How does ChatGPT recommend businesses?

ChatGPT recommends businesses through two separate pipelines. The first is model memory: what it learned about your market during training, from a snapshot of the public web that is months old by the time anyone asks. The second is live retrieval: when a question needs current or local facts, ChatGPT Search runs and composes its answer from the reviews, directories, comparison pages, and community threads it can reach — pages its own crawler, OAI-SearchBot, was allowed to read. There is no business database inside it, no submission form, and no ad slot in the organic answer. You get named when the sources a pipeline reads agree you belong there.

One question in, two pipelines behind the answer

When someone asks ChatGPT “who's the best injury attorney in Denver?” or “which CRM should a 12-person team buy?”, it does not look your market up in a directory. It writes the answer on the spot, and every business name in that answer arrived through one of two routes:

• The training pipeline (model memory).During training, the model absorbed an enormous snapshot of the public web. Businesses with years of consistent coverage — press, reviews, directories, reference pages — got encoded into that memory. Everyone else mostly didn't.
• The retrieval pipeline (live search). When the question calls for current, local, or specific information — and most buying questions do — ChatGPT searches the web mid-answer and builds its recommendation from the pages that come back.

The two pipelines age differently, read different sources, and respond to different work. Treating “ChatGPT visibility” as one undifferentiated thing is where most owners' effort goes wrong — and the audience reading these answers is no longer niche. Omnibound's 2026 statistics roundup puts ChatGPT at roughly 900 million weekly users as of February 2026.

Pipeline one: what the model memorized

Training doesn't store records; it compresses patterns. The model ends up with a statistical picture of your business assembled from everything the public web said about you before its cutoff date — not a fact sheet it can look up, but an impression it reconstructs on demand. Three consequences matter for owners:

• Thinly documented businesses are simply absent. A company that opened last year, or one whose footprint is a website and a handful of listings, usually left too faint a trace to be reconstructed. Memory-only answers skip it entirely.
• Descriptions freeze at the cutoff. Renamed services, moved offices, and discontinued products persist in memory long after you fixed them everywhere that matters. If ChatGPT keeps repeating your 2024 story, this is why — the correction paths are their own subject.
• Similar names blur together. A statistical impression of two similarly named companies can merge into one confident, wrong description. The thinner your documentation, the more the better-documented namesake wins the blend.

You cannot edit any of this directly — there is no correction form for a model's weights. The only lever is the public record the next training run will read: the same name, category, and story stated consistently across your site, your profiles, and the directories that carry you. Slow, compounding work — which is exactly why the second pipeline matters more in the short term.

Pipeline two: live search, and what is actually known about it

ChatGPT decides per question whether to search. Questions with buying intent — best, near me, how much, who should I hire — usually cross the threshold, because a good answer needs facts fresher than any training cutoff. What happens after that threshold is where most writing on this subject gets sloppy, so it is worth splitting into three parts and keeping them apart: what OpenAI documents, what outside testing measures, and what follows for you.

Part one: what OpenAI documents

OpenAI publishes the crawlers it operates, and that distinction is the most consequential technical detail on this page. OAI-SearchBot is the crawler behind ChatGPT Search — the surface that produces live, sourced answers. GPTBot is primarily training-data collection. They are separate agents doing separate jobs, which means blocking GPTBot is a defensible choice about training data and does notremove you from ChatGPT Search — while blocking OAI-SearchBot does exactly that. Anthropic splits its crawlers the same three ways: Claude-SearchBot for search appearance, ClaudeBot for training, Claude-User for pages a person asked Claude to fetch. Any “block the AI bots” recipe that names GPTBot and omits OAI-SearchBot is answering a different question than the one a business owner is asking.

What OpenAI does not publish is which search index — or indexes — sit behind ChatGPT Search. The documentation describes the crawler, not the retrieval provider, and there is nothing in it naming a single one.

Part two: what third parties measure

Because the retrieval side is undocumented, practitioners measure it from the outside: take the pages ChatGPT cites, and check what conventional search engines rank for the same query.

Read the number for precisely what it says: most of ChatGPT's cited pages also rank in Bing. It does not say Bing is the index, or the only index, or a requirement. An overlap that size is also what you would expect from two systems ranking the same open web on broadly similar signals. Anyone telling you Bing “feeds” ChatGPT is stating a cause where only a co-occurrence has been measured — and vendors who sell Bing work have an obvious reason to prefer the stronger sentence.

Part three: what follows practically

The cautious reading still leaves you with real work, which is why the uncertainty is worth stating rather than hiding. A page no general search engine has ever indexed sits outside the set that overlap is measured over — so the check is cheap and it eliminates a whole category of silence. One search on Bing does it: site:yourdomain.com. If your key pages are missing, IndexNow is the fastest route in — an open protocol that tells Bing a URL is new or changed within minutes of publishing. Submission is instant; indexing remains Bing's decision, and nobody can honestly say otherwise. Bing also carries Microsoft Copilot's answer layer — Microsoft's own product on Microsoft's own index — so this work has a second payoff that requires no inference at all. The full mechanics live in the Bing and Copilot guide, and the engine-specific tactics in the ChatGPT playbook.

