The shape of shopping inside ChatGPT — a 12-million-prompt study.
For three months we collected 12,184,000 anonymised shopping conversations across ChatGPT, Claude, Gemini and Perplexity. We wanted to answer one question: when a buyer asks an AI assistant for help, which brands actually surface — and why?
The dataset
Sessions came from a panel of 38,000 consenting users across the US, UK, Germany and the UAE. We filtered to messages with clear commercial intent ("best", "vs", "recommend", "under $X"), then clustered them into 412 product categories using a combination of embedding similarity and human review.
Every AI response was parsed for brand mentions, sentiment, citation links and ordinal position. We were strict: if a brand was only mentioned as a counter-example ("unlike Brand X"), we tagged it as a negative reference.
What we found
1. The top-3 capture 71% of attention.
Across categories, the first three brands named by the assistant account for the overwhelming majority of follow-up questions. "Tell me more about #2" is roughly 9× more common than "tell me more about #5". The list is shorter than a SERP, and the cliff is steeper.
2. Citations are the new backlinks.
Brands that earned 3+ citations from independent review sites in the last 90 days were 4.2× more likely to be named in the top-3. Press releases barely moved the needle. Editorial coverage, Reddit threads and YouTube reviews dominated.
3. Attributes matter more than adjectives.
Assistants strongly prefer brands whose product pages list structured attributes (materials, dimensions, certifications, warranty terms). Pages of marketing prose lost to pages with a simple spec table — even when the prose page had more inbound links.
The five sources that mattered most
- Wikipedia (entity confirmation)
- Reddit (real-user sentiment)
- YouTube transcripts (in-depth reviews)
- Independent review media (The Verge, Wirecutter, NerdWallet, etc.)
- Manufacturer spec pages with structured data
What to do on Monday morning
Audit your top 20 commercial pages for structured attributes. Pitch one independent reviewer per category every two weeks. And track your AI mention share monthly — the cost of being absent is now measurable.
“The brands winning inside ChatGPT aren't louder. They're easier to quote.”
