AI SEO: How to Get Your Pages Quoted in AI Answers

A page sits in the top handful of results for a question its author knows cold, and the traffic keeps thinning. Above it, an assembled answer resolves the question outright, citing two other sites. Neither of those two outranks the page. They were simply easier to quote.

That gap is what AI SEO is actually about, and it is not a new discipline bolted onto the old one. It is the same technical and editorial work judged against a different unit. Classic search ranks pages. An assembled answer selects passages. A page can be authoritative, well linked and correctly indexed, and still contain no sentence that survives being lifted out of it.

So the working definition we use is blunt: write pages that stay true and legible one paragraph at a time. Everything below is that idea applied to structure, markup, crawler access and authority, plus an honest account of why you cannot yet measure whether any of it worked.

Being ranked and being cited are different jobs

Ranking is a competition between whole documents for a position in a list. The reader clicks, arrives, and finds whatever context they need sitting around the answer. Citation works the other way round. A passage is pulled out, dropped into an answer written by something else, and attributed. The reader may never arrive at all.

What retrieval is doing to your page

Broadly, these systems retrieve candidate material, compose an answer from it, and attach sources to the parts that came from somewhere specific. Every step of that favours content usable with the least reinterpretation. A sentence that defines the thing plainly needs no rewriting. A paragraph beginning "as we saw earlier" cannot be used without carrying the earlier part along with it, so it tends not to be used.

Which is why advice circulating under the AI SEO banner sounds suspiciously like advice from fifteen years ago about writing clearly and marking up properly. It is that advice. The difference is that the penalty for ignoring it used to be a slightly worse snippet, and now it is not being quoted at all.

Ranking gets your page into the room. Being quoted requires a sentence somebody else can lift without editing it first.

What AI SEO is actually optimising for

The property that decides this is extractability: whether a machine can take one part of your page, present it on its own, and have it stay accurate and comprehensible. Extractability is not a score anyone publishes. It is an editorial habit, and it shows up in four places.

Lead with the definition

Under any heading that poses a question, the first sentence answers it, subject first, no run-up. "A redirect map is the file that records where every old URL should send its visitors after a migration." Not "when planning a migration, one of the things worth thinking about early is how URLs will be handled." The second version is friendlier and completely unusable, because there is nothing in it a machine can hand to a reader as an answer.

This carries a real editorial cost, and it is worth naming. Front-loading definitions makes prose flatter and occasionally repetitive, because you cannot rely on the reader having read the section above. We think the trade is worth taking on pages whose job is to answer something. On a brand or campaign page, whose job is persuasion, it is not.

Write sections that stand alone

Assume any single section will be the only part of the page anyone sees. That rules out "as mentioned above", pronouns pointing back at the previous paragraph, and the habit of naming a concept once at the top and calling it "the framework" for the next two thousand words. Name it again. Three words, and the section becomes portable.

Use the shapes that lift cleanly

Some formats survive extraction almost intact, which is why the pages that get quoted tend to look alike. When we plan organic content work, these are the units a page gets built from:

• A one-sentence definition. The single most extractable object on the internet. Put one under every heading that names a concept.
• A comparison table with a real header row. Two or three options against the criteria that separate them. Tables travel well because the relationships are explicit rather than implied by sentence order.
• A numbered procedure. Imperative steps, one action each, and no step that depends on a decision made three paragraphs earlier.
• A question-shaped heading with the answer directly beneath it. The heading supplies the query, the paragraph supplies the response, and the pairing is unambiguous.
• Short paragraphs carrying one claim each. A paragraph making three arguments at once cannot be quoted for any of them.

Your HTML is doing real work again

For a decade it was possible to be sloppy about markup and get away with it, because search engines had become good at inferring intent from a mess. Extraction is less forgiving. A list that is really a stack of line breaks is not a list. A comparison rendered as an image is invisible. A table built from styled divs looks identical to a reader and carries none of the structure that made it worth quoting.

The bigger risk is content that is not in the response at all until a script assembles it. Rendering is expensive, and not every agent that fetches your page will run your JavaScript. This is the same failure that has always sat underneath indexing problems, which is why it pays to be precise about where crawl, render and index actually diverge before blaming the writing. If the words are not in the first response, treat them as optional.

There is a content-operations angle here that is easy to miss. Pages assembled from structured fields produce predictable markup: a heading is a heading everywhere, a table is always a table, and nobody pastes styled HTML out of a word processor. A workspace like Acrosite takes that shape, generating files from structured content and committing them, which removes a whole class of markup drift that no amount of technical SEO can keep fixing page by page.

Who is allowed to fetch your pages

None of this matters if the agent cannot get the page. Several distinct crawlers now operate under separate names: the classic search crawler, crawlers gathering material for model training, and crawlers fetching a page live because a user has just asked something. They can be allowed and refused independently in robots.txt, and they are not interchangeable. Blocking the training crawler while allowing live retrieval is a coherent position. Blocking everything and hoping to be quoted is not.

