How ChatGPT Decides Which Businesses to Recommend

What decides which businesses ChatGPT names?

No one outside OpenAI knows the exact formula, and OpenAI has not published one. What OpenAI does say is that ChatGPT ranks search results using several factors meant to help people find relevant, reliable information, and that placement is not guaranteed. (OpenAI Help Center, "Searching the web with ChatGPT", checked October 19, 2026)

So an honest answer has three parts: what is documented, what is suspected but unconfirmed, and what nobody outside OpenAI can tell you. Each section below is labeled, so you can see which kind of claim you are reading.

What we know: what OpenAI documents

These points come from OpenAI's Help Center article above, its crawler documentation ("Overview of OpenAI Crawlers") and its October 31, 2024 announcement, "Introducing ChatGPT search". All were checked on October 19, 2026.

• ChatGPT searches the web on its own. The Help Center says it may search the web by itself when a question would benefit from current information. The article does not say how often a question about a local business triggers a search.
• It sometimes uses other search providers. OpenAI says ChatGPT search sometimes works with other search providers and usually rewrites a question into one or more targeted queries that it sends to them. The 2024 announcement says it draws on third-party search providers as well as content supplied directly by partners. OpenAI does not say which provider answers which question.
• Your location and memory change the search. ChatGPT uses an approximate location from the IP address, and the Help Center's own example turns a request for good restaurants near the user into a search for top restaurants in San Francisco. If memory is on, saved memories may also shape the rewritten search. Two people can ask the same question and get different searches.
• It shows sources, with a warning. Answers that use search may include citations, and OpenAI states that search results and citations can be incomplete, outdated or incorrect.
• Eligibility is a crawler setting, not a ranking. To make a site eligible for inclusion, OpenAI says to allow its search crawler, OAI-SearchBot, and to confirm that your host or content delivery network allows OpenAI's published IP addresses. Being eligible does not mean being named.

OpenAI documents several crawlers. Three matter here, and they have three different jobs:

OpenAI adds that a robots.txt change can take about 24 hours to take effect for search. NIXAR's own robots.txt allows OAI-SearchBot, checked October 19, 2026.

Google makes a similar point about its own AI features. Its documentation (Google Search Central, "AI features and your website", last updated December 10, 2025) says a page must be indexed and eligible to appear in Search with a snippet, that there are no additional technical requirements, and that you do not need to create new machine readable files, AI text files or markup. It also says "Indexing and serving isn't guaranteed."

What we think: suspected, not confirmed

We found no OpenAI statement confirming any of the following as a rule. They are suspected patterns.

• Consistent facts seem to help. When a business's name, services, city and contact details match across its site, its profiles and its listings, an assistant has less to reconcile. We found no OpenAI or Google page that documents it as a ranking recipe.
• Some writing changes help in lab tests. Aggarwal and colleagues ("GEO: Generative Engine Optimization", first posted November 16, 2023, accepted to KDD 2024) built a benchmark of queries and reported that some content changes raised visibility in generative engine responses, with effects that vary by domain. That is a lab benchmark, not proof about local businesses in ChatGPT today.
• Special files and tricks are unproven. Google's documentation above says AI text files are not required. We know of no controlled evidence that an llms.txt file raises ChatGPT citations. It is harmless to keep, but do not count on it.

What we don't know

• The formula, and how much weight each factor carries.
• How much of an answer comes from a partner's search results and how much from OpenAI's own crawl. The Help Center says both exist and does not give a split.
• Whether a Business Profile, a review count or a review score changes who gets named. The Help Center says ChatGPT may share your location with providers of nearby business listings, and does not say which listings.
• How often a local question triggers a web search at all, as opposed to an answer from what the model already holds.
• How stable an answer is from week to week, and how different it looks to different people.

What a business can control

None of these guarantees a mention. They are the parts that are yours, and most of them also help on Google.

1. Let the search crawler in. Check robots.txt for OAI-SearchBot, and check that your host or CDN is not blocking OpenAI's published IP addresses. Decide on GPTBot separately, because the two settings are independent.
2. Keep one set of facts. Use the same business name, services, city and contact details on your site, your profiles and your listings. Fix old listings rather than leaving two versions.
3. Write the page that answers the question. A strong page is specific to one service and answers what that business's customers actually ask, rather than generic copy about the industry.
4. Earn real reviews and mentions where your customers look. Ask every customer, at checkout or the closing moment, then follow up by text. Do not screen who gets asked, and do not offer anything for a review.
5. Verify your Business Profile if you qualify. Google's guidelines allow a service-area business one profile with a designated service area, and say to hide the address from customers. A mailing address that is not a working location is not eligible (Google Business Profile Help, "Guidelines for representing your business on Google", checked October 19, 2026).

For the full playbook on making a business easy for AI tools to cite, read our GEO guide.

How NIXAR tests it

NIXAR does not claim to know the formula, so we test what the assistants say instead. The method has five steps:

1. A fixed list of questions, written the way a customer types them.
2. Several assistants, each given the same questions in a fresh chat.
3. The same questions repeated on a schedule, weekly in Claude and monthly across ChatGPT, Gemini, Perplexity and Copilot, so a change in the answers is not a change in the wording.
4. Every answer logged to a table: whether the business is named, where it sits in the list, which other businesses appear, with the raw answer saved so cited sources can be read back.
5. A person spot-checks the automatic extraction before we quote any row, and results are compared only with an earlier run of the same method.

The limits are real. The sample is small. Answers vary from run to run and by account, location and memory. A script that runs an assistant without a person is not a real user. And every result is a snapshot of one day.

What the panel does and does not tell us: it shows whether a business is named and what an assistant says about it, on the day we ask. It does not tell us what works, and we do not treat it as proof of any tactic. An account's memory can carry into answers, so each question is asked in a fresh chat, signed out or in a temporary chat.

Your next step

Start with the parts you control: the crawler setting, one consistent set of facts and a page that answers your customers' questions. If you want help with them, the SEO and GEO service explains how NIXAR works on Google, Maps and AI answers and how we check what the assistants say. We cannot promise a mention, and no one honestly can.