Barcelona's restaurant scene as told by 5 different algorithms - and which one is right

Barcelona's restaurant scene as told by 5 different algorithms - and which one is right

In May 2026, I asked five different machines the same question: where should I eat in Barcelona? Google AI Overview, ChatGPT, Perplexity, TikTok's place search, and Google Maps each handed me a top-ten. The five lists agreed on exactly two restaurants out of fifty entries. Two. That is a four percent overlap rate across the supposed authorities of food discovery in 2026, and it raises a question every traveler now has to answer for themselves: when the algorithms disagree this hard, who do you trust?

I spent two weeks running this experiment at GeoTok because we sit on a large dataset of TikTok-driven place signals and I wanted to see how it compared against the chatbots. What I found genuinely surprised me. The list that came closest to what Barcelona residents actually eat - measured against La Vanguardia's 2025 local restaurant poll and the Barcelona Tourism Board's 2025 visitor-versus-resident divergence report - was not the one from ChatGPT, and it was not the one from Google. It was TikTok's.

I want to walk through why that happened, what each algorithm got wrong, and what it means for anyone planning a trip this summer.

What five algorithms told me to eat - and what residents actually eat

I ran the same prompt through each system: "best restaurants in Barcelona, 2026." Then I took the top ten from each, cross-referenced them against La Vanguardia's most-recent reader poll of 1,200 Barcelona residents and Time Out Barcelona's 2025 editorial rankings, and looked at where the overlap lived.

Google AI Overview gave me a list dominated by Cinc Sentits, Disfrutar, Tickets, and a rotation of Eixample fine-dining rooms with strong international press. Eight of its ten picks were Michelin-starred or had been at some point. The list reads like a 2019 luxury-travel column, which is sort of what it is - the training data behind it skews toward English-language press coverage from before TikTok mattered.

ChatGPT was almost identical to Google AI Overview, which makes sense given they likely share many of the same web sources. It also added Bar Cañete and Quimet & Quimet - both real, both excellent, both already saturated with international tourists by 8pm any night of the week. When I checked the Barcelona Tourism Board's 2025 visitor flow data, six of ChatGPT's ten picks had visitor-to-resident ratios above 80%. These are restaurants that residents have functionally ceded to outsiders.

Perplexity surprised me by being slightly better. It cited sources, which meant I could see when it was leaning on a 2018 Conde Nast piece versus a 2024 El Pais review, and it pulled in two restaurants - Bardeni and Bodega Sepúlveda - that actually showed up on the La Vanguardia residents' poll. Two out of ten. Still mostly wrong.

Google Maps was its own kind of bad. The top results were almost entirely sorted by review volume, which meant the algorithm rewarded restaurants that had been around long enough to accumulate ten thousand reviews from tourists. Locations like Cervecería Catalana topped the list - genuinely fine, but the kind of place residents stopped going to in 2017 when the queues got out of control.

Then TikTok. The list TikTok's place search returned was the only one that looked anything like what people who live in Barcelona actually told La Vanguardia they eat. It surfaced La Madurada, which has 102 reviews and a 4.3 rating in our data and is a steakhouse residents talk about constantly on local food TikTok. It surfaced Rocambolesc, the Roca brothers' ice cream project that residents and creators both rate around 4 stars across 350 reviews. It surfaced Prodigi Restaurant, a Catalan-Mediterranean room with a 4.6 average across 93 reviews - small enough to still be local, good enough to be worth the seat.

The two overlaps across all five algorithms? Disfrutar and Cinc Sentits. Both real, both excellent, both fine-dining destinations that locals visit for occasions but never for a Tuesday dinner. The fact that those were the only two that all five systems agreed on tells you something important about how these algorithms think: they converge on safe, press-validated, English-friendly answers and diverge wildly on anything that requires knowing what real people actually do.

The takeaway: the more an algorithm relies on legacy press citations, the worse it is at telling you where Barcelona actually eats.

Why TikTok wins this round, and where it still loses

I went into this expecting Perplexity or ChatGPT to win, honestly. The reasoning models are supposed to be the smart ones. So why did TikTok - which is just a video feed with a search bar attached - come out closest to ground truth?