Robots rules are also not the only gate. Bot-management policies at the edge, aggressive rate limits and firewall rules refuse agents silently, without appearing in any file anyone thought to check. When a stakeholder asks why competitors keep getting cited, read the access logs before rewriting a single paragraph. Sites that are invisible purely by configuration are more common than sites with a content problem.

Our position, for most marketing sites: allow retrieval. The counter-argument is real, though. Being quoted well can mean being answered without a visit, and if your content is the product rather than an advertisement for it, refusing access is a defensible commercial decision. Make it deliberately, write down which agents and why, and revisit it, rather than inheriting it from a security default nobody remembers setting.

Structured data: useful, and oversold

Structured data helps a machine agree with you about what a page is. It names the organisation behind an article, marks a question-and-answer block as exactly that, and separates entities that share a name. Genuine value, and cheap to add.

What it does not do is confer authority. Markup describes a page, it does not improve one, and marking up an answer nobody can see on the page is both against the published guidance and a good way to look untrustworthy. The rule we hold to is simple: mark up only what a visitor can actually read, and never emit a question-and-answer block for questions that exist purely in the markup.

This is also where audits generate the most noise, because missing optional properties are easy to detect and easy to list. A page of schema warnings is not a plan. Our parent company covers that pattern directly in how to read an audit you were sold.

Why authority still decides who gets quoted

Extractability makes a page usable. It does not make it the chosen one. When several pages answer the same question equally clearly, something has to break the tie, and the tiebreaker is roughly what it has always been: corroboration. Whether the surrounding site treats the subject seriously. Whether the same claim shows up in places already trusted. Whether the organisation behind the page is identifiable and consistent wherever it appears.

That is why a single excellent page on an otherwise thin site rarely gets picked, and why the cluster remains the working unit rather than the article. Coverage is what makes an answer credible: the definition, the comparison, the edge cases and the objections, all present and linked to each other. Building a cluster search can actually read is the same work as building one an assistant can trust. The related question of how many pages a site needs before search takes it seriously resolves the same way here: enough that the coverage looks deliberate rather than opportunistic.

The part nobody wants to admit: you cannot measure this yet

There is no reliable report telling you how often your pages were quoted. Referrers from assistants are partial and inconsistent. Some answers produce no visit at all, which means the most successful possible outcome leaves no trace in your analytics. Anyone selling a precise citation-share figure is selling a sample presented as a census.

That is not a reason to skip measurement. It is a reason to measure what does move and to label clearly which of it is a proxy. It is also the same miscrediting problem that has always distorted channel reporting, described in why your best performing channel is probably being miscredited. What we would put on the report instead:

• A manual citation check, run monthly. Ask the same set of questions your pages answer, record who gets cited, keep the log. Tedious, unglamorous, and the only direct evidence available.
• Branded search and direct arrivals. People who read an answer mentioning you often arrive later, by name. The shape of that curve says more than any single number.
• Referrals from assistant domains. Track them knowing they undercount. A rising floor still tells you something even when the absolute figure is wrong.
• Self-reported source on enquiry forms. One open question at the point of enquiry catches discovery paths no analytics tool can see.
• Position for question-shaped queries. Being retrieved usually requires being findable first, so ordinary visibility for the exact questions stays a leading indicator.

Where we would start on a real site

Sequence matters more than effort here, because most of the early wins are unblocking rather than authoring. In order:

1. Confirm access and delivery. Check robots rules, edge bot policies and rate limits, then fetch a page as a plain client and confirm the answer is present in the HTML.
2. Pick the ten questions worth money. Not the ten with the highest volume. The ten a buyer asks before choosing, which are usually more specific and less contested.
3. Rewrite those pages for extraction. Definition first under each heading, sections that stand alone, one comparison table, one procedure where a procedure genuinely exists.
4. Add markup for what is now visible. Article and organisation identity, plus a question-and-answer block only where those questions appear on the page itself.
5. Fill in the cluster around each one. The adjacent pages that corroborate the answer, linked to each other in prose rather than only from the site menu.

If a report has already landed on your desk listing forty AI SEO deficiencies, treat it the way you would treat any finding list and separate what is severe from what is urgent, a distinction we make in severity or priority after an audit. Most of that list will be markup properties, and almost none of it will be the reason nobody is quoting you.

The reassuring part is how little of this is speculative. Clear definitions, honest structure, markup that matches the page and enough coverage to be credible were good practice before any of these systems existed, and they will outlast whichever one is currently ahead. If you want a second opinion on a specific set of pages, an SEO audit is the usual starting point, or you can simply send us the questions you want to own and we will tell you what is standing in the way.