How Does ChatGPT Decide Which Businesses to Recommend?

Short answer: For anything current or local, ChatGPT almost never answers from memory — it runs a live search, pulls a handful of pages, and pieces together an answer from what those pages say. Businesses that get mentioned tend to have three things in common: their basic facts (name, service area, hours, services) are readable by machines through schema markup, those same facts are repeated consistently across their website, Google Business Profile, and a few independent sources, and their site doesn't accidentally block the AI crawlers doing the reading. Nobody outside OpenAI, Anthropic, or Google knows the full ranking logic, and this article won't pretend otherwise — but that pattern holds up consistently across real sites.

A few weeks ago I wrote about asking ChatGPT to recommend my own business and getting nothing back. That post was about what happened. This one is about why it happened — what's actually going on when an AI model decides one business is worth mentioning and another isn't, and what you can actually do about it this month, not in some vague future.

I'm not going to pretend I have this fully figured out. Anyone who tells you they've "cracked the algorithm" is selling something. What follows is what I've seen work across real client sites, cross-checked against what these companies have said publicly about how their models retrieve information — and I've tried to be specific about where I'm confident versus where I'm guessing.

Two different things are happening, and people mix them up

When someone asks ChatGPT "who's a good plumber in Austin," one of two things happens:

1. The model answers from training data — patterns it learned during training, with no live lookup. This is static and slow to change; it reflects how visible your business was across the web as of the training cutoff, and it updates only when a new model version is trained.
2. The model does a live search — ChatGPT with browsing, Perplexity, Google's AI Overviews, and similar tools actually query the web in real time, pull a handful of pages, and summarize them.

Most business owners assume it's #1 — that ChatGPT "just knows" things, like a search engine with opinions. In practice, for anything even slightly current or local, it's almost always #2 now. ChatGPT's browsing mode, Perplexity by default, and Google's AI Overviews are all doing live retrieval for local and "best X near me" style questions — they're not relying on frozen training data for anything time-sensitive.

That distinction matters because it changes where your effort should go. If it were mostly #1, the only lever you'd have would be waiting for the next model to be trained on a more visible version of the internet — which is slow and mostly out of your hands. Because it's mostly #2, your actual website, today, is part of what gets read on every single query. That's a lever you control directly, and it's the whole point of this article.

How the live-search version actually works, step by step

It helps to walk through what happens between someone typing a question and getting an answer, because each step is a place where a business can either show up or get skipped:

1. The model rewrites the question into search queries. "Who's a good plumber in Austin" might become two or three actual searches — something close to what you'd type into Google yourself.
2. It retrieves a short list of pages — usually somewhere in the range of five to ten sources, not the full first page of search results.
3. It skims each page for extractable facts, not the full text. This is closer to how a person scans a page in three seconds than how a person actually reads it.
4. It cross-checks facts across sources. If your GBP says one service area and your website says another, that inconsistency doesn't help — at best it gets ignored, at worst it makes the model less confident about mentioning you at all.
5. It writes the answer, usually citing two or three sources, sometimes with a link, sometimes without.

Every one of those five steps is a place your site can either help or actively work against itself.

What live AI search is actually reading

When the model does a live lookup, it's not reading your website the way a human does. It's pulling a small number of sources, often 5–10, and trying to extract clean facts fast. A few things consistently make that easier:

Structured data (schema markup). LocalBusiness, Service, Review, and FAQPage schema give the model machine-readable facts instead of forcing it to infer things from paragraphs. This is the single most concrete lever a business actually controls, and it's the one most small business sites get wrong or skip entirely.

Third-party confirmation, not just your own claims. A model has no reason to trust "we're the best plumbers in Austin" written on your own homepage. It has more reason to trust the same claim if it also shows up on your Google Business Profile, in a review site, in a local news mention, or on an industry directory. This is basically E-E-A-T for machines — consistent facts repeated across independent sources.

Whether AI crawlers can even reach your site at all. This sounds basic, but it's the one that surprised me the most. We wrote about a real case where a client's robots.txt was silently blocking GPTBot and ClaudeBot — the content was fine, the SEO was fine, but the site was invisible to AI crawlers because of three lines in a config file nobody had looked at in years. If you haven't checked this, it's worth five minutes before anything else on this list.

Recency signals. Reviews, GBP posts, and blog activity from the last few months seem to carry more weight than a static "About Us" page from 2019. Live search tools appear to weight freshness the way Google does, maybe more.

What I don't think matters much (yet)

This is where I'll give an honest opinion instead of hedging, because vague "it depends" answers aren't useful to anyone.

• Backlink volume for its own sake. In classic SEO, more links usually helps. For AI visibility, I haven't seen evidence that raw link count matters the way it does for Google rankings. What seems to matter more is whether a small number of trusted, topically relevant sources mention you — a local news write-up beats twenty low-quality directory links.
• Paid ads. ChatGPT and Perplexity aren't pulling from your Google Ads campaigns. This is one of the clearest gaps between traditional marketing spend and AI visibility, and it's part of why smaller businesses can actually compete here — money alone doesn't buy a mention.
• Keyword density. Stuffing "best plumber in Austin" fifteen times into a page doesn't help a language model the way it might have helped 2012-era Google. These models are reading for meaning, not counting occurrences.

I could be wrong about some of this in six months — this space moves fast, and I'd rather say what I currently believe and be corrected than give you a mushy non-answer.

What not to do

Since we're being honest: there are a few things being sold right now as "AI SEO" that I don't think work, and I'm not going to pretend otherwise just because it would be an easier sale.

• "Guaranteed ChatGPT mentions" services. Nobody controls OpenAI's live retrieval or training data closely enough to guarantee this. Anyone promising a specific outcome here is guessing, same as guaranteed #1 Google rankings always were.
• Mass directory submissions bought in bulk. Cheap, generic directory links might help the "does independent info about you exist online" signal a little, but junk directories with your business pasted into a template don't build the kind of consistent, trustworthy footprint that seems to matter. We don't sell this, and we wouldn't recommend buying it from anyone else either.
• Rewriting your whole site around "AI SEO" and abandoning normal SEO. These aren't separate disciplines competing for budget. A site that's genuinely well-structured, fast, and trustworthy for humans and Google is most of the way there for AI visibility too. Schema and crawler access are additions, not a replacement.