What the live pipeline actually reads

A retrieved answer is typically composed from a handful of pages. For business recommendations, four source families dominate what gets quoted:

• Review platforms.Pages that aggregate ratings and written experiences are dense, structured, and directly responsive to “who's good?” — the exact shape a composed answer needs.
• Directories and data aggregators.Industry directories and local listings corroborate the basics: that you exist, what category you're in, where you serve. For local questions this stack has its own rules — how AI assembles “near me” answers covers it.
• Comparison and roundup pages.“Best X in Y” articles pre-chew the recommendation work, so engines quote them constantly — even when a competitor or a publisher you've never heard of wrote them.
• Community threads. Reddit above all: engines treat threads of real people comparing providers as high-trust recommendation evidence. Why AI over-weights Reddit is its own story.

Notice what's far down the list: your own website.

Where your site does earn its keep is extraction. Structured data (schema) states your name, exact category, service area, and offerings in a format that leaves a machine nothing to interpret — including the specific business subtype rather than a generic label. Be clear-eyed about the size of that claim: Google states plainly that no special markup is required to appear in AI Overviews or AI Mode, and there is no such thing as “AEO schema.” What markup does is remove ambiguity about your prices, hours, and services — and it still earns classic rich results, which is reason enough on its own. Pages that answer one buyer question directly, in the first paragraph, do the same job in prose. Neither substitutes for third-party corroboration.

The ladder: mention, citation, recommendation

“Showing up in ChatGPT” is not one state. Practitioner writing on brand visibility describes a ladder, and it's the most useful frame we've found for reading your own standing:

1. Mentioned. Your name appears in the answer, usually among several. At least one pipeline knows you exist.
2. Cited. The engine quotes or links a source about you. You are now feeding the retrieval pipeline with material it considers worth showing — a much stronger position, because citations are repeatable in a way memory flickers are not.
3. Recommended. You are the named pick for a buying question — and you stay the pick when the question is rephrased and re-asked. This only happens when several independent sources agree on you.

The re-asking part is not a detail. According to the 2025 RankOS AI Visibility Benchmark, only about 30% of brands that appear in an AI answer show up again the next time the same question is asked. The sources behind the answers churn too:

Three myths that waste owners' time

Myth 1: “You can submit your business, or pay for placement.”

There is no portal that registers a business with ChatGPT the way Google Business Profile registers one with Maps, and as of mid-2026 no ad product sells a spot inside the organic recommendation. Both follow from the mechanism: there is no database to insert yourself into. The things that function like submission are getting your pages crawlable and indexed where general search engines can find them — Bing included, for the overlap reason above — and getting your name into the third-party sources the retrieval pipeline quotes.

Myth 2: “Build a Custom GPT and you're in.”

Advice circulating on several ranking pages tells owners to create a Custom GPT as their path in. A Custom GPT is a configured private chat that a user must deliberately open. It has zero effect on the default answers everyone else gets when they ask ChatGPT who to hire. It may be worth building for other reasons; as a recommendation tactic, it is a dead end.

Myth 3: “Rank first on Google and ChatGPT follows.”

Different retrieval step (Google's ranking is not what the live pipeline surfaces), different source diet (nine in ten citations third-party), different output (one composed answer, not ten links). Google rank is evidence you built things engines can value — it is not the input this one reads.

What this means you should do first

The mechanism dictates the order of work:

1. Establish which pipeline knows you. Ask ChatGPT about your brand without triggering a search, then watch what a live-searched buying question returns. The five-prompt check structures this in ten minutes.
2. If you're absent, find your specific reason. Not indexed by Bing, blocked crawlers, inconsistent entity data, and zero third-party corroboration produce identical silence with different fixes. The diagnostic walkthrough separates them.
3. Work the weeks-fast lane first. Bing indexing, schema with the correct subtype, and answer-shaped pages are retrieval-side moves that can register in weeks.
4. Then start the slow lane. Reviews, directory consistency, community presence, and earned mentions build the third-party agreement that recommendation requires — the playbook ordered by time-to-impact lays it out. Practitioner guides (Cited.so among them) put consistent recommendation at a 6–12 month build in most categories; the churn numbers above are why measuring continuously beats checking once.

Why we mapped this machine (a disclosure)

This mechanism is the foundation AEO Action's product is built on — which is both why we can describe it in this detail and why you should weigh the framing: the free audit asks ChatGPT, Gemini, and Perplexity ten of your buyers' questions — 30 live answers — and reports your standing as three numbers that map to the ladder above: mentioned, cited, named first, with the three most revealing answers shown word-for-word. Paid plans re-run the scan daily against the live engines, submit approved landing pages to Bing via IndexNow, and log a dated receipt when a page you shipped first gets cited — what that service covers, in full is its own page, and how it compares with Profound is the honest version of the shopping question. Sales pitch aside, the mechanics on this page are true whoever runs them.