Challenge · AI search visibility
Answer engines do not rank you. They quote you.
Optimising for a ranking gets you a position in a list. Optimising to be quoted gets you named inside somebody else's sentence, which is a different job.
20 August 2026 · 8 minute read
A search engine ranks pages and a person chooses one. An answer engine extracts passages and writes a paragraph, naming a few brands. Being ranked well helps you be extracted, but it is not the same thing, and the writing that gets extracted has specific properties: it answers the question completely in its first fifty to two hundred words, it is structured in self-contained blocks that survive being lifted out of context, it carries specific numbers rather than adjectives, and it cites sources. Published research on generative engine optimisation found expert quotations and explicit citations to be the two highest-impact changes, at around 41% and 40% improvement in visibility respectively.
What actually happens when someone asks an AI a question?
The question gets expanded into several sub-queries, documents are retrieved and filtered by authority and by how much new information they add, passages of roughly a hundred and twenty to a hundred and eighty words are pulled out of the survivors, and an answer is synthesised from those passages with attribution.
Every stage of that is a filter you can fail. You can fail retrieval by not ranking. You can pass retrieval and fail extraction because your page has no self-contained passage that answers anything. You can pass extraction and fail attribution because the passage does not name you.
How should a page be structured differently?
- Answer in the first paragraph. Not a build-up, not context, not a scene. The complete answer to the question the page is about, in the first fifty to two hundred words. This is the passage most likely to be quoted.
- Write in self-contained blocks. Each section should make sense lifted out on its own. If a paragraph depends on the one above it, it cannot be extracted without breaking.
- Make headings questions. Real questions, phrased the way a person would ask them, because that is what gets matched.
- Use numbers. "Significantly faster" is unquotable. "40 to 60 percent since 2023" is a sentence a model can lift.
- Name yourself in the extractable passage. A passage that answers well but never names the brand gets you extracted and not attributed, which is the worst outcome: you did the work and a competitor gets named.
- Cite sources. Explicit citation to reputable external sources is among the highest-impact changes measured.
The counter-intuitive one
Keyword stuffing actively harms performance in generative engines, where it used to be merely useless in search. Repetition reduces the information density of a passage, and information density is what the extraction step is selecting for.
Why does information gain matter so much?
Generative engines are trying to avoid redundancy. If your page says what forty other pages already say, it adds nothing to the answer and there is no reason to select it. This is measured as how far a document diverges from everything else already covering the topic.
The practical consequence is uncomfortable for most content teams: a well-written summary of what is already known is close to worthless. Original data, a genuinely new argument, or a first-hand account are what get selected, which is a higher bar than most content calendars are built for.
Does traditional SEO still matter?
Yes, and more than the framing of this article might suggest. Roughly 38% of citations in Google's AI Overviews come from URLs already ranking in the top ten organic results. Ranking is not sufficient and it is close to necessary.
The honest summary is that this is additive rather than a replacement. You still need the fundamentals; you now also need the page to be extractable once the fundamentals have got it retrieved.
The hardest part of this for a content team is not the format. It is that summarising what is already known stopped being valuable. If your page adds nothing, there is no reason for a model to select it.
How do you know if any of it worked?
By asking the questions and reading the answers, repeatedly, and recording which brands get named and which sources get cited. There is no impression count and no click for an answer you were not in, so the only way to measure absence is to look for it deliberately.
Do this monthly at minimum. The channel is volatile in ways that have nothing to do with your content, as August 2026 demonstrated when Reddit lost about 86% of its ChatGPT citations in four days because of a change nobody outside OpenAI knew was coming.
If you are an agency
This is a genuine service line and one clients are actively asking for. It is also one where the deliverable is unusually clear: a set of tracked category questions, a monthly record of who gets named, a diagnosis of why not, and content built to fix it. Be careful about promising outcomes; the sensible commitment is measurement and remediation, not guaranteed inclusion.
How we do this ourselves
The AEO/GEO Monitor tracks the questions your buyers ask across ChatGPT, Perplexity, Gemini and Google AI Overviews, keeps the full answers with their sources, diagnoses why you are absent, and produces the briefs, drafts and structured data to fix it. It then re-checks against evidence rather than asserting the fix worked. A market is always required, because an answer for the wrong country is an answer about somebody else's customers.
How the AEO/GEO Monitor works