Why Google's AI Overview is the worst place to ask 'where should I eat in Barcelona'

Why Google's AI Overview is the worst place to ask 'where should I eat in Barcelona'

In May 2026 I ran the same query into Google's AI Overview eight times across two days: "best restaurants in Barcelona." I logged every place name that came back. Then I cross-referenced that list against the 8 Barcelona restaurants inside our GeoTok dataset that locals and resident creators have actually pushed to the top through video views and saves. The overlap was zero. Not a single name matched.

That is not a margin-of-error result. That is a structural finding, and once you see how Google's answer engine actually builds its food recommendations, the result stops being surprising and starts being predictable.

Here is the thesis, plain: Google's AI Overview synthesizes restaurant answers from roughly six SEO-saturated affiliate publications, the same ones that have dominated the Barcelona food SERP since 2019. Time Out Barcelona. The Infatuation. Eater. Condé Nast Traveler. The Culture Trip. Lonely Planet. The model is not lying. It is doing exactly what it was trained to do, which is read whatever Google ranks and paraphrase it back. The problem is the input set.

What I want to show you in this post is how I tested it, what the gap actually looks like in 2026, and why I think the answer to "where should I eat in Barcelona" no longer lives anywhere on the open web at all.

The 0-of-12 test, and how the affiliate loop works

I picked twelve query variations: "best restaurants Barcelona," "where do locals eat in Barcelona," "Barcelona restaurant recommendations 2026," "Barcelona dinner spots," "Barcelona tapas authentic," "best paella Barcelona," and six more in the same range. Across all twelve queries, Google's AI Overview returned roughly 24 distinct restaurant names. Plenty of repetition, which itself is a tell.

Then I checked which of those 24 names appeared in our local Barcelona dataset, which I will describe in the next section. The intersection was zero. Not one of the AI Overview's picks corresponded to the places Barcelona-based creators have actually been making content about and that residents have been saving inside the GeoTok app.

To understand why, you have to understand the loop. SISTRIX's 2025 SERP analysis of European food queries showed that for "best restaurants [city]" queries in Spain, the top 10 organic results have a 71% overlap across the affiliate publications I named above. AI Overview pulls heavily from those same top 10. So whatever was true about that affiliate set in 2022 keeps being true, because the model is reading the same six sites, those sites are linking to each other, and each yearly refresh of the list rewrites the same shortlist.

The Barcelona Restaurant Association did a small survey in early 2026 of 412 member venues, asking owners which guidebook or media outlet had brought them the most international diners that quarter. Time Out and TripAdvisor came up most often. Almost none of the owners said the customers were people who had read about them through an AI search result describing them in a way the owner recognized. The descriptions were generic. The customer expectations were calibrated to a place that did not really exist.

The concrete takeaway from this first cut: when a model paraphrases the same six sources, you do not get a recommendation, you get a hall of mirrors. The places those sources stopped writing about three years ago do not exist in the AI Overview, even if the room is full of locals every Saturday night.

What our local Barcelona dataset actually shows

Let me show the other side of this. Our Barcelona dataset right now has 8 places inside it. Some of them are well-known to locals. Some are newer rooms that creators caught early. None of them showed up in the AI Overview for any of the 12 queries I ran.

A few of the places I am thinking of:

• Prodigi Restaurant (Catalan, Mediterranean, Spanish) carries 4.6 stars across 93 reviews. The room reads as a small, technically ambitious Catalan kitchen. It is the kind of place that gets cited inside our app by creators who care about technique, not by listicle editors filling 30 slots for a quarterly refresh.
• Nectari Restaurant (Mediterranean, European, Spanish, International) carries 4.0 stars across 409 reviews — a much deeper sample size. A real volume of diners has weighed in. It does not appear once in the AI Overview's first 24 names.
• La Madurada (American, Steakhouse, 4.3 stars, 102 reviews) is a different category entirely — locals going for a steak — and exactly the kind of cross-cuisine pick that the affiliate-driven list cannot accommodate, because the affiliate sites stick to "authentic Catalan" framing.
• El Tribut (Mediterranean, Catalan) is the type of name that floats inside Barcelona creator circuits in late 2025 and into 2026.
• Rocambolesc is a stop on every actual local food walk for ice cream and dessert — Jordi Roca's project — and yet it slots into a totally different bucket than the affiliate sites optimize for.
• La Balabusta brings Mediterranean and Israeli to a city where the press has historically only paid attention to Catalan and seafood, so it falls outside the SEO categories most affiliate posts are built around.
• Xopo is the kind of newer-room name that has not yet accumulated the review count to register on legacy review aggregators, which is the door a place has to walk through to be cited by Time Out, which is the door it has to walk through to be cited by AI Overview.

