ChatGPT named none of 50 independent coffee roasters. Here is the study.

I asked ChatGPT and Perplexity five ordinary questions about buying coffee online, then checked whether 50 randomly chosen independent roasters showed up in the answers.

In a study I ran on November 6, 2026, ChatGPT named none of 50 randomly chosen independent US coffee roasters when asked five common questions about buying coffee online. Perplexity named one. These stores were not hiding from AI. Of the 50, 36 publish product price and availability in structured data, and only one blocks an AI crawler.

I write a lot about AI agents picking a short list before anyone buys. I wanted a number for how short that list is for an ordinary small merchant, so I picked one category and measured it. Coffee suits the test because there are so many independent sellers: Google alone surfaced 248 roasters selling their own beans online.

How the study worked

I wrote the method down before collecting any data. Ten fixed coffee shopping searches on Google in the US returned 418 sites. I removed marketplaces, grocers, publishers, directories, equipment sellers, multi roaster services, large chains and sites that would not load, which left 248 independent roasters. A computer drew 50 of them at random with a fixed seed, so the draw can be repeated exactly.

Then I asked each engine these five questions, once each:

1. What are the best online stores to buy coffee beans?
2. Where should I buy fresh roasted specialty coffee online?
3. What are the best coffee subscriptions in the US?
4. Which independent coffee roasters ship nationwide and are worth ordering from?
5. I want to buy a bag of good whole bean coffee online for home brewing. Which roasters do you recommend?

ChatGPT ran through OpenAI's API with web search turned on, and Perplexity ran through its own API. A store counted as named if an answer mentioned its brand or its domain, or cited one of its pages. I also checked each site for an llms.txt file, for Product schema with price and availability, and for robots.txt rules that block AI crawlers.

What the engines named instead

ChatGPT recommended marketplaces and subscriptions such as Trade, MistoBox, Bean Box, Dayglow and Atlas, plus a short list of nationally known roasters including Onyx, Counter Culture, Heart, SEY, George Howell, Verve and Olympia. Perplexity named a similar group. The one sampled store that made it was Roast Coffee Company of Medford, New Jersey, which Perplexity named and linked in its answer to the second question.

The sources ChatGPT cited show where it looked. Across the five answers, the site it cited most was Roast Magazine, a coffee trade publication. Kiplinger and Bon Appétit appeared too, along with the marketplaces and the well known roasters. For its fifth answer ChatGPT asked clarifying questions and named no one.

What the technical checks showed

• 36 of 50 stores publish Product schema with price and availability.
• 41 of 50 serve a readable llms.txt file, and 4 serve it as plain text, which was the stricter rule I set before starting.
• 1 of 50 blocks an AI crawler in robots.txt (it blocks GPTBot only).
• 0 of 50 were named by ChatGPT, and 1 of 50 by Perplexity.

My read is that the decision happens before an agent ever reaches a product page. The engines named stores that other sources already talk about: the marketplaces that carry them, the magazines and guides that review them, and roasters with a national reputation. A clean product feed helps once an agent is already looking at a store, but in this sample it did not get anyone onto the list. That is the Found and Chosen part of Found, Chosen, Booked, and it is where small merchants are losing right now.

If you sell online, the useful next step is to get covered by the places these engines already cite in your category, such as trade publications, buying guides and marketplaces, and then keep checking whether the answers change. I explain how the shopping agents choose in agentic shopping and what a store can build in AI agents for ecommerce. For the wider category, start with what is agentic commerce.

Limits of this study

• Each question was asked once per engine, on one day. AI answers vary between runs, so a repeat could name different stores.
• One category, US results only. These numbers describe these 50 stores on this date, not all small retailers.
• Five questions is a small set. Different wording, a city in the prompt or a follow up question could change who gets named.
• The schema check looked at the homepage and the first product page linked from it. For 11 stores no product link was found, so they count as no.
• Crawler blocking was measured from robots.txt only, not from firewall or server rules.

The written method, the sample, every raw answer and the scripts are on file. If you want to check or repeat the work, ask me for them.

FAQ

Does ChatGPT recommend small online stores?

In this study it rarely did. Asked five common questions about buying coffee online on November 6, 2026, ChatGPT named none of 50 randomly chosen independent US coffee roasters and Perplexity named one. Both engines favored marketplaces and roasters with a national reputation.

Does blocking AI crawlers explain why stores are not recommended?

Not in this sample. Only 1 of the 50 stores blocked an AI crawler in robots.txt, so blocking cannot explain why 49 of them went unnamed.

Does product schema get a store into AI answers?

Not by itself. Of the 50 stores, 36 published price and availability in Product schema, and ChatGPT named none of them. Structured data helps once an agent is already considering a store, but in this sample it did not put anyone on the short list.

Which sources did ChatGPT cite for coffee recommendations?

Across the five answers, the site it cited most was Roast Magazine, a coffee trade publication. It also cited marketplaces and subscriptions such as Trade, MistoBox and Dayglow, roasters with national reputations such as George Howell, Onyx and Heart, and outlets such as Kiplinger and Bon Appétit.

How was the study done?

Ten fixed Google US coffee searches produced 248 eligible independent roasters, and 50 were drawn at random with a fixed seed. Each engine was asked five buyer questions once. A store counted as named if the answer mentioned its brand or domain or cited one of its pages.