How ChatGPT Decides Which Businesses to Recommend
Training versus live search, the providers behind retrieval, OAI-SearchBot access, and what OpenAI has never published about ranking.
Two ways ChatGPT can know your business
A ChatGPT answer can come from what the model learned during training, from a web search performed during your conversation, or from a mixture of both. Learned knowledge is not a live directory: it can be incomplete and it goes stale. Search is different. ChatGPT may decide that current information would improve its answer, or the person asking can select Search themselves. A question like "which tour guide should I hire in Las Vegas" is far more likely to trigger retrieval than "what is a tour guide". Treat any answer with no visible citations as unverified. To test properly, run the same question once normally and once with Search selected, then record whether the businesses named, the reasons given and the sources cited actually change.
The retrieval path is not one fixed index
OpenAI describes ChatGPT Search as using third-party search providers alongside content supplied directly by partners. Its Enterprise and Edu documentation states specifically that disassociated search queries may be sent to Bing. That supports saying Bing can be part of the path. It does not support saying every ChatGPT answer is a Bing result, and it does not support saying Google is never involved. OpenAI also says it rewrites your prompt into one or more targeted queries, and may issue further, more specific queries after reviewing what came back. The practical consequence is that your pages need to match the language customers actually use: service, place, audience and constraints. Two prompts that sound similar to you can retrieve a different set of candidates.
You have to be eligible before you can be chosen
OpenAI identifies OAI-SearchBot as the crawler used to surface websites in ChatGPT Search. It states that a site opting out will not be shown in search answers, though it may still appear as a navigational link. That makes access a hard prerequisite rather than a tactic. Check robots.txt, CDN rules, firewall settings, bot challenges and published OpenAI IP ranges before you rewrite a single sentence of copy. Keep the facts that matter in crawlable HTML rather than only inside scripts, logins, image files or forms. Do not confuse GPTBot with OAI-SearchBot: OpenAI documents these as separate controls, one for model-training use and one for search. Allowing the crawler makes you eligible. It does not guarantee placement, and OpenAI says so plainly.
Write pages that answer the retrieval question
A retrieved page is more useful when it states plainly what the business does, where it operates, who it serves, and what evidence supports the claim. Put the official business name, service area, phone, opening hours, service names and starting prices into readable text. Use headings that resemble real questions, but write for customers rather than inserting artificial prompt text. Link each service to its own page and keep outdated offers, locations and staff claims corrected. This is operational advice, not a published OpenAI scoring rule: OpenAI only documents broad aims such as relevance and reliability. The sensible goal is to make a retrieved page clear and checkable, so a model can use it without having to guess or extrapolate.
Build a consistent public record
When ChatGPT searches for a local business it may encounter your website, directory pages, map data, reviews, news and niche sources. OpenAI has not stated that any particular directory, review count, link metric or citation volume receives a fixed weight. What it has said is that placement uses multiple factors toward relevance and reliability. Inconsistent names, addresses, phone numbers, categories or descriptions make verification harder for people and systems alike. Claim the listings that matter in your market, remove duplicates, align core facts, and correct major third-party errors. Ask for genuine feedback through normal, policy-compliant processes rather than manufacturing reviews. The deliverable worth keeping is an audit trail: what you changed, the URL, the date, and any conflicts you could not resolve.
Local answers depend on context
ChatGPT can use approximate location inferred from an IP address, optional precise device location, saved memory, and locality terms in the prompt itself. A person asking "near me" may therefore get a different set of businesses from someone in another city, and a VPN can distort the result further. Test local visibility with explicit locations rather than treating one generic prompt as universal. Record the exact wording, the region or device, the date, whether Search was forced, and whether a map or reservation feature appeared. Make every service-area claim on your own pages precise, and where you do not serve a location, say so. Location is documented as a relevance input. It is not evidence that ChatGPT runs a published local-pack algorithm or a fixed radius.
What citations do and do not tell you
When ChatGPT uses web search, it may attach clickable inline citations and a Sources panel containing cited sources and other relevant links. A citation is useful because it lets a customer inspect the page behind a statement. It is not an endorsement, and it does not prove every sentence in the answer is supported. OpenAI warns that results and citations can be incomplete, outdated or incorrect. Check whether the cited page actually names the business, supports the service and location, and is current. Track which sources appear across repeated tests, but do not chase citations as a vanity metric. A good monitoring sheet records the prompt, date, answer, businesses named, citation URLs, any factual errors, and whether the source was the business itself or an independent publisher.
Freshness varies by mode and by page
Ordinary ChatGPT Search is intended to provide current web information, but current does not mean every page is refreshed immediately. OpenAI also documents an offline-search configuration for eligible workspaces that uses its own indexed and cached content instead of live external search. Coverage and freshness vary by site, language, region, page type, crawl access and technical setup, and OpenAI gives no refresh guarantee for a specific URL. This matters when a business changes its address, hours, pricing or availability. Update the source pages and listings, then re-test after a reasonable interval rather than promising instant propagation. Before comparing two results, establish which mode and workspace produced each one, because an offline-index result can look like a platform preference when it is really a configuration difference.
Product feeds are a separate pathway
OpenAI has published specific guidance for shopping and product discovery. Merchants can supply feeds, and complete, current data with richer descriptions, images, prices, availability and reviews helps products appear more accurately and can improve placement. That is a structured commerce pathway. It is not proof that a service business receives any product-feed benefit, and it is not evidence that local recommendations run on the same system. If you sell products, maintain the feed and your own product pages as two separate assets, checking SKU, price, stock, imagery and review data. If you sell services, do not claim a product feed will make ChatGPT recommend you. Keep the distinction explicit in reporting so a commerce feature is never overgeneralized into a service promise.
What is documented, and what we are inferring
Documented by OpenAI: ChatGPT can use learned knowledge or web search; Search may rewrite prompts, use third-party providers, and in Enterprise and Edu may send disassociated queries to Bing; OAI-SearchBot controls search eligibility; ranking uses multiple factors aimed at relevance and reliability; location can shape local retrieval; citations and a Sources panel expose supporting links; and offline workspaces may use cached content with variable freshness. Not documented: the provider mix for every consumer request, a fixed Bing-versus-own-index rule, entity resolution, ranking weights, candidate or review thresholds, and any guaranteed freshness window. Advice about accurate pages, consistent listings, reviews, structured data or third-party mentions is practical inference and independent practitioner guidance, not a verified OpenAI formula. Label every client claim as verified or inferred, and cite the source.
Did this guide answer your question?
See how visible your business is to AI
Get your free 0–100 AEO score across ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews.