Nobody outside OpenAI knows how ChatGPT chooses which businesses to recommend, and OpenAI publishes exactly two sentences on it. Those two sentences say that ChatGPT ranks search results using multiple factors intended to help people find relevant, reliable information, and that placement is not guaranteed. No published factor list. No weightings. No appeal. Everything else written on this question, including the confident numbered lists, is inference. Some of it is reasonable. Some of it is invented.
This piece separates the three layers: what OpenAI documents, what is visibly happening between your buyer's question and the answer, and what people are guessing. If what you actually want is to see what ChatGPT currently says about your own business, that is a different job and it takes ten minutes: can ChatGPT find your business.
- OpenAI publishes two sentences on ranking: multiple factors intended to surface relevant, reliable information, and placement is not guaranteed. There is no published factor list and no weightings.
- The only actionable instruction OpenAI publishes is eligibility: allow OAI-SearchBot to crawl, and make sure your host or CDN is not blocking its IP addresses. The second half is the one that catches people out.
- ChatGPT rewrites the question before it searches, so you are competing to answer a query you never see rather than the one your customer typed. That is why results feel arbitrary.
- Of the factor lists circulating, consistency and third-party mentions are reasonable inference, authority is too vague to act on, and engagement metrics are invented outright.
- The same system can describe you confidently and wrongly, with no correction form and no appeal. Make the facts easy to retrieve and hard to contradict, and check twice a year.
What OpenAI actually publishes
The whole of OpenAI's public statement on ranking sits in its help documentation, under a heading about making your website available in search results. It reads, in full:
ChatGPT ranks search results using multiple factors intended to help users find relevant, reliable information. Placement is not guaranteed.
OpenAI, Searching the web with ChatGPTThat is it. That is the entire published ranking statement for a system a growing number of buyers now ask for recommendations before they ask anyone else.
The same page carries one concrete eligibility requirement, which is the only actionable instruction OpenAI publishes on the subject: allow OAI-SearchBot to crawl the site, and confirm that your host or content delivery network allows traffic from OpenAI's published searchbot IP addresses. The second half of that sentence catches people out far more often than the first. A site can have a perfectly permissive robots.txt and still be invisible, because a firewall or CDN rule is turning the crawler away at the door before robots.txt is ever read.
OpenAI names the crawlers separately in its bots documentation: OAI-SearchBot for search, ChatGPT-User for pages fetched because a user's question prompted it, and GPTBot for content that may be used in training. Only the first two decide whether you can be recommended today.
What happens between the question and the answer
The mechanism most owners picture is wrong in a way that matters. They imagine ChatGPT holding a ranked list of businesses in a category and reading off the top three, the way a search results page works. That is not what happens.
When ChatGPT searches, it rewrites the question first. OpenAI's documentation describes this plainly and gives its own worked example: a researcher's question about a class of cancer drugs is rewritten into a narrower query, results come back, and then further, more specific queries follow based on what returned. Your buyer's question goes through the same machinery. Somebody asks for a good accountant for a small import business, and ChatGPT sends out a rewritten query that neither of you ever sees.
That has a consequence worth sitting with. You are not competing to be the best answer to the question your customer asked. You are competing to be a good answer to a query generated from it, and you cannot read that query. It is the single biggest reason these results feel arbitrary from the outside.
Three further details are documented rather than guessed, and each one changes how you should read a result:
- It uses outside search providers. In the section covering information shared with search providers, OpenAI points readers to the privacy policies of Microsoft and Shopify, and for restaurants it names OpenTable, Resy and Yelp. So part of what ChatGPT knows about businesses arrives through other companies' indexes and listings rather than through anything you ever handed to OpenAI.
- Location is approximate by default. ChatGPT estimates general location from the IP address, at country, state or city level, and device location sharing is optional and switched off by default. A buyer two towns over may well be getting results drawn from a much wider area than you are picturing.
