AEO skill spec: what makes an answer engine cite your page

This is the specification for the answer engine skill: making a page usable as a source when somebody asks an assistant a question instead of typing it into a search box. The assistants in scope are the general ones people ask questions of all day, ChatGPT, Claude, Perplexity and Gemini, along with the answer panels that search engines now place above their own results.

An assistant answers from two different places, and the difference decides what you can influence.

What the model learned

Everything absorbed during training, held as parameters rather than as documents. There are no citations here and no live updates: the model produces what its training made likely, up to its cutoff. What the web says about you before that point is part of what it knows.

What the model fetched

Text pulled in at the moment of the question, by searching an index, by crawling, or by opening a URL the user supplied. This is where citations come from, and it is the path a published page can enter within hours rather than at the next training run.

On the retrieval path the sequence is roughly the same everywhere, even though the implementations are not public. A query is derived from the conversation, often rewritten and often split into several. Candidate documents come back. Those documents are broken into passages, and passages, not pages, are what gets ranked and passed to the model. The model then writes an answer over the passages it was given and attaches citations to the sources it used.

Two consequences follow directly, and they drive the rest of this spec.

A page has to clear three separate bars, in order. Failing any one of them makes work on the other two worthless, which is why the skill checks them in this sequence.

• Retrievable, the assistant is allowed and able to fetch the page, and something in the index connects it to the question
• Quotable, a passage on the page answers the question on its own, in a form a model can lift without repairing it
• Corroborated, the claim is supported elsewhere, so a system that weighs agreement has a reason to trust it

The major providers publish the user agents they use and state that they honour robots.txt for them, and the important detail is that a provider does not have one crawler. It has several with different jobs: one that gathers text for training, one that builds a search index, and one that fetches a page because a user asked about it right now. They carry different names and are controlled separately.

• Blocking the training crawler is a legitimate business decision and does not block the agent that fetches pages to answer questions, so check which one you actually disallowed
• At the time of writing the published names include GPTBot, ClaudeBot and PerplexityBot alongside their search and user-initiated counterparts, and at least one control is a robots.txt token rather than a crawler, so read the current documentation instead of copying a list from an article
• A blanket disallow added years ago for a different reason will keep a site out of assistants entirely, and nothing in any interface tells you
• Server rules count too: a firewall or bot filter that returns a challenge page to an unfamiliar agent produces a fetch that succeeds technically and carries no content

After access comes discovery. Assistants that search reach pages through an index, so the work described in the search specification, being crawlable, having a clean canonical URL and being linked from pages that are already crawled, is the same work. Assistants that fetch a URL directly need only that the URL resolves and renders its content server-side.

A retrieved passage is read out of context, next to passages from other sites, by a model that will paraphrase it. Writing for that is a real constraint, and it is mostly the discipline of never depending on the paragraph above.

Name the subject

Every paragraph states what it is about rather than opening with it, this or the above. A passage whose subject lives two paragraphs earlier is unusable the moment it is separated from them.

Answer, then qualify

The first sentence under a heading answers the heading. Conditions, exceptions and history follow it. A passage that spends four sentences arriving at the answer is often cut before it gets there.

Make claims checkable

Give the figure, the date it applies to and where it came from, in the same sentence. A number with no date and no source is exactly the kind of claim a careful system declines to repeat.

Define terms in place

The first use of any term the audience might not share carries a one sentence definition, because the passage may be the reader's only contact with your vocabulary.

Use question-shaped headings

Headings written as the question a person would ask make the match between passage and query more direct, and they force the section beneath to actually answer something.

Keep structure in the text

Comparisons, steps and criteria belong in prose, lists and headings that survive being converted to plain text. Anything that only exists as layout, or only appears after a script runs, may not reach the model at all.

Say what is uncertain

Marking the limits of a claim makes a passage safer to quote, not weaker. It also stops a system from finding the contradiction somewhere else and choosing the other source.

The third condition is the one you cannot write your way into on your own page. Both paths reward agreement across independent sources: a model trained on text that repeatedly associates your name with a category learns that association, and a retrieval system assembling an answer from several documents has an easier time using a claim that its other sources do not contradict.

• Be described consistently everywhere you appear, because a product described three different ways in three places gives a model no stable fact to learn
• Treat mentions without links as real, since training text and passages carry names whether or not anything is hyperlinked
• Publish the things other people need to cite you for, the definition, the method, the comparison, rather than only the pages that sell
• Keep your own facts current in the places that aggregate them, because a stale description elsewhere outweighs a fresh one on your site when everything else agrees with the stale version

How much any of this weighs is not documented by any provider, and the corroboration effect is inferred from how these systems are built rather than confirmed by them. It is a sound bet, not a measured one, and this spec does not pretend otherwise.

The overlap with search work is large, which is why teams assume the two are the same job. These are the places where that assumption produces wrong decisions.

A single observation therefore proves very little, and a pattern across many is the smallest useful unit of evidence. New pages need retrieval to work for them, and a long-standing reputation is what the trained path uses.

Run these against the page before it ships. Each is a check with an answer, not a judgement call.

There is no position to track. There is no equivalent of a results page you can fetch and score, and any tool that reports a single citation rank is reporting one sample of a non-deterministic system.

What works instead is a panel of your own questions, the ones a buyer would actually ask, run on a schedule across the assistants you care about, with the answers and the cited sources kept. That gives you a record you can compare over time. Alongside it, your server logs and analytics show visits that arrive from assistants that pass a referrer, which is a floor rather than a total, since some sessions arrive with nothing attached.

Anything written about this subject should carry one of these, and this spec is not exempt.

Documented

Crawler names and robots.txt behaviour, the existence of a training cutoff, and each provider's stated policies. Verifiable today and still liable to change.

Structural

That retrieval operates over passages, that models answer from parameters when they do not fetch, and that a source not retrieved cannot be cited. These follow from how the systems are built rather than from any announcement.

Observed

That clear, self-contained, well-sourced passages tend to be quoted more readily. Consistent with practice and with how the systems work, and not confirmed by any provider.

Unverified

Anything expressed as a share, a multiplier or a ranking factor. Nobody outside these companies can measure it, and a number attached to it is a sales device rather than a finding.

The skill returns four things per page