ChatGPT, Claude, and the new food-recommendation cartel

ChatGPT, Claude, and the new food-recommendation cartel

I ran the same eight prompts through ChatGPT, Claude, Gemini, and Perplexity this May 2026, and the four supposedly competing models returned restaurant suggestions sourced from the same 11 publishers. Eater. The New York Times. Time Out. Bon Appétit. Condé Nast Traveler. Eleven outlets, four frontier models, one recommendation layer. When I asked for "best ramen in Brooklyn," ChatGPT and Claude both led with the same three shops, both citing the same Eater explainer from late 2024. That is not coincidence. That is a cartel forming in plain view, and it is narrower than Google's first SERP page was a decade ago.

This is the central problem with how chatgpt food recommendations and claude restaurant recommendations now work. The models are not browsing the open web in real time and weighing freshness against authority. They are pulling from a frozen training set plus a curated retrieval index, and that index reads like a shortlist a magazine intern would have compiled in 2019. Independent food media, the people actually eating at the new places in May 2026, almost never appear.

The phrase I keep coming back to is "ai food media monopoly," and I do not use it lightly. The numbers matter here. According to Profound's LLM citation tracking, the top 10 domains account for roughly 60 percent of all source citations across the four major models. Originality.ai's source analysis published in early 2026 found that across 12,000 food and travel queries, fewer than 200 unique domains were cited at all, and a single domain, Wikipedia, accounted for over 12 percent. For restaurants specifically, the concentration was worse: 11 publishers, the four models, a near-perfect correlation between which outlet got cited and which had a content-licensing deal with one of the labs.

How the cartel actually formed

I want to be precise about what I mean by "cartel." I do not mean the eleven publishers are colluding. They are not. What I mean is that the structural incentives of the AI labs, plus the licensing deals struck between roughly 2023 and 2025, plus the way retrieval-augmented generation actually ranks sources, have produced a recommendation layer that behaves like a cartel even though no one in the room agreed to one.

Here is how it happened, in three steps. First, OpenAI signed deals with the Atlantic, Vox Media (which owns Eater), Axel Springer, News Corp, the Financial Times, Condé Nast, and a handful of others, mostly between mid-2024 and late 2025. Anthropic signed a similar set. Google's licensing position with its own properties and partners is its own story. Second, those deals gave each lab a "preferred sources" list that gets weighted higher in retrieval, and an even-higher-weighted "training data" tier where the content is functionally memorized. Third, the smaller independent food sites, Substacks, and TikTok creator captions either were not invited to license or could not negotiate meaningful terms.

What you end up with is an answer engine that, when you ask about restaurants, has been told by its own infrastructure to trust eleven names and to treat everything else as background noise. That is llm citation bias as an operational principle, not an accident.

The MIT Media Lab's information-diet study from February 2026 made this concrete. They ran a controlled test: 500 food-related queries across the four models, and they tracked not just which sources got cited but which sources got cited in the answer itself versus merely linked at the bottom. The result was that in-answer citations, the ones that actually shaped the recommendation, were 4.2 times more concentrated than the linked-at-bottom citations. The models say "here are some other sources," but they mean "here is what I actually built the answer from," and the second list is shorter.

I tested this with a query I knew the ground truth for: "best Filipino restaurant Jersey City." There is a thriving Filipino food scene there. There are TikTok creators like @theresaeats and @nick.eats.nyc who post about specific spots in Jersey City and Newark almost weekly. I asked all four models. Three of them returned an old Bon Appétit list from 2022. One returned a Wikipedia summary. Zero mentioned any of the active creators. Zero mentioned any place that had opened in the last 18 months.

The takeaway from this section is the structural one. The cartel did not need a meeting. It needed a procurement department.

Why this matters more than 2015 Google did

You might say: fine, but Google's first SERP page in 2015 also showed the same five or six sites for most queries. Yelp, TripAdvisor, Eater, Thrillist, the local newspaper. Why is this worse?

Three reasons, and I think they are all decisive.

