In this article 6
Short answer
When an AI platform with search answers “which online accountant would you recommend?”, it does four things. It rewrites the question as one or several search queries. It retrieves a handful of documents (web pages) from an index. It reads those documents and writes an answer from them, sometimes mixing in what it already knew. And it shows some of the pages it used as sources. Your website can come in at step two (if it is crawlable and relevant) and at step four (if the model uses it and cites it). It does not get into step three by magic: it gets there because it was in step two.
Step 1. The question becomes queries
The model does not search for your literal question. It breaks it down: “online accountant freelancer Madrid”, “best online accountants 2026”, “online accountant price for freelancers”. That is why one question can trigger several searches and retrieve pages that answer different parts of it.
What it means for you: your page does not need to contain the exact question; it needs to answer one of its parts clearly.
Step 2. Documents are retrieved
From an index: Bing’s, Google’s or the platform’s own, depending on the platform. The system retrieves the pages it considers most relevant and most trustworthy for those queries. This is where the fundamentals rule: the page has to be in the index, it has to be crawlable, its title and first lines have to say what it is about, and the text has to be in HTML.
What it means for you: if you are not indexed or you block search crawlers, you do not get past this point.
Step 3. The model writes
With the retrieved documents and what it learned in training, the model generates an answer. This is where randomness comes in (two runs can differ), and where the model chooses which brands to name and in what order, guided by what the documents say.
What it means for you: if three retrieved pages mention you with clear facts, you are more likely to appear; if they mention you with contradictory facts, it is more likely to describe you wrongly.
Step 4. Sources are shown
Some platforms cite the pages they used: ChatGPT with search, Perplexity, AI Overviews, Copilot. They do not always cite all of them, nor always the same ones. A citation means the page was retrieved and judged useful that day; it does not mean it caused the mention.
What it means for you: the sources cited in your category are your map of where to be.
When there is no search
If the platform answers without searching (ChatGPT with search turned off, a model used through the API without tools), there is only step three, with what the model learned. There are no sources, no possible update until a new version, and your presence depends on what existed when it was trained. It is a different product, and it is measured separately.
Why this makes GEO measurable
Because steps two and four leave a trace: you can see what is retrieved and what is cited. And because step three can be observed by repeating the same question on different dates. What you cannot do is control step three. Nobody can.