Highlights
- The dataset: This report draws on over 200,000 AI source citations, collected with prompts specific to the restaurant and coffee shop industry. The underlying AI responses, from four AI models, were gathered in 12 countries throughout Q2 2026.
- 56% local in Sweden vs 19% in Argentina: The same Spanish-speaking question lands very differently by market.
- ChatGPT cites sources in just 28% of its answers: The thinnest citation behavior of the four models.
- ~79% of top sources differ: Across all six model pairs, range 14% to 29%.
- 69% of Microsoft Copilot's citations are from commercial domains: The most marketplace-heavy model by a wide margin.
- 43% of Grok's citations are from editorial domains; 34% user-generated: Reviews and community threads.
Which sources should you target to get cited as a restaurant or coffee shop?
A small set of giants recurs across the models, but each model weights them differently. tripadvisor.com, instagram.com, guide.michelin.com, reddit.com, and yelp.com show up repeatedly; what changes is the supporting cast.
| Model | Top 5 cited domains | Citation character |
|---|---|---|
| Grok | tripadvisor.com, guide.michelin.com, instagram.com, reddit.com, yelp.com | Reviews, fine-dining guides, community |
| Microsoft Copilot | instagram.com, facebook.com, tripadvisor.com.mx, es.restaurantguru.com, tripadvisor.es | Social + local directories |
| Google AI Overview | google.com, instagram.com, reddit.com, tiktok.com, tripadvisor.ca | Own surfaces + social video |
| ChatGPT | reddit.com, timeout.com, thatsup.se, cntraveler.com, tripadvisor.es | Community + editorial food media |
Grok is the only model anchored by a fine-dining guide, guide.michelin.com is its second most-cited domain. Microsoft Copilot's top list is almost entirely social platforms and localized restaurant directories, reflecting its heavy commercial tilt. Google AI Overview is the only model whose own properties (google.com surfaces, youtube.com, TikTok) sit at the very top. And ChatGPT leans on reddit.com more than any single publisher, a Reddit thread is more likely to feed a ChatGPT restaurant answer than any review site.
How have AI source rankings changed over time in the restaurant and coffee shop industry?
Splitting the roughly 16-week window at its midpoint (2026-05-04) and re-ranking the most-cited domains in each half surfaces clear movement at the top of the list. We read this as relative position within each half.
Reddit rose from #4 to #1 and is cited across all four models. TripAdvisor fell from #1 to #4, and guide.michelin.com from #3 to #5, while yelp.com, thefork.com, theworlds50best.com, and wanderlog.com dropped out of the top-25 entirely. Localized TripAdvisor editions surged into the top tier in the second half: tripadvisor.com.mx, tripadvisor.com.ar, tripadvisor.ca, tripadvisor.cl, and tripadvisor.fr. We report these as relative rank movements within the window rather than a firm ranking trend.
What type of content do AI models cite for Restaurants & Coffee Shops?
Classifying every cited domain by category exposes the sharpest behavioral split in the dataset. The shares below are of each model's total cited content.
| Model | Commercial | Editorial | UGC | Institutional | Reference | Other | Unclassified |
|---|---|---|---|---|---|---|---|
| Microsoft Copilot | 69.4% | 6.5% | 11.6% | 0.5% | 0.9% | 1.1% | 10.0% |
| Grok | 14.7% | 43.1% | 34.2% | 2.2% | 1.3% | 0.3% | 4.1% |
| Google AI Overview | 29.6% | 29.2% | 31.5% | 3.1% | 1.1% | 1.1% | 4.4% |
| ChatGPT | 9.0% | 40.1% | 35.8% | 4.1% | 5.3% | 0.4% | 5.4% |
Three patterns stand out. Microsoft Copilot is the commercial model by a mile, nearly 70% of its citations are commercial domains (restaurant directories, booking platforms, brand sites), roughly double the next model and seven times ChatGPT's share. Grok and ChatGPT are the editorial-and-community models: each sends more than 40% of its citations to editorial food and travel media and another third to user-generated content like Reddit. Google AI Overview spreads its citations most evenly. For a brand, a directory or booking-platform listing matters far more for Microsoft Copilot visibility, while earned coverage in food media and an active Reddit presence carry the most weight for Grok and ChatGPT.
