The CITE framework: how to write a page an AI will quote
Four tests any page has to pass before a language model can cite it. A working checklist you can run against your own commercial pages this afternoon.
Most advice about writing for AI search stops at “write helpful content”, which is true and useless. It does not tell you what to change on the page in front of you.
CITE is the checklist we run against a commercial page before we consider it finished. Four tests, in order. A page that fails an earlier test cannot be rescued by the later ones.
| Test | The question | |
|---|---|---|
| C | Crawlable | Can the system fetch and read this at all? |
| I | Inspectable | Can it find the answer without rendering or guessing? |
| T | Transferable | Can a single sentence be lifted out and still be true? |
| E | Evidenced | Is there a reason to trust the sentence it just lifted? |
C: Crawlable
Everything else is academic if the page cannot be fetched.
Check these four things, in this order:
- AI crawler access.
GPTBot,OAI-SearchBot,ClaudeBot,PerplexityBotandGoogle-Extendedare governed separately fromGooglebotinrobots.txt. A site can rank perfectly in Google search while being invisible to every system that writes answers. Blocking them is a legitimate publisher decision; blocking them by accident is the most common unforced error in this category. - Rendering without JavaScript. Disable JS and reload. If your service description, pricing or contact details disappear, many AI crawlers see the same nothing.
- Status codes. The page returns 200, not a redirect chain ending in one.
- Canonical honesty. The canonical points at the page a visitor actually lands on.
Fails this test if: you have never checked your robots.txt against those
user agents.
I: Inspectable
The system can fetch the page. Now it has to locate the answer inside it.
An extraction system reads structure before it reads prose. Help it:
- The question is a heading. Not “Our Approach” but “How long does an AI visibility audit take?” People phrase questions to assistants the way they speak. Your headings should match that, not your internal taxonomy.
- The answer is directly under the heading. First paragraph, not the fourth. Marketing pages are usually built to persuade, which means they build towards a conclusion. Extraction reads the top of the section and moves on.
- Comparisons are tables. A model answering “X vs Y” will take a table over six paragraphs every time.
- Processes are ordered lists. They survive summarisation. Prose describing five steps frequently does not.
- One claim per paragraph. A paragraph carrying three claims is hard to quote without quoting all three.
Fails this test if: someone skim-reading your page could not find the answer in under ten seconds. A model is a worse skim-reader than a person.
T: Transferable
This is the test most pages fail, and the one nobody runs.
Pick three sentences at random from your page. For each, ask: could this appear verbatim, on its own, as the answer to a question somebody would actually type?
Compare:
We’re passionate about delivering world-class results for our clients.
Nothing there can be quoted in response to any real question. It carries no fact. Now:
Engagements run month to month with 30 days’ notice, and the audit is delivered within ten business days.
Every clause is liftable and attributable. A model answering “do marketing agencies lock you into contracts?” has something concrete to use.
The failure mode is not bad writing. It is writing that is about the company rather than stating things about the company. Brand atmosphere is fine in moderation. A page made entirely of it cannot be cited.
The quick audit: count the sentences on your key page containing a number, a timeframe, a named thing, or a specific commitment. If it is under a fifth, the page is atmosphere.
Fails this test if: your page would read identically on a competitor’s site with the logo swapped.
E: Evidenced
The model found a quotable sentence. Now: why should it repeat yours rather than a larger competitor’s saying the same thing?
This is where most of the durable advantage sits, and it is the slowest part.
- Attribution. Content credited to a real, identifiable person is stronger evidence than content from nobody. Inventing an author persona is a credibility risk, not a tactic.
- Dates. Published and updated dates, accurate. Undated content is harder to trust and recency is a plausible retrieval input.
- Primary sources. Where you cite a fact, link where it came from. Where the fact is yours, say how you measured it.
- Corroboration. A synthesised answer weighs what several sources say. If only your own domain makes a claim about you, it is weak evidence. Being described accurately by credible independent sources does more than any amount of self-description.
- Originality. Material that exists nowhere else has to be cited to be used. A summary of what five other articles said has no reason to be referenced by anything.
Fails this test if: every claim on the page is self-reported and undated.
Running it
Take your five most commercially important pages. Score each out of four. Fix in order, lowest letter first, because the tests are sequential: a beautifully evidenced page that fails C is invisible, and a crawlable page that fails T has nothing worth taking.
In our experience the distribution is lopsided. Most sites pass C and I with small technical fixes, and fail T and E comprehensively. That is encouraging, because T is a rewriting problem you control entirely.
What CITE does not do
It does not guarantee a citation. Nothing does. These systems do not publish their ranking factors, they change frequently, and the same prompt returns different sources to different users.
What it does is remove the reasons a system cannot use your page. That is a smaller claim than most of this category makes, and it is one we can actually stand behind.
CITE is one of the frameworks behind our generative engine optimization work, and the full audit method is published free in the Academy.