Your brand doesn’t appear in ChatGPT, Perplexity or Google’s AI Overviews because “the model learned it”. It appears because a retrieval system finds one of your pages (or one that talks about you) in real time, passes it to the model, and the model grounds its answer in it and cites it. That’s the real path of GEO: retrieval and grounding, not training. Once you understand this difference, you stop chasing myths (“training ChatGPT”) and start working on what actually moves the needle.
The three mechanisms
| Mechanism | What it is | Can you optimise for it? |
|---|---|---|
| Training | The model learns from a massive corpus up to a cut-off date. Its “factory” memory. | Not in practice. You can’t retrain ChatGPT by optimising your website, and it doesn’t generate citations or traffic. |
| Retrieval | The engine searches an index in real time (Google, Bing or its own) and retrieves documents. | Yes. This is the GEO path. It depends on being indexed and accessible. |
| Grounding | The model backs each claim with the retrieved sources and adds the citation. | Yes, indirectly: by making your passages extractable and verifiable. |
Training: what the model knows “by heart”
A language model is pre-trained on an enormous corpus of text up to a knowledge cut-off date. Everything it learns is encoded in its weights: its parametric memory. The relevant point for you: even the newest models are born months behind the present. GPT-5, for example, has a cut-off of 30 September 2024 despite launching in August 2025 (OpenAI’s model documentation).
Why does this mean training isn’t your GEO channel?
- Slow, opaque cycles. You don’t know when your site was crawled, what was filtered out or when the next model will be released.
- No attribution. Parametric memory doesn’t generate links: even if the model “knows” about you, it doesn’t cite you or send you traffic.
- Inaccuracy. That memory is fuzzy; a model can get your prices, services or location wrong.
- They’re separate channels. OpenAI distinguishes between GPTBot (crawling for training) and OAI-SearchBot (crawling to appear in ChatGPT Search). You can block the first and allow the second: you stay out of training but remain eligible to be cited.
“I get clients convinced they need to ‘get into ChatGPT’s training’. It’s a waste of energy. You can’t retrain a model from your website, and even if you could, it wouldn’t cite you. What you do control is whether AI finds you when it goes out to search in real time. That’s where the game is played.” – Rafa Villaplana, SEO consultant and CEO of Linking & Growing
Retrieval: the real path of GEO
Retrieval is the moment when the engine, instead of answering from memory alone, runs live searches against an index and retrieves documents that it passes to the model as context. Each engine draws on a different source:
- Google’s AI Overviews and AI Mode: Google’s index. Its official guidance (2026) says these features surface content from its search index, relying on its core ranking systems.
- ChatGPT Search: a hybrid model that mainly uses Bing’s index plus its own index and cache (fed by OAI-SearchBot). In the antitrust trial against Google (April 2025), ChatGPT’s head testified that it relies on Bing data.
- Perplexity: its own crawler (PerplexityBot) and index, combined with third parties.
- Gemini: uses “Grounding with Google Search”, in other words Google’s index.
The implication is direct and harsh: if you’re not in the index an engine uses, you don’t exist for it. Google is blunt about it: to be eligible for its AI features, a page must be indexed and eligible to be shown with a snippet. And OpenAI warns that sites blocking OAI-SearchBot won’t be shown in ChatGPT Search answers.
Grounding: from “what I remember” to “what this source says”
Grounding is the ability to connect the model’s output to verifiable sources. An answer is grounded when every claim is supported by a retrieved document and can be traced back to it. Google’s 2026 guidance explicitly treats the two concepts as one: retrieval-augmented generation (RAG), also known as grounding, improves quality, accuracy and freshness by showing prominent, clickable links to the pages that support the information.
What matters for you: citations happen at passage level. The system links specific sentences in the answer to specific chunks of your page. It doesn’t cite “your website”: it cites a paragraph that answers a sub-query well.
RAG step by step in an AI search engine
Here’s the full process, in marketing language rather than engineering:
- Deciding to search. The model decides whether the question needs external information.
- Rewriting and query fan-out. It breaks your question down into several sub-queries. Google describes it as a set of concurrent, related queries generated by the model: for a question about fixing a lawn full of weeds, it runs searches such as best weed killers, removing weeds without chemicals and how to prevent them. A note for non-English markets: according to a Peec AI study (Search Engine Journal, February 2026), almost 78% of prompts in languages other than English included at least one sub-query in English, so English-language content in your industry can also compete for your topics.
