Overview
A model citing your page has to lift a span of text out of it and drop that span into an answer where none of your surrounding context exists.
A sentence that needs the previous sentence to make sense cannot be lifted, so it will not be cited. Most well-written prose fails this, because good prose deliberately chains sentences together. Ranking rewards flow. Citation rewards sentences that stand alone.
This is a structural edit, not a keyword edit.
The quote test
Take any sentence carrying a claim. Paste it alone into a blank document.
Does it still assert something complete and true, with no ambiguous referent?
If not, it is unciteable, no matter how good the page is. Run this on every claim-bearing sentence before and after editing.
The passes
1. Resolve the referents. Sentences opening with this, that, it, these, they, such are the largest single cause of failure. Restate the subject even when it feels repetitive on the page. Repetition costs a reader half a second and buys the whole citation.
2. Answer first. The paragraph under a heading opens with the answer in one sentence, then the context. Models weight the first sentence after a heading disproportionately, and a paragraph that builds to its point gets its throat-clearing quoted instead.
3. Headings as real questions. Match how someone would ask out loud, not your internal taxonomy. “How much does it cost to self-host n8n?” beats “Pricing considerations.”
4. One claim per sentence. Compound sentences get truncated mid-quote, and a half-quote either misrepresents you or gets dropped. Split them.
5. Attach specifics. Numbers, dates, versions, named methods, named alternatives. A sentence with a specific gets quoted; a general sentence gets paraphrased, and paraphrase does not carry a citation.
6. Date the perishable claims. Prices, version numbers, model capabilities, anything hedged with “currently”. Each one carries its date in the sentence itself, so a model can tell whether the claim is still live. Stale and uncited beats stale and cited wrong.
7. Confirm it’s reachable. None of the above matters if the content isn’t in
the server-rendered HTML. Check: view source and find the actual sentences, check
robots.txt for AI crawler blocks, and check for a Content-Signal header
declining AI input. JS-gated content is invisible to most of them.
Output contract
Return, in this order:
- Blockers. Anything from pass 7. If the page is unreachable, say so first and stop; the rest is wasted work until it’s fixed.
- The claim table. Every claim-bearing sentence, its quote-test result, and which pass it failed.
- The rewrite. The restructured page.
- The five most quotable sentences in the result, listed plainly. These are what you’d expect to see turn up in an answer. If none of them feels like something worth quoting, the problem is that the page has no claims, and no amount of restructuring fixes that.
Common mistakes
| Mistake | Why it fails |
|---|---|
| Treating this as keyword optimization | Models aren’t matching keywords, they’re extracting spans. Structure is the lever. |
| Adding an FAQ block and calling it done | An FAQ of questions nobody asks, answered vaguely, is unciteable too. Same passes apply inside it. |
| Removing all flow to maximize extractability | Humans still read the page and convert on it. Self-contained sentences and readable prose are compatible. It costs some repetition, not readability. |
| Optimizing a page with no original claim | Citation requires something worth citing: your data, your method, your number. A summary of consensus gets no citation because the model already knows it. |
| Skipping pass 7 | The most common real cause of zero citations is that nothing can fetch the page. Check it first, always. |