Answer Engine Optimisation
Definitions
Answer engine optimisation is the practice of making a source easy for a language model to retrieve, quote correctly, and attribute. The unit of success is not a ranked position but a citation inside an answer the user reads instead of visiting the page.
The three terms in circulation overlap and are often used interchangeably. They differ in what they optimise for.
| Term | Optimises for | Unit of success | Consumer |
|---|---|---|---|
| SEO | Ranking in a list of links for a query. | Position and click through. | A person scanning results. |
| AEO | Being retrieved and quoted inside a generated answer. | Citation, and whether the quote is accurate. | A model assembling an answer. |
| GEO | The same outcome as AEO. The term is used mostly in vendor marketing. | Citation. | A model assembling an answer. |
The practical difference is the click. Search sends traffic and the page converts it. An answer engine consumes the page and may send nothing, so the return on being cited is reputational rather than measured in sessions. That changes what is worth optimising.
How Answers Get Built
Being indexed is not being cited. A source passes through four stages, and most optimisation advice addresses only the first.
A crawler fetches the page. Separate crawlers exist for training corpora, for live search grounding, and for user initiated fetches, and they can be permitted independently.
What a publisher controls: robots.txt directives per user agent, and server side rendering so content exists without executing JavaScript.
The page is split into passages and stored, usually with both lexical and vector representations. Passages are indexed independently of the document they came from.
What a publisher controls: Heading structure, since it usually determines where the splits fall.
A query pulls candidate passages. The competition here is against every other page in the corpus that covers the same ground, not against the rest of your own site.
What a publisher controls: Vocabulary that matches how the question is asked, and coverage of the specific question rather than the general topic.
The model chooses which retrieved passages to use and whether to attribute them. Retrieval without selection produces no visible outcome at all.
What a publisher controls: Self contained passages, stated scope, explicit dates, and unambiguous claims that are easy to quote without distortion.
Selection is the stage that decides most outcomes and the one publishers influence least directly. A model choosing between five retrieved passages favours the one that answers the question completely in the fewest tokens, states its scope plainly, and carries a date.
Signals That Matter
| Signal | What it does | Status |
|---|---|---|
| Server rendered HTML | Content exists in the initial response. Crawlers that do not execute JavaScript see the full text rather than an empty shell. | Established |
| Structured data | Schema.org types such as TechArticle and BreadcrumbList give explicit author, date and headline rather than leaving them to be inferred. | Established |
| Canonical URLs | Consolidates duplicate addresses so a citation resolves to one place. | Established |
| Freshness metadata | A visible and machine readable revision date lets an answer qualify its claims and lets a retriever prefer current material. | Established |
| Per crawler robots rules | Grants or refuses access separately for training, search grounding and user fetches. Refusing a crawler reliably removes you from that surface. | Established |
| Sitemap with lastmod | Advertises the full URL set and signals which documents changed, improving recrawl targeting. | Established |
| llms.txt | A markdown index of the site intended for model consumption. Cheap to generate and harmless to publish. | Unconfirmed |
| Content licensing declarations | States reuse and attribution terms in a machine readable form. Adoption by answer engines is inconsistent. | Unconfirmed |
Treat the unconfirmed signals as cheap insurance rather than established practice. Publishing llms.txt costs an hour and is trivially maintained by a build step. Whether any major answer engine reads it is not established, and claims that it materially changes citation rates are not supported by evidence a publisher can verify.
Chunk Addressability
Retrieval operates on passages, not documents. A long page is split before it is embedded or scored, and the split usually follows heading structure. The page therefore competes as a set of independent passages, each of which has to stand alone.
Three properties decide whether a passage survives that split intact.
The passage makes sense without the paragraphs above it. Pronouns and references that point outside the section break when it is extracted.
Test: Read the section alone. If it opens with "this" or "the above", it will not survive extraction.
The heading describes the content rather than performing. Retrieval frequently weights heading text, and a heading is often the only context a passage carries.
Test: Would the heading alone tell a reader whether this section answers their question?
Long enough to answer something completely, short enough that the answer is not diluted by unrelated material sharing the same chunk.
Test: One section, one question. Two questions in one section means both retrieve weakly.
Descriptive headings do the most work here. A heading that names its content, Failure Modes or Latency Budget, tells a retriever what the passage contains. A heading written as a rhetorical flourish carries no signal and strands an otherwise good passage.
Anchored headings extend this. When each section has a stable id, an answer engine can cite the exact passage rather than the page, and a reader arriving from that citation lands on the paragraph that was quoted.
