1. Two stages that are constantly confused
A generative answer arises in two separate steps, and they fail for different reasons:
- Retrieval. The engine looks for documents that fit the question and places them in its working context. This is where it is decided whether your page is considered at all.
- Citation. From the fetched documents the model chooses what it uses and names. This is where it is decided whether you appear in the answer.
The critical survey of GEO research from 2023 to 2026 explicitly demands that both stages be reported separately — because the literature has so far mixed them, and thereby makes statements about the second stage that get read as statements about both.
Almost every GEO recommendation you will find concerns stage two. Almost every problem a website actually has sits in stage one.
2. The fallacy: it is not about a top-10 ranking
This is where the shortcut usually goes wrong: "So I need good rankings first, then GEO." That is not how it works, and the numbers show it clearly.
A 2026 investigation found that 53 per cent of the domains cited by Google AI Overviews are not in the organic top 10. Classic placement and citation in the AI answer are precisely not neatly correlated.
The condition is a different and a lower one: not high up, but retrievable at all. That is exactly the part of technical SEO that counts — and the quietest, because it shows nothing when it is missing.
3. What retrievability means in practice
Four levels decide whether a page is a candidate for retrieval. None of them is new; all are older than GEO:
- Access. Is the crawler of the respective answer engine allowed to read at all? The control for that is `robots.txt`, and the rules for AI crawlers are different from those for the classic search engine crawler. Anyone who blocks wholesale or leaves everything open wholesale has, in both cases, not made a decision but inherited a default. We have measured this elsewhere: llms.txt in 2026: barely read — and what actually works instead.
- Delivery. Does the server answer with 200, and reliably and quickly enough? A timeout on retrieval is, for the machine, the same as a page that does not exist.
- Readability without detours. Is the content in the delivered HTML — or does it only come into being once a browser executes scripts? For classic search this has been a risk for years; for retrieval by answer engines it is a risk you do not need to take.
- Unambiguity. Canonical address, clean language markup, no two addresses for the same text. What is ambiguous gets cited inconsistently — if at all.
4. Why text optimisation without this basis can do harm
It does not stop at "ineffective". An arena study of 2026 measured what happens when only the body text is optimised: presence in the top 20 fell by around 9 per cent, top-10 presence after reranking by 16 per cent, actual citation by 6 per cent.
The mechanism behind it is obvious as soon as you keep the two stages apart: a text that is rewritten for the answer stage — denser in claims, poorer in what carries retrieval — gets worse at the retrieval stage. And the retrieval stage comes first.
5. Part of the answers cannot be reached at all
One more number that sorts expectations: in 57.8 per cent of the evaluated ChatGPT runs no search was triggered at all. The model answered from what it had learned anyway.
For planning this means: a substantial part of the answers in your subject area cannot be influenced by any measure on your website — neither technical nor editorial. Anyone who is clear about that asks the right question: not "how do I get into every answer", but "am I retrievable when a search happens".
6. The order that follows from the research
Not as a wish list, but as a sequence — each step is what makes the next one meaningful:
- Settle access. Decide which AI crawlers may read, and record the decision in `robots.txt`. Deliberately, not inherited.
- Check delivery. Status codes, response times, redirect chains. Whatever sticks here sticks for all stages.
- Secure the content in the HTML. Whatever only appears after script execution, free from that dependency.
- Establish unambiguity. Canonical addresses, clean language references, no duplicates.
- Only now work on the text. And then on the one lever that is robustly proven: answer the question concretely and completely.
- Measure with repetition. Seven to eight queries per question as a minimum, several phrasings, several points in time.
7. How you can tell that stage one is the problem
Three signs that are reliable in practice:
- Your page ranks decently in classic search but appears in no AI answer. Then it is rarely the text.
- The answer engines cite the same two or three third-party sources on your topic throughout, regardless of phrasing. Then your document is probably not in the context at all.
- Your server logs show visits from the classic search engine crawlers but hardly any from the AI crawlers. That is the most direct measurement you have, and it costs nothing but a look at the log.
8. What can be measured and what cannot
Finally, the boundary you should know. The critical survey records that there is so far no study showing a stable causal effect on organic findability that holds over time and across several platforms. Anyone who sells you such an effect is selling you something the research does not provide.
What you can measure, by contrast, is the condition: does the crawler get in, does it receive a clean answer, is the content in the HTML. Those are questions with a yes-or-no answer, and they can be settled in an afternoon. For visibility in AI answers they are currently the only part you truly control.
9. FAQ: frequently asked questions about GEO and technical SEO
Do I need good Google rankings first before GEO works?
No. 53 per cent of the domains cited by Google AI Overviews are not in the organic top 10. The condition is not placement but retrievability: the crawler must be allowed to read, the server must answer cleanly, and the content must be in the HTML.
Why does text optimisation not work for me?
Because all known GEO effects were measured on documents the engine had already retrieved. If your page is not retrieved, none of these measures takes hold. Check access, delivery and readability without script execution first.
Can wrong optimisation lower my visibility?
Yes. A 2026 study measured, for optimisations that touched only the body text, a drop in top-20 presence of around 9 per cent and in citation of 6 per cent. A text rewritten for the answer stage can lose at the retrieval stage.
Do I have to allow AI crawlers in robots.txt?
That is a business decision, not a technical one. What matters is that it gets made: anyone who enters nothing leaves it to the default of their provider or their protection service — and then does not know who is reading.
How do I see whether AI crawlers visit my site at all?
In the server log. That is where the requests appear with their identifier, and where you can see without detour which crawlers come, which status codes they receive and how quickly they are answered. It is the most direct measurement available.
Is an llms.txt enough for visibility in AI answers?
No. According to our own measurements the file is hardly ever fetched; the crawlers fetch HTML. What controls who reads is `robots.txt` — not `llms.txt`.
10. Sources and status
Status of the evaluation: September 2026.
- Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023–2026), arXiv:2607.14035 — source for the separation of retrieval and citation, the 53 per cent on AI Overviews domains, the 2026 arena values, the share of runs without search and the finding on missing evidence for organic findability.
First part of this series: Generative Engine Optimization without the myth — what the research actually knows about AI citations. On the technical foundation: llms.txt in 2026: barely read and Search Console generative AI report: measuring your AI visibility.