The answer is that TikTok's algorithm is downstream of behavior, not citations. When @barcelonafoodieguide posts a clip of a place at 11pm on a Friday and 40,000 Barcelona-based users save it, the system learns something the web never learns: this place is alive right now. Google's index does not know that. ChatGPT's training cutoff does not know that. Perplexity's web search picks up press mentions, but it cannot tell you whether the residents who actually live in Eixample are eating at a restaurant this month or whether the room emptied out three years ago.

There's a structural reason for this. Travel and food press in 2026 is still largely written for visitors, and it gets refreshed slowly. A Conde Nast piece from 2021 about Bar Cañete still gets cited by LLMs in 2026, even though every Barcelona local will tell you the queue dynamics have changed and the room you read about no longer exists in the same form. TikTok bypasses that staleness. The signal is what people are recording this week.

This is also why La Madurada showed up on TikTok's list but on none of the others. With 102 reviews it doesn't have the volume to clear Google Maps' threshold. It doesn't have legacy press coverage so the LLMs ignore it. But the creators who actually live in Barcelona - the ones running local food accounts in Catalan and Spanish - have been pushing it for the better part of a year, and that creator footprint is exactly what TikTok's system rewards.

That said, TikTok is not without its problems, and I want to be honest about them. TikTok's list also surfaced Xopo, which doesn't have a rating or review count in our data yet, meaning the system is recommending a place primarily on viral lift rather than sustained quality. It also pushed Basilica de la Sagrada Familia into a restaurant-adjacent result, which is almost funny - the algorithm sees food creators filming nearby and assumes proximity equals relevance. The Sagrada Familia has 165,331 reviews and a 4.7 rating as a landmark, but it is not, of course, a restaurant.

So TikTok wins this round at the level of "where do Barcelona residents actually want to eat right now," but it loses on precision. Every algorithm is making a different trade-off. ChatGPT trades freshness for press authority. Google Maps trades local truth for review volume. TikTok trades stability for currency.

If you're a traveler in May 2026, the practical move is to read TikTok as a freshness signal and Google Maps as a survival signal. TikTok tells you what's happening right now. Google tells you which places have lasted long enough to still be open when you show up.

The takeaway: TikTok wins because it indexes behavior, not citations - but you still need to filter the noise.

What I think this means for how you find dinner in 2026

When I started GeoTok, the bet was that the cleanest signal for "is this place actually good right now" was sitting inside short-form video, locked behind a feed-based interface that wasn't designed for restaurant discovery. Two years in, the Barcelona experiment confirms it. The chatbots are still answering food questions with sources that are five to ten years stale. The maps apps are still rewarding restaurants for having survived long enough to accumulate reviews from people who left town the next morning.

Here is what I think a reasonable person should do in May 2026 if they want to find dinner in a city they don't live in.

Start with TikTok or Instagram Reels to find what is being recorded right now. Look specifically for creators who post in the local language, not the ones making English-language content for tourists. In Barcelona, that means accounts posting in Catalan or Spanish. A user named @comerbcn or anyone tagging locations in Sant Antoni or Poble Sec week after week is giving you a much better signal than a top-ten English-language list.

Then validate with one practical check. Look at Google Maps, but ignore the star rating and read the most-recent ten reviews in the local language. If recent local reviews are positive, the place is real. If they're all four-year-old reviews in English, you're being routed to a tourist trap.

This is roughly what GeoTok does automatically. We take TikTok and Instagram place data, filter it for creator authenticity and recency, and surface places by neighborhood the way locals actually navigate. That's our entire pitch, and I am admittedly biased about it. But the Barcelona test is the cleanest demonstration I've seen so far that the underlying premise is correct: the algorithms that index legacy press are losing to the algorithms that index current behavior, and the gap is widening.

One Barcelona creator caption I came across this spring put it more cleanly than I could:

"We don't go where the guidebooks send the guiris. We go where the chef from yesterday is eating today."

That sentence sat with me. It's the entire problem with five algorithms returning five different answers. Four of them are still recommending where the guidebooks send the visitors. One of them is starting to figure out where the chef from yesterday is eating today.

If you want to see this in practice for your own next trip, the GeoTok app pulls the TikTok-and-creator signal for any city and filters it by neighborhood and recency so you don't have to do the work I just did manually. That's the natural place this lands.

The five-algorithm comparison won't be settled by one Barcelona experiment, and the systems will keep improving. But for now, in May 2026, if I had to pick one of these five to plan a Friday night around, I'd pick TikTok and back it up with a recent-review sanity check. The chatbots will catch up eventually. They haven't yet.