How to structure a page so an AI system can quote it accurately.

A system summarising a topic has to do something specific with your page: find a passage that answers the question on its own, lift it out of everything around it, and present it without adding a claim you did not make. A page that reads well to a person can fail all three, because prose written for a reader leans on the sentences before it.

That is the difference worth naming early. Ranking is a competition for position. Being quoted accurately is a question of whether any part of the page survives being removed from the page. Most content is never tested against that, because nobody reads it the way a retrieval system does — in fragments, without the run-up.

The practical consequence is that the work is closer to editing than to optimisation. Nothing here involves tricking a system. It involves writing pages where the answer is actually present, in one place, stated once, and true on its own.

A usable test: take any heading on the page, read only the two sentences beneath it, and ask whether they answer the heading. On most pages they do not — they introduce the answer, which arrives three paragraphs later.

These stack. A page that fails an earlier condition cannot be rescued by satisfying a later one, which is why so much effort spent on the last two produces nothing — it was spent on a page that never cleared the first.

A machine can use a page when it is:

• CRAWLABLE

  Reachable, permitted, and rendered as text without depending on a script running. This is the only purely technical condition, and it is binary: a page that fails it is invisible regardless of quality.

• UNDERSTANDABLE

  The subject, the audience and the scope are stated rather than implied. A page that never says plainly what it is about forces a system to guess, and a guess becomes the description of you.

• RETRIEVABLE

  Some specific passage answers a specific question completely, without the surrounding paragraphs. This is where most otherwise good pages fail.

• ATTRIBUTABLE

  It is clear who is asserting the claim and on what basis. Unsourced confidence reads as filler to a person and as a risk to a system that has to stand behind what it repeats.

• WORTH CITING

  The page contains something that is not available in the ten pages already saying the same thing. Form cannot manufacture this, and no amount of structure substitutes for it.

The single highest-value change on most pages is moving the answer to the top of its own section. Not a summary of the section, and not a promise that the answer is coming — the answer, stated flatly, in one or two sentences directly under the heading that asks for it.

This is uncomfortable to write, because it reverses the order most people compose in: context, reasoning, conclusion. Retrieval reverses it back. The conclusion is what gets lifted, and the reasoning is what convinces the reader who stays.

The same sentence also has to work with the heading removed. A passage that begins “It depends on three things” is meaningless in isolation; “Custom software is justified when the process is a genuine differentiator” is not.

A heading that reads “Pricing” tells a system the region of the page. A heading that reads “What changes the price of a custom application” tells it which question the region answers, and lets it match a phrasing a person actually used.

That is not a licence to write a page of question-shaped headings nobody asks. Google’s guidance on structured data is explicit that markup must describe visible content, and its content guidance discourages material produced primarily for search systems rather than for people. A heading invented to carry a phrase, above a paragraph that does not answer it, fails both tests and reads badly besides.

The workable version is narrow: where a section genuinely answers a question people ask, phrase the heading as that question or as its direct answer. Where it does not, leave it as a topic and let the section earn its place another way.

Form is useful where it removes ambiguity that prose cannot. A comparison across a fixed set of factors is genuinely clearer as a table, because the table asserts that the same factor was considered for both options. A definition set out as a definition is unambiguous in a way an in-sentence gloss is not.

Where form is applied to content that was already clear, it adds nothing and costs reading. A list of five sentences that were a paragraph is a paragraph with bullets in front of it, and the reformatting has not made the underlying point more available to anybody.

The test is whether the structure carries information. A table whose second column repeats its first in different words is not structure; it is layout.

Worth the space:

• A COMPARISON ACROSS FIXED FACTORS

  Two or three options answered on the same axes, so a difference is visible rather than argued.

• A DEFINITION STATED AS ONE

  A term the page depends on, defined in a sentence that is true without the paragraph around it.

• AN ORDERED SEQUENCE

  Steps where the order is load-bearing, with what each one decides — not a list of activities.

• A BOUNDED SET

  Four failure modes, three conditions, five questions: a list whose completeness is the point.

Structured data describes what is already on the page in a form a machine does not have to infer. That is the whole job. It is useful for facts with a settled shape — what an organisation is, who wrote something, when it was published, what an article is about — and it is checkable against the visible page.

What it does not do is add credibility the page has not earned, and Google’s general structured data guidelines are direct about it: markup has to represent visible content and must not describe things the page does not contain. Ratings, credentials or claims asserted only in markup are a misrepresentation before they are an optimisation error.

