Answer engine optimization,
and where it stops working.
AEO is the practice of structuring content so an AI system picks it as the basis for an answer. The techniques are real and mostly old. What almost nobody publishes is the boundary: the point where a perfectly structured page still does not get quoted, and what decides that instead.
Which pages assistants actually pull into answers.
This is the part that changes what you should structure. These are the most-fetched URLs from live-answer traffic on two unrelated live sites — requests made by an assistant while composing a reply for a real person, background crawlers excluded.
| Page | Site | Live fetches |
|---|---|---|
| A "most reliable brands" data page | Site A | 33 |
| A statistics page built from the client's job records | Site A | 31 |
| Homepage | Site A | 28 |
| A brand-specific component failure article | Site A | 18 |
| A brand failure-rate data page | Site B | 20 |
| A "most reliable brands" data page | Site B | 15 |
Windows of eight and five days. Site A: 3,498 AI agent requests, 207 of 1,703 crawled URLs pulled into answers. Site B: 3,178 requests, 53 of 1,290. Service and location pages made up the large majority of what was crawled and almost none of what was consulted — so the highest-return AEO work is usually not on the pages a business most wants to promote. Full breakdown →
Five rules that make a passage quotable.
Answer in the first sentence, then explain
An extractable passage does not build up to its point. The heading asks a question, the sentence underneath answers it completely, and the paragraphs after it are the reasoning. Most pages are written the other way around, which makes them unquotable without the surrounding context.
Attach a base to every number
Not "most customers decline expensive repairs" but a percentage across a stated number of visits. A figure with a denominator can be lifted intact and stays true outside your page. A figure without one is unusable to a machine and unverifiable to a human — which is also why numbers with bases attract links and numbers without them do not.
Make each passage survive being cut out
The test I use: delete everything above and below a paragraph. If it still makes sense and still supports its claim, it can be quoted. If it depends on "as mentioned above" or a chart three screens up, it cannot.
Use structure the parser already understands
Question-shaped headings, tables instead of prose for anything comparative, FAQ schema mirroring the visible questions word for word. None of this is new — it is the featured-snippet playbook, and it works here for the same reason.
Then stop, because structure is only half
Formatting makes a page quotable. It does not make an engine choose you over the three other quotable pages answering the same question. That choice leans on consistency of your details across the web, third parties saying the same thing, and having facts nobody else has. That part is not a formatting job.
What AEO will not fix
- Structuring a page you have no authority to make claims on. Extraction is a mechanism, not a substitute for having something to say.
- FAQ blocks written to fill a schema slot. "It depends" never gets quoted, and a schema whose text does not match the visible page is a liability.
- Optimising service and location pages for extraction. In two site logs, assistants pulled almost exclusively data pages; service pages were crawled and then never consulted.
- Treating AEO as separate from having something citable. Splitting them into two invoices is a pricing decision, not a methodology.
The measurement half matters as much as the writing half, because a structured page either starts getting fetched or it does not, and that answer arrives in days rather than quarters. How the tracking works →
Answer engine optimization, asked plainly.
Structuring content so an AI system selects it as the basis for a specific answer: a direct answer in the opening sentence, question-shaped headings, tables instead of prose for comparisons, self-contained passages, and schema that mirrors the visible page. It is the narrower half of the work — the answer layer — where generative engine optimization is the broader question of whether your business gets named at all.
The techniques largely predate the acronym — this is the featured-snippet playbook. What changed is the payoff. Extraction used to win you a snippet above ten blue links; now it decides whether a model has a clean passage of yours to quote in an answer that shows no links at all. Same craft, higher stakes, and a different measurement problem, since there is often no click to count.
Barely, and I have the logs to say so. Across two live sites, the URLs assistants pulled live to answer real questions were overwhelmingly data pages — brand reliability, failure rates, cost figures. Service and location pages made up the majority of what got crawled and almost none of what got consulted. Structuring a page that has no factual claim to make does not create one.
Two instruments. Your own server log tells you whether assistants started fetching a specific URL after you rewrote it — that moves within days and is a census rather than a sample. A prompt-monitoring tool tells you whether you appear in answers, sampled from outside. Neither replaces the other, and both are more honest than the screenshot of a chatbot that most of this industry runs on.
Yes, with two conditions: the schema text matches the visible questions word for word, and there is exactly one FAQPage block on the page. I have found live pages serving two different FAQ blocks with contradicting questions, which is worse than having none — it makes the page an unreliable source at exactly the moment it is being read as one.
No, and I would be suspicious of anyone who does. On a real engagement the sequence is: make the page extractable, make its claims corroborated, then measure whether assistants actually pull it. Selling step one alone produces beautifully structured pages that nobody quotes.
Find out whether you own anything an engine would quote.
For most sites the honest answer is “not yet”, and knowing that is worth more than another month of restructuring pages nobody consults.