Attribution
Attribution is the only return on being cited, so the mechanics of it are worth as much attention as the content. Four things determine whether a citation points somewhere useful.
- Canonical URLOne address per document, declared explicitly, so a citation and a reader arrive at the same place.
- Stable anchorsSection ids that do not change when copy is edited. A cited anchor that later moves sends readers to the wrong part of the page or to nothing.
- Explicit authorshipAuthor and publisher in structured data, so attribution names a person rather than a domain.
- Revision datesA machine readable dateModified lets an answer say when a claim was current, which is the difference between being cited accurately and being cited as though stale material were live.
Canonical URLs matter more than they do for search. A model that has seen the same passage at two addresses will cite whichever it retrieved, and duplicate content splits the attribution rather than the ranking.
Being Used, Not Only Read
Everything above optimises for a source being retrieved and quoted. WebMCP addresses the opposite half: letting a site expose callable tools to an agent already in the browser, so it can act rather than only read. The page becomes the tool server, running client side inside the user's existing authenticated session, which is what separates it from the server side protocol of the same family.
| Property | Retrieval and citation | WebMCP |
|---|---|---|
| What the agent does | Reads text and quotes it into an answer. | Calls a declared tool and acts on the result. |
| Where it runs | A crawler or fetcher, outside any user session. | Client side in the page, inside the user's logged in session. |
| Authentication | None, so only public content is reachable. | Inherited from the session, so authenticated actions are reachable. |
| What you publish | Clean semantic HTML, structured data, stable anchors. | A declared set of tools with names, descriptions and parameters. |
| Failure mode | Misquoted, stale, or attributed to someone else. | An action taken that the user did not intend or review. |
| Maturity | Established, and the signals are well understood. | Incubation, origin trial, no stable specification. |
Status matters more than usual here. It is a Draft Community Group Report from the W3C Web Machine Learning Community Group, first published February 2026 and originating in a joint Google and Microsoft proposal. It is explicitly not a W3C Standard and not on the standards track, it is under incubation, and browser support so far is an origin trial rather than general availability. Treat it as something to understand and prototype, not something to depend on.
It also suits transactional sites far better than reference ones. A site where the useful verbs are search, filter, book or purchase has obvious tools to expose. A knowledge base mostly has one, which is search, and an agent can already reach that through ordinary retrieval. The honest position for a content site is that the retrieval work above pays now and this does not yet.
Failure Modes
The page renders correctly in a browser and returns an empty shell to any crawler that does not execute JavaScript. The site appears healthy in every human check.
Response: Prerender or server render. Verify by fetching the raw HTML and searching it for body text, not by loading the page.
A revised page continues to be quoted from an older crawl, so a claim you have since corrected is attributed to you as current.
Response: Accurate lastmod in the sitemap and a visible revision date. Neither forces a recrawl, but both make the staleness detectable.
The substance is used and the citation points elsewhere, or nowhere. Common when the same material exists at several addresses or is syndicated.
Response: Canonical tags, one address per document, and distinctive phrasing that is traceable back to the source.
The page is cited for a claim it does not make, because a summary drifted or two sources were merged. The citation is real, the content is not.
Response: Unambiguous, self contained statements. Hedged or heavily qualified prose is the easiest to distort in summary.
Headings are rewritten, anchors change, and existing citations decay. The page still works, and every deep link into it stops resolving.
Response: Treat anchor slugs as a public contract once external references exist. Rename copy without renaming ids.
Effort is directed at a vendor visibility score that is itself inferred, rather than at retrievability or accuracy.
Response: Instrument what is real, which is crawler activity in server logs, and sample citations manually.
Measurement
The honest position is that this is poorly measurable, and any tool claiming a precise answer engine visibility score is inferring it. Separating what is observable from what is not prevents optimising against a number that does not mean what it appears to.
- Crawler hits per user agent, from server logs
- Which URLs each crawler has fetched, and how recently
- Whether raw HTML contains the content, verifiable directly
- Referral traffic where an answer engine sends it and passes a referrer
- Citations found by asking the engines questions manually
- How often a page is retrieved but not cited
- Which passage of a page was used
- Whether a page is in a given training corpus
- Citation share against competing sources
- Whether any engine reads llms.txt
Server logs are the most reliable instrument available. Crawler user agents identify themselves, so hit rate and coverage per crawler are directly measurable even when downstream citation is not. Start there before buying a visibility product.
The practical stance: do the things that are cheap and defensible on their own merits, structured markup, clean canonicals, descriptive headings, accurate dates, and treat citation share as something to sample manually rather than track.