GEO courseModule 1 of 10: Start HereLesson 1.2
How AI Answers Are Built, and Who’s Reading Your Site
An AI assistant answers either from what it learned in training, or by searching the web live and writing an answer from what it finds. Which one happens, and which crawler fetched your pages, decides whether you can appear at all.
01What it is
A quiz contestant either answers from memory or phones a friend who looks it up. A memory answer can be out of date. A looked-up answer is only as good as what the friend finds.
Memory or live search
- From training: the model learned from a huge body of text up to a cutoff date. If your agency opened, rebranded or changed its fees after that, or was rarely written about, it may know nothing, or something out of date.
- From live search (often called grounding): the assistant searches the web, reads the results and writes from them. ChatGPT, Claude, Gemini, Copilot and Perplexity all do this when they judge it helps, and you can usually switch it on.
- For testing, this matters: if the assistant didn't search, it can only mention you if you were in its training data. Switch search on where you can, and note whether it was used.
From your page to an AI answer, in six steps
| Step | What happens | What goes wrong |
|---|---|---|
| 1. Crawl | Bots such as Googlebot, Bingbot and OAI-SearchBot fetch your pages | robots.txt, a firewall or a CDN setting blocks them without anyone deciding to (lesson 3.1) |
| 2. Index | The page is stored so it can be found | A stray noindex, a duplicate, or content hidden behind scripts or logins (lessons 3.2 to 3.4) |
| 3. Retrieve | Someone asks; the system searches for matching pages | Your facts are unclear, or you're confused with another agency (module 4) |
| 4. Select | It picks the passages and sources it trusts | A portal, review site or rival is chosen instead of you (module 7) |
| 5. Summarise | The model writes a short answer | It merges two agencies or repeats an old fee (module 8) |
| 6. Cite | It may show its sources | No sources shown, or someone else's page about you |
Query fan-out: one question becomes several searches
Google confirms that AI Overviews and AI Mode can split a question into several related searches and combine the results. Ask "should I use a letting agent or manage it myself?" and it may quietly search for letting fees in your town, landlord legal duties, full management against let-only, and reviews of local agents. Pages that answer those sub-questions can be used even if they never match the original wording. Lesson 6.2 turns this into a plan.
The platforms, and their crawlers
| Platform | What reads your site | Worth knowing |
|---|---|---|
| Google AI Overviews and AI Mode | Googlebot (normal Search crawling) | Part of Google Search. No special file or markup needed. A Search Console switch can exclude your site (lesson 3.3) |
| Gemini app | Google's index, controlled by the Google-Extended setting | Google-Extended covers Gemini training and grounding. It doesn't affect Google Search |
| ChatGPT | OAI-SearchBot (search), GPTBot (training), ChatGPT-User (a person's request) | Block OAI-SearchBot and OpenAI says you won't be shown in ChatGPT search answers |
| Microsoft Copilot and Bing | Bingbot | Bing Webmaster Tools now reports how often Copilot cites you |
| Perplexity | PerplexityBot (search), Perplexity-User (a person's request) | Shows numbered sources, which makes it easy to see who it used |
| Claude | Claude-SearchBot (search), Claude-User (a person's request), ClaudeBot (training) | Every web-search answer comes with citations |
| Apple (Siri, Spotlight, Safari) | Applebot | Applebot-Extended only controls training; it doesn't crawl anything |
Different tools, different sources
BrightLocal's study of about 200,000 local searches (September 2026) found Google Business Profile was the single biggest source across Google's AI features and ChatGPT combined (about 29% of citations), but ChatGPT itself leaned on Yelp and Bing and barely used Business Profiles. ChatGPT and Google's AI features named the same businesses only about a fifth of the time. It's a vendor study with no UK breakdown, but the lesson travels: test each platform separately, and never assume one tells you about the rest.
02Keep it up
That's lesson 1.2
Worked through it? Mark it done and we’ll keep track on this device.
← Back to 1.1 What GEO Is: Mentions, Citations, Recommendations and Clicks