Do AI models cite local-language content for Restaurants & Coffee Shops?
Using a country-code top-level-domain (ccTLD) proxy, counting how often a cited domain ends in the local country's TLD, we measured how "local" each market's citations are. This understates true local-language content, because plenty of local publishers use .com; treat these as a floor, not a ceiling. English-language markets are excluded.
The same Spanish question produces different citation geographies depending on the country setting, Spain, Mexico, and Argentina all skew strongly toward .com, suggesting global platforms (TripAdvisor, Instagram) and the localized TripAdvisor .com-family editions crowd out purely local publishers in Spanish-language restaurant search.
Which AI model relies most on local sources for Restaurants & Coffee Shops?
Splitting localization by model shows the market-level averages hide a strong per-model effect. Here Microsoft Copilot relies on local sources the most; Grok defaults to global .com domains almost everywhere.
Microsoft Copilot cites local-country domains between 38% and 72% of the time across the six languages, by far the strongest localization in the dataset, consistent with its heavy reliance on local restaurant directories. Grok sits at the bottom in most languages, citing local domains under 10% of the time in French, German, and Dutch. Google AI Overview's German (n=38), Dutch (n=15), and Italian (n=28) cells rest on tiny samples and are directional only; it had no French or Swedish citations in scope. A brand serving non-English markets should expect Microsoft Copilot to surface local directories and publishers, while Grok will favor large international platforms regardless of country.
Should Restaurants & Coffee Shops optimize for each AI model separately?
Yes, almost entirely. We took each model's 20 most-cited domains for the restaurants and coffee shops category and measured how much those lists overlap, pair by pair. The average overlap across all six model pairs is 21.5%. Pooled across pairs, shared domains account for 21.2% of the combined set, so about 79% of top-cited domains differ between any two models.
The overlap is lopsided. The three search-grounded models, Google AI Overview, Microsoft Copilot, and Grok, cluster loosely around a shared core of TripAdvisor, Instagram, and the Michelin Guide. ChatGPT sits apart, sharing barely a seventh of its list with any of them. A presence strategy built around one model's preferred domains leaves a brand largely invisible in the others.
How many sources does each AI model cite per answer?
Counting cited sources per response (among responses that cite at all) reveals a large gap in citation appetite.
| Model | Mean sources / response | 95% CI | Responses with citations |
|---|---|---|---|
| Grok | 27.2 | 26.5–28.0 | 3,789 |
| Microsoft Copilot | 14.0 | 13.9–14.2 | 5,620 |
| Google AI Overview | 13.9 | 13.5–14.3 | 973 |
| ChatGPT | 5.8 | 5.6–5.9 | 1,804 |
Grok cites roughly 4.7x as many sources per answer as ChatGPT and about 2x as many as Microsoft Copilot or Google AI Overview. Two practical implications follow. First, the long tail of Grok's citations is enormous, a brand has many more "slots" to win in a Grok answer than in a ChatGPT one. Second, ChatGPT cited sources in only 1,804 of its 6,430 logged responses for this category (about 28%), most ChatGPT answers about restaurants are generated from model knowledge without surfacing citations at all, which is why its citation sample is the thinnest in this study.
Context
This analysis draws from Temso's AI visibility monitoring platform, which tracks how brands appear in AI model responses across ChatGPT, Microsoft Copilot, Grok, and Google AI Overview. The dataset covers restaurants-and-coffee-shops prompts, where to eat, best restaurants, top cafes and coffee shops, and dining recommendations, across 12 country markets and 7 languages over roughly sixteen weeks in spring 2026.
