How do I make ChatGPT recommend my restaurant?

You can't buy a place in ChatGPT's answer — OpenAI's own documentation says placement is not guaranteed, and there is no ad slot. But you can earn one, and the good news is that we know how. Our census study of every café, restaurant and bar in two markets — 4,776 venues, 2,208 AI answers — measured which factors get a venue recommended. Five of them are things a restaurant owner controls directly. This article is those five steps, in the order the evidence supports.

What decides whether AI recommends a restaurant?

Documentation — how much is written about you, and whether the systems can find and read it. AI recommendation works in two stages. First the assistant has to find you in the material it reads; here the measured factors were your own website (1.9× the odds), review volume (1.6× per step), a complete profile with price level (1.5×), and mentions across the web (1.4× per step). Only once you're inside the answer does the second stage begin, where star rating finally matters (1.17× per step toward being named first). A 4.9 venue with forty reviews loses to a 4.6 with nine hundred — not because it's worse, but because there is less of it to read.

This is why the question “how do I make ChatGPT recommend me?” has a real answer: not a trick, but a short list of inputs with measured weights. Here they are.

The five steps that measurably raise your odds

1. 01Have your own website — and make it answer what a guest asks. Where you are, when you're open, what you serve, what you're known for, in plain text. In our citation logs a venue's own site was the single most-cited source of all: the systems don't check that it exists, they read it and quote it. Make sure it's crawlable — OpenAI's search bot must be allowed in for a site to be eligible at all.
2. 02Complete every field of your Google Business Profile. Hours, menu, photos, attributes — and the price level, which on its own was associated with 1.5× the odds. Ten minutes, measured, yours.
3. 03Build a steady stream of detailed reviews. Volume was 1.6× per step at the entry door and 1.3× inside it, and detail gives the assistant sentences to answer questions with (“quiet at lunch, easy parking”). Ask every guest, at the natural pause, with the review one tap away — the full method is in our pillar guide.
4. 04Be present beyond one platform. TripAdvisor, Facebook, local guides, blogs, forums — mentions across sources were 1.4× per step, and those are exactly what the assistants cite alongside the big platforms.
5. 05Keep the record current. The systems recommended permanently closed venues 93 times in our study, because old listings outlive businesses. Fresh reviews very likely help at the entry door too (1.4× — just shy of our deliberately strict statistical bar, so we say “very likely,” not “proven”). Recent reviews and a live profile are how the machine knows you're open this month.

How do I check whether it's working?

Ask the way guests ask, and ask enough times to see a pattern. Write six occasion-based questions (“quiet work lunch,” “date night with good wine,” “brunch with kids”), put them to at least three systems, and repeat them across a week — a University of St. Gallen team measuring ChatGPT visibility found only 32–43% overlap between identical prompts on the same day and recommends at least seven runs per question. In our study any two systems agreed on only 33–54% of their top-20, so one screenshot tells you nothing; a month-over-month pattern tells you everything. Or let our free AI-visibility checkup run the protocol for you and send back where you appear, where you don't, and which of the five steps to take first.

What doesn't work?

Shortcuts. “GEO hacks” — rewriting pages with more statistics and authoritative phrasing — were replicated across domains and models by a Parameter Lab team, and only 3 of 54 combinations produced a significant gain; Google's guidance for its AI features says there are no special optimizations. Bought or rewarded reviews are restricted by every platform and, in the US, finable — the honest alternative is step three above. Offering every guest a private channel for problems alongside the public buttons is not gating; filtering who gets asked is. The five steps are slower than a hack and the only thing the data actually supports.

Frequently asked questions