How does AI even find out about my business?

AI gets its information about businesses from four routes, each with its own personality:

1. Training data. The web as it existed when the model was trained: sites, articles, discussions, directories. This is the engine's long-term memory, refreshed only when models update, which means it can lag reality by months.
2. Live web search. Search-native engines like Perplexity and Google's AI answers, and increasingly the assistants too, search at question time and read current pages before answering. This is where recent changes surface fastest.
3. AI crawlers. Dedicated bots, GPTBot for OpenAI, ClaudeBot for Anthropic, PerplexityBot and others, visit websites directly to collect and refresh what the engines know.
4. Licensed and structured sources. Paid deals for high-value archives, most famously Google's arrangement to train on Reddit's human conversations, plus the directories and platforms engines treat as reliable.

The strategic takeaway is that these routes cross-check each other. An engine reads your site, then looks for the rest of the web to agree. A business visible on one route and absent from the others reads as thin, which is why the fix is a consistent record everywhere the routes look, never a single page.

AI crawlers are almost certainly visiting your website already, at a scale most owners have never checked. Vercel's analysis of traffic across its own hosting network measured AI crawlers as a major new presence on the web: OpenAI's GPTBot alone made 569 million fetches across that one network in a single month, with the other AI crawlers adding substantial volume of their own, already a meaningful fraction of what Google's own crawler does.

Three practical things follow:

• You can verify it tonight. Your hosting or analytics logs will show user agents like GPTBot, ClaudeBot, PerplexityBot, and Google-Extended. Their presence means the engines are actively reading you; their absence is worth investigating.
• You control the door. Your robots.txt file can welcome or block each crawler by name. Blocking has legitimate uses, but for a business that wants to be recommended, blocking the crawlers is asking to be forgotten.
• A visit is not comprehension. Crawlers fetch what your site serves; whether they can extract meaning from it is a separate question about structure and clarity, and it's the more common failure.

Third-party sources teach AI the part it trusts most: what other people say when the business isn't in the room. Your own website is testimony; the third-party record is the cross-examination, and engines weigh it accordingly.

The evidence for how seriously the industry takes human discussion is written in contracts. Google pays Reddit a reported $60 million a year specifically to train its AI on the platform's conversations, because millions of people candidly discussing what worked, what failed, and who they would recommend is exactly the ground truth engines cannot generate themselves.

Beyond the marquee deal, the third-party record engines read includes:

• Reviews, on the platforms your industry actually uses.
• Directories and professional listings, which confirm existence, category, and location.
• Press, articles, and podcast appearances, which attach your name to your expertise on domains with their own authority.
• Community discussion, forums, Reddit, professional groups, where recommendations happen in the wild.

Each mention is a witness. A business its own website describes one way and the third-party record confirms is verifiable; a business only its own website describes is a claim waiting for corroboration.

You can't tell AI about your business directly, at least not in the way owners hope. There's no portal where you register with ChatGPT, no form that updates Claude, no fee that gets you into Perplexity's answers. The engines deliberately learn from the open record rather than from self-submission, because self-submission is exactly the channel spam would flood.

What does exist is narrower and still worth doing:

• Make your site maximally legible to the crawlers that visit: clear answers, structured data, no walls between them and your content. This is the closest thing to telling the engines directly.
• An emerging convention called llms.txt lets a site offer AI systems a plain-text guide to its most important content. Adoption is early and uneven, but it costs little and signals current stewardship.
• Feed the surfaces engines already trust: complete professional profiles, accurate directory listings, active presence where your industry gets discussed.

The mental shift that helps: publishing evidence where the routes already look IS how you tell the engines, and the routes carry it in. Anyone selling guaranteed AI placement is selling something the architecture doesn't offer.

To find out what AI currently knows about you, ask the engines themselves, the same way a prospect would, and audit what comes back in layers.

1. The direct lookup. 'Tell me about [your business]' on two or three engines. You're grading four possible states: rich and accurate, thin, outdated, or wrong, and each points at a different intake route failing.
2. The buyer's question. 'Who should I hire for [what you do] for [who you serve]?' Whether you appear, and what reasons get attached, shows how your evidence competes, not just whether it exists.
3. The source check. On engines that cite, note where their information about your category comes from. Those domains are the third-party surfaces worth showing up on.
4. The crawler check. Your own logs, for GPTBot, ClaudeBot, and friends, confirm whether the reading route is even reaching you.

Run the layers quarterly and screenshot everything, because the answers move as the routes refresh. Running all four layers systematically, across engines, with the failing route identified and the fix ordered, is exactly what our free AI Visibility Scan does.