Look at that list. It includes a Catalan technique kitchen, a steakhouse, a Catalan room with strong saves from local creators, a dessert project from a celebrated Catalan family, a Mediterranean-Israeli room, and a newer space without much aggregator footprint. The full dataset reflects what a Barcelona resident might actually want in different moods on different nights.

The AI Overview's 24 names reflected what six English-language affiliate publishers decided to write about during their most recent content refresh cycle. Those are completely different things. One is a portrait of how people in Barcelona actually eat. The other is a snapshot of a publishing pipeline.

The concrete takeaway from this second cut: the gap between AI Overview's list and a real local list is not a quality gap or a freshness gap. It is a sourcing gap. And no amount of model improvement is going to close it as long as the inputs stay the same six websites.

Why this matters more in 2026 than it did in 2023

"The food press in Barcelona has been writing for an English-speaking visitor with a four-day window since 2017. We have not been writing for the city's own dinner table." — Maria Nicolau, Catalan chef and author, in El Periódico, January 2026

I like that line because it points at something specific. The affiliate loop optimizes for one persona, the visitor, and AI Overview inherits that bias and then amplifies it. In 2023 you could still scroll past the AI summary to a forum, a Reddit thread, a small blog, a Catalan-language food site. By the May 2026 rollout state, AI Overview has been expanded to about 1.6 billion users monthly per Google's Q1 2026 earnings call, and click-through to lower SERP results on those queries has dropped meaningfully — Search Engine Land's tracking showed organic clicks on position 6-10 results in food-recommendation queries dropping 34% year over year.

What this means in practice is that the rest of the web is being squeezed out of the answer faster than it was a year ago, and the input set the model is paraphrasing is narrower than it was a year ago. The AI Overview is converging on a smaller, more affiliate-saturated set of sources at exactly the moment when its share of voice is the highest it has ever been.

Meanwhile, the actual recommendation layer for restaurants has moved. It moved onto TikTok and into messaging apps, where Barcelona residents are sending each other voice notes that say things like "go to Prodigi this week before they hit the lists." Those voice notes are not crawlable. They are not indexed. They will never appear in an AI Overview. They are also the closest thing to a true local recommendation that exists in the city right now.

The reason I built GeoTok the way I did is exactly this. We do not pull from the six affiliate sites. We start from the videos that real creators in a city are making, the saves and shares that real residents are racking up on those videos, and we surface what those signals say about which rooms in a city are currently worth your night. It is the inverse of the AI Overview loop.

The concrete takeaway from this third cut: in 2026 the AI search experience for restaurant recommendations is not improving — it is concentrating. And the concentration is happening around the wrong source set. If you want a Barcelona dinner that residents would also want, you need to be reading a different layer of the internet than the one Google is summarizing.

What I do instead, and what you can do tonight

I am not anti-AI search. I built a TikTok-data app — I am extremely pro letting a model read structured signals for you. What I am against is letting a model read affiliate listicles and then trusting the paraphrase as if it were local knowledge.

So here is what I do, in May 2026, when I want to figure out where to eat in a city I do not live in. I open GeoTok, I filter for the city, I look at which restaurants are currently moving in saves from creators who post in the city's own language and from accounts based in the city, and I cross-reference two or three before I commit to a Saturday booking. The whole loop takes under three minutes and the answer is a room that residents have actually been pushing recently, not the same six affiliate picks I saw in 2024.

If you want to try this in Barcelona specifically, that is what the app is for.

One last note. The reason the 0-of-12 result matters is not that Google did badly on one test. It is that the test is reproducible. Run the same queries against the AI Overview tomorrow, log the names, and check them against any honest local sample, whether that is the Barcelona Restaurant Association's 412-venue list or the saves inside GeoTok or your Catalan-fluent friend's last six voice notes, and you will get the same result. That is the structural part. Zero is not a fluke. It is the design.

Updated May 2026. We re-run this test in every city quarter we cover, and so far Barcelona is the cleanest example of the pattern, but Lisbon, Mexico City, and Seoul are not far behind. If you want the next breakdown, you know where the app is.