- Retrieval. Each sub-query runs against the index and returns candidates.
- Reranking. Candidates are reordered by relevance, quality and freshness. What competes here isn’t your whole page, but its passages.
- Synthesis with citations. The model writes the answer grounded in the chosen passages and attaches a citation to each claim.
This fan-out mechanism explains the most revealing figure in today’s landscape: only 37.9% of the URLs cited in AI Overviews are in the top 10 for that same query (Ahrefs, March 2026), down from 76.1% in July 2025. Citations increasingly come from the SERPs of the sub-queries, not the main query.
What the data says (2025–2026)
- Citations are decoupling from rankings: 37.9% of AI Overview citations are in the top 10 (Ahrefs, March 2026). On average, only ~12% of URLs cited by assistants are in Google’s top 10 for the original prompt (ChatGPT, 8%; Perplexity is the exception at 28.6%).
- AI visibility is volatile: an AI Overview lasts 2.15 days on average and has a 70% chance of changing between observations (Ahrefs). Running the same prompt again returns the same list of brands less than once in 100 times (SparkToro + Gumshoe, 2026).
- Brand mentions carry more weight than backlinks: a correlation of 0.664 vs 0.218 with visibility in AI Overviews; YouTube mentions, at 0.737, are the strongest single factor (Ahrefs, 75,000 brands).
- Freshness matters: URLs cited by AI are 25.7% newer than organic ones (Ahrefs).
“The fact that your position changes every two days scares a lot of people, but I read it the other way round: it means the door is open. In classic SEO, knocking the number one off its spot takes months. In AI, if you build the right content and mentions, you can get in within weeks. You do have to measure differently, though: not your position on a given day, but how often you appear on average.” – Rafa Villaplana, SEO consultant and CEO of Linking & Growing
Five myths worth burying
- “I can train ChatGPT on my brand by optimising my website.” False. Optimising your website doesn’t touch any model’s weights; it makes you retrievable.
- “If ChatGPT doesn’t know me, I need to get into its training.” No. Being indexed and allowing OAI-SearchBot is enough.
- “I need llms.txt or special schema for AI.” Google states that no special machine-readable files, AI text files, markup or Markdown are needed, and Ahrefs found no increase in citations after adding schema to 1,885 pages.
- “If I’m in the top 3, I’ll show up in AI.” It helps, but 62% of AI Overview citations come from outside the top 10.
- “Buying mentions will make me visible.” Google warns that inauthentic mentions are filtered out by the same anti-spam systems.
What to do with all this (checklist)
- Audit technical eligibility per engine: indexing in Google (with snippets allowed) and Bing, inclusion in AI features in Search Console, and robots.txt and your WAF allowing OAI-SearchBot and PerplexityBot.
- Map the fan-out of your intents: for each service, list the likely sub-queries (comparisons, prices, alternatives, “how to choose”) and cover them.
- Rewrite for extractability: a direct answer at the start of each section, sourced data, expert quotes, tables, self-contained definitions.
- Invest in first-party data: original studies and figures are content no model can generate and that others will cite.
- Build a presence beyond your website: trade press, YouTube, reviews, communities. These are the sources fan-out retrieves.
- Measure by frequency, not position: how often you appear on average, per engine, plus Search Console’s generative AI report.
Frequently asked questions
Can you “train” ChatGPT to talk about your brand?
Not from your website. A model’s training happens in closed cycles run by its maker and can’t be influenced by optimising your site. What you can achieve is being retrievable: when ChatGPT searches in real time, it finds your page and cites it.
What’s the difference between GPTBot and OAI-SearchBot?
GPTBot crawls content for training OpenAI’s models; OAI-SearchBot crawls so your page can appear in ChatGPT Search. They’re separate channels: you can block training and still be eligible for citations in search.
Why does AI cite a page of mine that isn’t even on Google’s first page?
Because of query fan-out: AI breaks your question down into sub-queries and retrieves results for each one. Your page can rank well for a sub-query even if it isn’t near the top for the main query. That’s why only 38% of AI Overview citations are in the top 10.
Do you need an llms.txt file for AI to cite you?
There’s no evidence that it helps. Google has said it doesn’t need special files for AI, and the available studies find no relationship between having an llms.txt and getting more citations.
How do I measure whether AI is citing me?
With two layers: dashboards that run your key prompts many times and measure the percentage of times you appear (not your position on a given day, which changes constantly), and Google Search Console’s generative AI performance report, which shows impressions in AI Overviews and AI Mode.