- Memory rewrites the query. OpenAI states that when memory is enabled, saved memories may be used when rewriting a search query, and gives the example of someone who has mentioned being vegan getting a restaurant search rewritten accordingly. This is precisely why testing from your own logged-in account tells you almost nothing, and why the ten-minute check is run in a temporary chat.
Where the factor lists you read elsewhere come from
Search this question and you will find a dozen pages listing the factors ChatGPT uses to pick businesses. Relevance. Authority. Reviews and reputation. Consistency across the web. Engagement metrics. They are written as findings, in the tone of something discovered. None of them carries a source, because there is nothing to source them to.
It is worth separating them, because the categories behave very differently once you try to act on them.
The honest summary is that the reasonable inferences are probably worth acting on and the rest is noise. The trouble with the pages presenting all five as equal findings is not that every item is wrong. It is that they give you no way to tell which is which, and they are almost always written by somebody selling the fix.
What the system gets wrong, and how confidently
There is a second half to this question that almost nobody selling AI visibility will raise, which is how reliable the recommendation is once you get one.
Researchers at the Tow Center for Digital Journalism tested it directly. They gave ChatGPT 200 quotations drawn from articles across 20 publishers and asked it to identify the source of each one. It returned partially or entirely incorrect answers on 153 of those occasions, and admitted it could not answer accurately only 7 times. The ratio is the finding, more than either number on its own. The system very rarely says it does not know. It produces a confident answer either way.
That study was published in November 2024 and the models have moved a long way since, so treat those exact figures as a snapshot rather than a current measurement. The behaviour it identified has not gone anywhere, and OpenAI says as much in the same help documentation: search results and citations can be incomplete, outdated, or incorrect.
For a business owner that cuts in both directions. The system that might recommend you can also describe you confidently and wrongly to a buyer who has no reason to check and no way to know. There is no correction form and no appeal. The only lever anyone actually has is what the thing can find, which is the entire reason the input side is worth any attention at all.
The quoted ranking statement and the eligibility requirement are taken verbatim from OpenAI's help documentation, Searching the web with ChatGPT, read on 24 August 2026. The crawler names and their stated purposes come from OpenAI's published bots documentation. The query rewriting, search provider, location and memory details are all drawn from that same help page rather than inferred. The figures of 200 quotations, 153 incorrect responses and 7 admissions come from the Tow Center for Digital Journalism study by Klaudia Jazwinska and Aisvarya Chandrasekar, published 27 November 2024, and are flagged in the text as a snapshot because the models have changed since. The status column in the claims table is our own editorial judgement, not a measurement. The $250 a month AI search visibility line and the $2,400 a month Grow retainer are our published prices, listed in full on the pricing page.
What this means if you run a small business
You cannot optimise against two published sentences, and anybody offering to is selling a certainty they do not have, whichever acronym it comes under: GEO, AEO and AIO are one job. What you can do is narrower, more boring, and older than any of this.
Make the facts about your business easy to retrieve and hard to contradict. In practice that means the plain statements exist somewhere quotable, and the independent sources carrying your name agree with them. That is not a ranking trick and it is not new. It is the condition under which a retrieval system can say anything about you at all, and it happens to be the same condition under which a human researching you reaches the right conclusion.
It also means publishing things a lot of businesses would rather not. Prices are the clearest case, because a page carrying a real number can be quoted and a page saying contact us for a quote cannot. Ours are itemised in full on the pricing page, and why so few agencies will name a figure is its own piece.
Two honest limits to finish on. Being named in an answer is not a sale, and the distance between those two things is your own website's problem: how people actually decide once they arrive did not change because the referrer did. And for a good number of businesses this is a quarterly check rather than a monthly line item. The ongoing version costs $250 a month inside our Grow retainer at $2,400 a month, published in full like everything else on the pricing page, and we would genuinely rather tell you to run the free check twice a year than sell you a retainer you do not need yet.