It measures what AI models cite, not clicks, conversions, or what users actually read. The localization measure is a ccTLD proxy and undercounts local publishers who use .com. ChatGPT cites sources in a minority of its answers, so its sample is smaller and its patterns should be read as suggestive. All figures are observational.
Methodology
How we measured this
We analyzed 17,247 AI responses and over 200,000 cited web sources collected for the restaurants and coffee shops category between 2026-03-10 and 2026-06-29, across four models, 12 countries (US, CA, GB, AU, ES, MX, AR, SE, NL, FR, DE, IT), and 7 languages (English, Spanish, Swedish, Dutch, French, German, Italian). Only sources actually cited in a response were counted. Cross-model overlap was measured on each model pair's top-20 most-cited domains (shared domains divided by the combined set), reported as an average across all six pairs and as a pooled proportion with a 95% confidence interval. Domain categories (Commercial, Editorial, UGC, Institutional, Reference, Other) were assigned from a global domain registry.
Localization is measured by a country-code-domain proxy, the share of cited domains ending in the market's local country code, so those figures are directional, form a floor on true localization (local publishers often use .com), and exclude English-language prompts. Temporal analysis split the observation period at its midpoint (2026-05-04) and compared domain rankings in each half. Model coverage varied across the observation window, so temporal movements are reported as relative ranks within each half. Cells with fewer than 100 observations are flagged as thin and treated as directional (Google AI Overview's non-English localization cells).
Frequently asked questions
Do different AI models really cite different sources for the same restaurant question?
Yes, dramatically so. Any two of the four models share only about 21% of their top-20 most-cited domains, meaning roughly 79% of the sources differ. There is no single set of "authoritative" sources for where-to-eat questions across AI assistants.
Which domains are cited most for restaurants and coffee shops?
TripAdvisor is the most-cited domain overall, alongside Instagram, the Michelin Guide, Reddit, and Yelp. But each model weights them differently, Grok leans on TripAdvisor and Michelin, Microsoft Copilot on Instagram and local directories, ChatGPT on Reddit and food media.
Which model cites the most sources per answer?
Grok, by a wide margin, about 27 cited sources per response, versus roughly 14 for Microsoft Copilot, 14 for Google AI Overview, and 6 for ChatGPT. Grok offers a brand many more citation "slots" to win.
Why does ChatGPT look different from the other models?
ChatGPT cited sources in only about 28% of its answers in this category, so it surfaces citations far less often than the search-grounded models. When it does cite, it relies on Reddit and editorial food and travel media rather than commercial directories, and it shows the highest reference-site share (5.3%) of any model.
Does local language matter for restaurant search?
It depends heavily on the market. In Sweden and Italy, roughly half of cited domains are local-country domains. In Spanish-speaking markets, the figure drops to about 19%–27%, global platforms and localized TripAdvisor editions dominate. This is a conservative ccTLD-based estimate that undercounts local publishers on .com.
Which model is best for reaching non-English markets?
Microsoft Copilot relies on local sources the most, it cites local-country domains 60%–72% of the time in German, Dutch, Swedish, and Italian, reflecting its heavy reliance on local restaurant directories. Grok relies on local sources least, defaulting to global .com platforms in most languages. Google AI Overview's non-English volume was too small in this window to compare reliably.
Should a restaurant brand prioritize a directory listing or editorial coverage?
Both, but for different models. A directory or booking-platform listing (TripAdvisor, RestaurantGuru, OpenTable) matters most for Microsoft Copilot, where nearly 70% of citations are commercial. Earned coverage in food and travel media, plus an active Reddit presence, matters most for Grok and ChatGPT, where editorial and user-generated content together make up over 75% of citations.
Did the source rankings change much over the sixteen weeks?
Yes, in relative terms. Reddit rose to #1, TripAdvisor slipped to #4, and sites like Yelp and TheFork dropped out of the top tier. These are relative rank movements within the observation window.
