Where assistants send people in San Francisco
66 answers from ChatGPT, Claude and Gemini, asked on 2026-11-02, to the questions people ask and to questions about the top businesses in the index. 94% of them link a business's own site or its booking software, and 30% link a marketplace, most often OpenTable (9%).
of answers link the business's own site or its booking software
link a marketplace
link OpenTable, the marketplace named most
answers, 66 to the questions people ask and 0 about businesses
Of the 66 answers to the questions people ask, 94% link a business's own site or its booking software and 30% a marketplace, OpenTable most often.
Updated 2026-11-03. The San Francisco index · Places named, question by question · Markdown · JSON · Share image · All cities.
The destinations, most linked first
Each row is a place an answer can send people. Share is the percentage of all 66 answers that link to it. Each assistant's column counts only that assistant's answers.
| Destination | Kind | Answers | Share | ChatGPT | Claude | Gemini |
|---|---|---|---|---|---|---|
| The businesses' own sites | own site | 61 | 92% | 22 100% | 17 77% | 22 100% |
| Press, guides and forums | press and guides | 11 | 17% | 0 | 0 | 11 50% |
| OpenTable | marketplace | 6 | 9% | 0 | 6 27% | 0 |
| karpo.ai | other site AI technology platform that does not fit business, directory, or booking categories. | 5 | 8% | 0 | 0 | 5 23% |
| waitlistwizard.com | vendor, direct | 5 | 8% | 0 | 0 | 5 23% |
| TableAgent | marketplace | 4 | 6% | 0 | 0 | 4 18% |
| Tock | vendor, direct | 4 | 6% | 0 | 0 | 4 18% |
| DoorDash | marketplace | 3 | 5% | 0 | 2 9% | 1 5% |
| Resy | marketplace | 3 | 5% | 0 | 0 | 3 14% |
| Uber Eats | marketplace | 3 | 5% | 0 | 0 | 3 14% |
| 7x7.com | other site | 2 | 3% | 0 | 0 | 2 9% |
| DoorDash Storefront | vendor, direct | 2 | 3% | 0 | 1 5% | 1 5% |
| Michelin Guide | directory | 2 | 3% | 0 | 0 | 2 9% |
| perfectvenue.com | vendor, direct | 2 | 3% | 0 | 0 | 2 9% |
| restaurent.com | marketplace An online platform for discovering and booking event spaces and private dining rooms. | 2 | 3% | 0 | 0 | 2 9% |
53 more destinations with fewer answers each are in the JSON.
What each assistant does in San Francisco
ChatGPT: answered by gpt-6-luna (22).
Claude: answered by claude-sonnet-5-5 (22).
Gemini: answered by gemini-3.8-flash (22).
Places named when people ask
Each assistant was asked the way people ask, with web search on, for up to three places and a link to book or order with each one directly. The table shows how many of the places it named came with a link to the place's own site, to a marketplace or to a directory, or with no link at all.
| Assistant | Own site or direct | Marketplace | Directory | No link | Places |
|---|---|---|---|---|---|
| ChatGPT | 97% | 0% | 0% | 3% | 63 |
| Claude | 68% | 8% | 0% | 24% | 63 |
| Gemini | 97% | 0% | 0% | 3% | 68 |
Who gets named, and with which link
Asked on 2026-11-02. A place in this index links to its report. A place with its own site that isn't in the index yet links to a scan. "No link" means the assistant named the place but didn't link its site. Each question links to the full list of places named.
How this is counted
Everything here is counted in answers, using the latest answer from each assistant to each question, asked with web search on. Some questions are the ones people ask ("best pizza delivered right now", "book me a table for Saturday"). The others are three questions about each of the top-ranked businesses in the index: the official website, where to book and where to order. Every link in an answer is read. A destination counts once for each answer that links to it, whether as the place to book or as a source the assistant cited. So each share is the share of answers that link that destination, and the shares don't add up to 100%.
All the businesses' own sites share one row, whatever the business. A booking or ordering vendor on the business's own domain counts as direct. Marketplaces and directories are named. Listing, review and map sites (Tripadvisor, Google Maps, Yelp) count as directories, even where the scanner files them with the marketplaces. Newspapers, magazines, guides, forums and tourism sites share one row. Anything left is listed by its registered domain, and a model sorts each of those unknown hosts once into directory, press, marketplace, vendor, business or other, with the reason shown. Text that only looks like a host because a space is missing ("them.here") is dropped. Business names are matched against the index. A place that isn't in the index links to a scan of its site, if the assistant gave one.
Updated 2026-11-03. The data is open, and the JSON version of this page has every destination. Methodology.