How AI assistants decide what to say about your business

When someone asks ChatGPT, Claude, Perplexity or Google's AI Mode about your business, the answer comes from two places: what the AI model learned during training, and what it finds by searching the web at the moment the question is asked. How ChatGPT chooses sources for that second part is only partly public, but the basics are clear enough to act on.

The practical takeaway is simple: AI assistants repeat what the web says about you. If your website, your Google Business Profile and the directories that list you all agree, there's a good chance the answer will be right. If they disagree, or if your site is hard to read, the assistant may fill the gaps with old or wrong information.

Two ways an AI assistant "knows" about you

1. Training data: what the model already learned

Large language models are trained on huge collections of text, much of it gathered from the public web, up to a cutoff date. From that, a model picks up general knowledge, including facts about well-known businesses.

For most small businesses, this layer is thin. A model may know a bakery exists in a certain town, or it may know nothing. What it learned can also be out of date: an old address, a menu you changed two years ago, an owner who has since retired. And because a model generates text rather than looking it up, it can state things confidently that were never true.

AI companies use crawlers to collect training data. OpenAI's is called GPTBot and Anthropic's is ClaudeBot. Both companies say you can opt out of training by blocking them in your robots.txt file, and both say blocking the training crawler is separate from their search features (see OpenAI's crawler page and Anthropic's help article).

2. Live retrieval: what the assistant looks up right now

For questions about current details like hours, prices or "best plumber near me," modern assistants increasingly search the web first, then write an answer from what they find. This is often called retrieval, or grounding.

Here's what each company has said publicly, as of October 2026:

• ChatGPT: OpenAI's help center says ChatGPT may search the web automatically when a question would benefit from current information, and that it sometimes partners with other search providers. To be eligible for ChatGPT's search results, a site must allow OpenAI's search crawler, OAI-SearchBot. OpenAI says it ranks results using multiple factors and that placement is not guaranteed.
• Claude: Anthropic says Claude-SearchBot indexes content to improve search results and Claude-User fetches pages when a person asks Claude a question. Blocking either may reduce your visibility in Claude's answers.
• Perplexity: Perplexity says PerplexityBot crawls the web to index it for its search engine (Perplexity's crawler docs).
• Google AI Overviews and AI Mode: Google's AI features documentation says a page must be indexed and eligible to show with a snippet in Google Search to appear as a supporting link. It also says these features may use "query fan-out": running several related searches on subtopics to build one answer.

Crawler names and policies change, so check those pages directly before you change your robots.txt.

How ChatGPT chooses sources (and what nobody knows)

No AI company publishes its full ranking method, just as Google never published every search ranking factor. What we can say with confidence:

• You have to be reachable. If your robots.txt, firewall or security plugin blocks an assistant's search crawler, that assistant can't read your pages when it searches. This is the one requirement OpenAI, Anthropic and Google all state plainly.
• Search results feed the answer. When an assistant searches, it picks pages from search results, reads some of them and writes a summary. Pages that rank well in ordinary search for the question have a better chance of being read.
• Citations show what was used, not what is true. Many answers link to the pages they drew on. OpenAI's own help page warns that search results and citations can be incomplete, outdated or incorrect.

Beyond that, be skeptical of anyone who claims to know the exact formula. Google's own advice for its AI features is that there are no special optimizations needed beyond normal search best practices, and nobody has shown a secret switch for the other assistants either. Our review of the evidence on llms.txt is a good example: as of October 2026, there's no good evidence that it affects which sites AI search tools cite.

Why consistent facts across the web matter

Imagine an assistant searching for "Harbor Dental Portland hours." It might read your website, your Google Business Profile, a review site and a directory listing. If three say you close at 5pm and one still says 6pm from two years ago, the answer could go either way.

The assistant has no way to call you and check. It works with what it finds. So the most reliable thing you can do is make sure what it finds agrees.

Start with these:

1. Your name, address and phone number. Spelled the same way everywhere: your website footer, contact page, Google Business Profile, Apple Maps, Bing Places, Yelp and industry directories.
2. Your hours, including holiday hours. Update them in every place at once.
3. What you do. If you stopped offering a service, take it off your site and your profiles. Old service pages are a common source of wrong answers.
4. Old pages and old sites. A forgotten landing page or a second domain from a previous rebrand can keep feeding outdated details into search results. Redirect it or update it.
5. Prices, if you publish them. A stale price on one page is easy for an assistant to repeat.

Make your own site easy to quote

Your website is the one source you fully control, so make it the clearest one:

• Say the key facts in plain text. Who you are, what you do, where, for whom, and how to book or buy. Put them near the top of the relevant page, in words, not only inside images.
• Make sure content shows without JavaScript. Some crawlers don't run JavaScript, so text that only appears after scripts load may be invisible to them.
• Use clear headings and short answers. A heading like "Do you take walk-ins?" followed by a direct answer is easy for people and machines to quote accurately.
• Add structured data. Schema markup such as LocalBusiness or Organization states your name, address, hours and contact details in a format machines read reliably.
• Publish an llms.txt if you like. It's a short summary of your site and key pages. It's cheap and it can't hurt, but treat it as housekeeping, not a strategy. Our guide to what goes in a great llms.txt file shows examples.

What to do when an assistant gets you wrong

First, find the source. If the answer includes citations, open them: the wrong detail is often sitting on one of those pages. Fix it there if it's yours, or contact the site owner or directory if it isn't.

If there's no citation, the error may come from training data, which you can't edit directly. The best you can do is make sure the current, correct facts are clear and consistent everywhere, so that when the assistant does search, it finds the right answer. Most assistants also have a feedback button on each answer; it's worth using.

Don't expect instant changes. Search indexes and AI products update on their own schedules.

Check what crawlers can see

Our free AI readiness check looks at the basics that decide whether assistants can read your site at all: whether AI crawlers are allowed, whether your content is visible without JavaScript, structured data, your sitemap, and your page titles and descriptions. It takes a minute and doesn't need an account.

Bottom line: AI assistants combine what they learned in training with what they find by searching. You can't control the first, and nobody outside these companies knows exactly how the second ranks sources. What you can control is being reachable, being clear, and making sure every place that describes your business says the same thing.