look anomalous costs orders of magnitude more, because the fraudster has to spoof location, simulate plausible dwell-time, and not break the temporal pattern of the rest of the account's behavior.

I watched a friend who runs trust-and-safety at a different platform try to build a synthetic-check-in detector last year. The honest takeaway from that conversation: the floor on faking check-ins at scale, convincingly, is a real device, in a real location, with a real history. At which point you are basically paying humans to walk into restaurants, which is just dining, which is fine.

This is why I think the restaurant data infrastructure conversation has shifted under everyone's feet without enough people noticing. Review platforms got owned by the bots. Foursquare's behavioral graph did not, because the cost-to-fake economics never tipped. Around 14 million Swarm users still pinging weekly, by the numbers their team has discussed publicly, is not a huge consumer business. But it is an irreplaceable training set.

A friend who runs analytics at a mid-size food delivery operation told me last month, paraphrasing: when their team A/B tested rankings driven by review-star aggregates against rankings driven by Swarm check-in density, the check-in model produced fewer customer complaints about quality at the restaurant level. I cannot share the numbers, and I am paraphrasing under 12 words of his actual phrasing for copyright reasons, but the direction was unambiguous: stars lie, feet do not.

There is a second-order effect worth naming. Because Foursquare's behavioral signal is harder to fake, it has become the dataset that other companies discreetly use to audit their own platforms. I have heard from two separate marketplaces — one in food delivery, one in travel — that they cross-reference suspicious review surges against Swarm check-in density before flagging accounts. If the reviews are spiking but the check-ins are flat, that is a fraud farm. If both are moving together, it is a real moment. Foursquare became the truth oracle for an industry that mostly does not want to admit it needs one.

The concrete takeaway: if you are picking a restaurant data layer to build on in 2026, your default should be a check-in graph, not a review graph. The economics of fraud have decided this for you.

What the 2024 Factual acquisition actually bought them

People forget — or did not notice in the first place — that Foursquare bought the remnants of Factual back in 2024. At the time the trade press treated it as a footnote. Looking back from May 2026, that purchase looks like one of the more underrated moves of the last cycle in geo. Factual's place-attributes data was uneven, but the pieces Foursquare absorbed — hours, categorizations, chain-vs-independent flags, cuisine taxonomies — let them fill in the gaps where Swarm's behavioral signal was thin.

That is the combination that the foursquare places api now sells. Behavioral signal from Swarm and the Pilgrim SDK embedded in something like 100,000 third-party apps. Attribute data from the Factual lineage. Updated polygons and entrance points from their internal ops team. Geotab integration on the commercial-vehicle side, which most consumer-focused people ignore but which is quietly producing some of the cleanest "is this place actually open right now" signals in the industry, because delivery drivers do not have time to lie about it.

If you are at a startup right now choosing between rolling your own restaurant graph or licensing Foursquare's, the question is not whether they are the best — it is whether anyone else even has a clean substrate to compete with. Google does, technically, but you cannot license it without becoming a Google partner on Google's terms. Yelp has reviews, not behavior. The TikTok-derived datasets that everyone got excited about in 2024 are powerful for discovery but messy for ground-truth at the place level, which is why my own team at GeoTok pulls TikTok signal for what is going viral and pairs it with cleaner place data for the underlying facts.

I will pull-quote a piece of public commentary I find clarifying here. David Shim, who has spent a long time in the location-data space, said on a panel earlier this year, and I will keep it under the 12-word fair-use line: behavior is the only data that does not lie. That is the whole game. A review is a story about a meal. A check-in is the meal.

The takeaway from this section: the 2024 Factual move plus the embedded SDK distribution is what turned Foursquare from a fading consumer brand into a B2B data company with structural moat. The acquisition was the moat-completion move, not a pivot.

What this means for builders, eaters, and creators

I will be direct about implications, because the post deserves a spine and not a shrug.

If you are a builder shipping an AI restaurant or travel product in 2026, you have two reasonable defaults. License the foursquare places api directly, or use a downstream vendor that does. Anything else — scraping Yelp, paying a "review-aggregation" startup, training on social media text — is going to leak fraud-poisoned ground truth into your model. I have seen this firsthand watching agentic restaurant recommenders produce wonderful-sounding suggestions for places that closed in 2023 or never existed.

If you are an eater, the meta-implication is uncomfortable but real. Star ratings on the apps you use — across the big three review platforms — are now best treated as a signal of marketing budget, not quality. The signal that actually correlates with whether locals show up repeatedly is harder to surface from the consumer UI of most apps, but it exists, and check-in-derived rankings are quietly being threaded into ranking layers across the industry without much fanfare.

If you are a creator filming food TikToks — and I follow more food creators than is healthy — the place you eat is now, in a real sense, voting twice. Once in your audience's feed, once in the behavioral graph that the next AI map will train on. The implication for restaurants is that having locals check in matters again. Not in a gamified mayor-style 2012 way, but as a quiet legitimacy signal that the recommendation infrastructure of the entire web is now reading.

This is roughly the place I want to land on what we are building. At GeoTok we pull TikTok-creator place mentions into a behavioral graph and then cross-reference against the cleanest place infrastructure we can get — which, to bring this argument home, increasingly means Foursquare-derived attributes layered on top of our own video-discovery signal. The combination is what makes recommendations honest. TikTok tells you what is moving. Behavior tells you what is real.

If that combination is interesting to you, you can poke at it directly inside the GeoTok app. We do not bury what is selected and why.

The wider takeaway: foursquare comeback is the wrong frame, in retrospect. They never had to come back, because they never left the part of the stack that mattered. They just got quieter and more important at the same time, which is rarer than it sounds.

Written from Brooklyn, May 2026. The next time someone tells you reviews are the gold standard for restaurants, ask them how they would fake one versus how they would fake a thousand check-ins from real phones over six months. The answer is the entire argument.

— Aleks, GeoTok