The honest summary is that structured data removes ambiguity about facts. It does not create authority, it does not improve a weak page, and a page carrying markup it cannot support is worse off than one carrying none.

This is where most writing on the subject stops being reliable. The field is full of specific-sounding assertions about what AI systems reward, stated without a source and frequently with a multiplier attached. Treat all of it as inference until somebody shows the documentation.

What is documented is narrower and more useful. Google states that to be eligible for its AI features a page has to be indexed, eligible to appear with a snippet, meet the technical requirements and comply with search policies — and that there are no additional requirements or special optimisations beyond ordinary search practice. It also states plainly that meeting every requirement does not mean a page will be crawled, indexed or served.

So the achievable goal is eligibility and clarity, and the unachievable one is selection. Nothing in this article causes a page to be cited by any named system, and anybody claiming otherwise is describing something they do not control.

1. 01

   Documented behaviour

   Owner

   The platform, in published guidance

   Advances when

   Can be stated as fact, with the source beside it

2. 02

   Published standard

   Owner

   A standards body or specification

   Advances when

   Can be stated as fact about the standard, not about any system’s use of it

3. 03

   Reasonable practice

   Owner

   The practitioner

   Advances when

   Can be recommended, with the reasoning shown

   Ends when

   Stops being defensible the moment it is presented as a ranking factor

4. 04

   Inference

   Owner

   Whoever is asserting it

   Advances when

   Can be offered as a view, labelled as one

   Ends when

   Becomes a fabrication when a number is attached to it

This is best learned on one page rather than planned across a site. Pick the page that answers a question the business is genuinely asked, and work through it in this order — the order matters, because three of these steps are wasted if the first one fails.

1. 01

   Confirm it can be read at all

   Decides

   Does the page render its content as text without a script running?

   Output

   A page whose content exists in the response, or a defect to fix before anything else.

2. 02

   State the subject in the first paragraph

   Decides

   Would a stranger reading only the opening know what this is and who it is for?

   Output

   An opening that defines rather than introduces.

3. 03

   Turn each heading into the question it answers

   Decides

   Which question does this section actually settle?

   Output

   Headings that match how the question is asked, or a section that has no question and should be cut.

4. 04

   Move each answer directly under its heading

   Decides

   Can the first two sentences stand alone?

   Output

   One liftable passage per section.

5. 05

   Remove the dependencies

   Decides

   Does any passage rely on a pronoun or a scope set earlier?

   Output

   Passages that survive being read out of order.

6. 06

   Attribute or soften every claim

   Decides

   What is this resting on — documentation, standard, practice or opinion?

   Output

   Claims that are sourced, or reduced to what is genuinely known.

Each of these is common enough to be worth naming, and each makes the page worse for the person reading it as well as for the system reading it — which is usually the clue that the technique was wrong.

• QUESTIONS NOBODY ASKS

  Headings written to carry a phrase, above content that does not answer them. Detectable by reading the heading and the paragraph together.

• THE SAME ANSWER IN FIVE PLACES

  Repeating a passage across pages so that whichever one is retrieved carries it. This splits the signal and leaves no page that is the obvious source.

• MARKUP AHEAD OF CONTENT

  Structured data describing ratings, credentials or events the page does not contain.

• FRESHNESS THEATRE

  Changing a date without changing the content. It is trivially detectable, and it costs the credibility the page had.

• STRUCTURE OVER SUBSTANCE

  A perfectly formatted page with nothing in it that a competitor could not have written. Form is a delivery mechanism; it is not the thing being delivered.

1. Being retrieved and being quoted accurately are different problems: the second depends on whether any passage survives being lifted out of the page.
2. Five conditions stack — crawlable, understandable, retrievable, attributable, worth citing — and effort spent on the last two is wasted if the first fails.
3. The highest-value change on most pages is moving the answer to the top of its own section and making it true without the sentences around it.
4. Tables, definitions, sequences and bounded sets earn their space when they carry information prose cannot; reformatting clear prose adds nothing.
5. Structured data describes what is visibly on the page. It removes ambiguity about facts; it does not create authority and must not assert what the page lacks.
6. Google documents that AI-feature eligibility needs indexing, snippet eligibility, technical compliance and policy compliance — with no special optimisations — and states that serving is not guaranteed.
7. Separate documented behaviour, published standards, reasonable practice and inference. A number attached to an inference is a fabrication.