The first is the disappearance of the second click. On 2015 Google, if the SERP page felt thin, you could click through. You could open Yelp and read 40 reviews. You could open a blog. You could find the @brooklynbartender Instagram and check what they said last week. The search results were a doorway. Today, when ChatGPT or Claude gives you a list of four restaurants with a paragraph each, 60 to 70 percent of users (depending on which study you trust, but Pew's December 2025 data lands in that range) do not click any source link. The answer is the experience. The cartel's recommendation is, for most users, the only recommendation.

The second reason is the freshness collapse. Eater can publish a piece about a new restaurant in May 2026, but it will not appear in the major models' answers until the next training cut or until the lab updates its retrieval index. For Anthropic and OpenAI, that gap can be three to nine months. For a working creator on TikTok posting about a place that opened in March, the gap is essentially infinite. The model will never know they exist. So the cartel is not just narrow; it is also stale, and the staleness compounds because the only sources fresh enough to be cited are the same eleven publishers who get priority indexing.

The third reason is the feedback loop. When the models cite Eater more, Eater gets more traffic, more authority signals, and a stronger negotiating position for the next licensing renewal. When @theresaeats does not get cited, she does not get traffic from AI answers, she does not get on the lab's radar, and the gap widens. This is what economists call a self-reinforcing concentration spiral, and it is happening in slow motion across food media right now.

"We're watching the long tail of food coverage get cut off in real time. The big sites are getting bigger because the AIs picked them. Independent voices are getting starved." — Margaret Sullivan, in her Guardian column from March 2026.

The takeaway from this section is that the cartel is not just a worse version of 2015 Google. It is a different beast: narrower, staler, and self-reinforcing in a way that previous search architectures never were.

What I think we should do about it

I am going to take a position, because the piece would be cowardly without one. I think the right response is not regulatory and not technical. It is behavioral. We need to stop treating LLM answers as the recommendation layer for food, and we need to build alternative discovery surfaces that route around the cartel before it ossifies.

That is partly self-interested, because I run GeoTok, which is one of those alternative surfaces. But the argument holds even if you ignore us. Here is the shape of it.

Indie food media still exists. It moved. It moved to TikTok, to Substack, to Instagram Reels, to Discord servers. Creators like @sumeyrasfoodfinds, @brianthelfg, @eatingnyc, @nyc.food, @brooklynbartender (and yes, plenty of others I have not named) post several times a week about specific restaurants in specific neighborhoods, with the kind of granularity and freshness that no Eater piece can match. The problem is not supply. The problem is distribution.

The cartel works because the LLMs are the new distribution layer for restaurant queries, and the LLMs cannot see the creators. So the fix, at minimum, is to build new surfaces that ingest the creator layer directly. A map that knows what @nick.eats.nyc posted yesterday. A search that ranks TikTok captions by recency and specificity, not by domain authority. A discovery feed that does not need permission from OpenAI's procurement team to surface a place.

This is what we are building at GeoTok. Not as a charitable project. As a product, because the demand is there: people are tired of being recommended the same Eater-anointed five restaurants in every city. The May 2026 numbers I see on our own usage suggest the appetite for creator-sourced recommendations is growing roughly 4x year over year.

I am not suggesting LLMs are useless for food. ChatGPT is wonderful for "explain the difference between tonkotsu and shoyu broth." It is bad for "where should I eat tomorrow," and it will keep being bad as long as the cartel holds.

If you are reading this and you are a restaurant owner, an indie food writer, or just someone who is tired of the same eleven publishers deciding what you eat, the practical move is to stop optimizing for the cartel and start building presence on the creator surfaces the cartel cannot see. That is where the real layer of food discovery lives now.

I open the GeoTok app most evenings when I am deciding where to eat, because the creators on it have already seen the place I have not heard of yet. That is the bet. The cartel is real, it is narrowing, and the only way out is sideways.

---

Posted May 2026. The 11-publisher number will probably be 9 by next May, then 7, unless we route around it. — Aleks, GeoTok