How GEO works
The mechanism Google documents, and the part nobody outside the labs can see.

Generative engine optimization works by making your pages easy to find, retrieve and quote at the moment an AI system answers a question. There are two kinds of evidence for how that happens: what Google and OpenAI publish, which is short and mostly about access, and everything else, which is inference. This note keeps the two apart.

The mechanism, in four steps

Google describes its AI features in terms that match what most answer engines appear to do:

1. The question is expanded. Google calls this query fan-out: the model generates a set of related searches at the same time, so one question becomes several.
2. Pages are retrieved. The results come from search, not from the model's memory. Google says this grounding is rooted in its core Search ranking and quality systems, so a page generally has to be indexed, and eligible to show a snippet, before it can be used.
3. Passages are selected. The system reads what it retrieved and picks the pieces that answer each part of the question.
4. An answer is written, with some sources shown. Not every source that was read is shown, and not every page that is shown was actually relied on.

OpenAI documents less. Its public material covers crawler behaviour: a site that blocks OAI-SearchBot is not eligible for ChatGPT's search results, and its different crawlers are controlled separately. We covered what that does and does not tell you in ChatGPT SEO: what OpenAI documents.

What Google says you do not need

This is the part that most GEO advice contradicts. Google's guide to generative AI features says, as we read it:

• Structured data is not required for generative AI search, and no special schema is needed.
• You do not need new machine-readable files, AI text files, special markup or Markdown to appear.
• llms.txt is not used by Search. Creating one will neither help nor harm.
• Rewriting content in small "chunks" for AI is unnecessary.

Google also notes that "No third-party tool has access to our internal ranking or AI systems." That applies to every visibility score you will be sold, including anything we report.

What the models said today

On 16 October 2026 we asked ChatGPT (GPT-5.4) and Perplexity (Sonar Pro), both with web search on: "How does generative engine optimization actually work? What is documented by Google or OpenAI versus what is just speculation?" We ran it through DataForSEO's API, once per model.

• ChatGPT called GEO "mostly an industry label". It said that for Google it reduces to SEO plus content quality and crawlability, and it listed llms.txt and chunking as unsupported tactics, citing Google.
• Perplexity gave a seven-step retrieval pipeline and sorted common advice into documented, supported-but-not-guaranteed and speculative. Its ten sources were a mix of vendor blogs, Wikipedia and academic papers. It mentioned a study reporting visibility gains from adding statistics and quotations. We have not checked that study, so we do not repeat its numbers.

Both models, asked directly, were more sceptical than most of the pages ranking for this topic. That is worth knowing before you pay for a proprietary "GEO framework".

What is supported but not guaranteed

Nobody publishes a weighting for these, so treat them as good practice, not mechanics:

• Put the answer in the first two sentences, then the evidence.
• Keep one fact stated one way across your site and the profiles buyers use.
• Put the content in the HTML, not behind a script.
• Link to primary sources, and date what you publish.

All of these also help ordinary search, which is the point. For us, GEO is mostly good technical SEO and pages written to be quoted, plus checking what the models say about you.

What costs us something

If Google is right that no special markup is needed, then part of what agencies sell as "GEO implementation" is optional, including parts of ours. We still add structured data, because it is cheap and it helps ordinary rich results. We do not claim it moves AI answers.

Nor can anyone promise you a citation. Answers vary by prompt, by location and by day. The only honest measure is a fixed set of buyer questions, asked repeatedly, with the results recorded. That is what the generative engine optimization work starts from, and why we put the same check in front of every note on this site.