How To Audit A Page For AI Citability In Fifteen Minutes

Citability is not a score. It is a property you can check by hand, and the check is faster than installing anything.

The underlying question is simple: if a retrieval system lifted any single paragraph from this page and used it to answer a question, would the result be true, attributable, and recognisably yours? Most pages fail on the first of those, and the failure is always in the same places.

The Checklist

Work down it with the page open. Twelve items, roughly a minute each.

1. Does each H2 correspond to a question a person would type? “System Requirements” is a label. “How much RAM do you need to run a local model?” is a query. Both are fine as headings; only one matches retrieval.

2. Is the first paragraph under each heading a complete answer? If it is context-setting, the retrieval system extracts your context-setting.

3. Does any claim depend on a qualifier in a different paragraph? This is the highest-frequency failure. Move the qualifier into the claim’s sentence.

4. Are the numbers specific? “Lightweight” invites a model to substitute its own guess. “1.1 GB on disk” does not.

5. Is every figure attributed in the same sentence as the figure? Not in a footnote, not two sentences later.

6. Are perishable claims dated in the body text? Write “as of August 2026” in the prose. The page-level date does not travel with the chunk.

7. Are entities named in full at least once per section? A chunk that begins mid-document has no idea what “it” refers to.

8. Would any paragraph be false if read alone? Read three at random with the rest covered.

9. Is there anything on the page that only you could have written? Your own figures, your own tests, your own incident. Without this the page is a paraphrase of the corpus and there is no reason to cite it over the corpus.

10. Does the page state a position? Hedged pages get summarised into nothing.

11. Is the page one topic? Pages covering three subjects get retrieved for one and quoted for all three.

12. Can a person still read it? Self-contained sentences repeat context. Overdone, they read like a legal document, and a page nobody finishes is not a page anybody links to.

The Fix That Handles Half The List

Items 3, 4, 5, and 7 are one edit repeated.

Before:

Local processing offers significant benefits for teams working with confidential material. Performance is generally acceptable on modern hardware. It should be noted that older machines may struggle, as certain instruction set extensions are required.

After:

Local processing keeps draft text on the device, but it requires a CPU with AVX2 support and about 2 GB of free memory for a 1.7B model. Machines from before roughly 2013 generally cannot run it at all.

Everything that constrains the claim now travels with the claim. That single habit is the whole of the technique described in how to write so AI search quotes you correctly, and it is the one item on this list worth doing even if you skip the other eleven.

The Test That Beats The Checklist

Paste a section into an assistant. Ask it the question the heading asks. Read the answer as though you were a stranger who had never seen the page.

If it comes back a fair representation, the structure works. If it comes back confident and subtly wrong, the specific wrongness tells you which qualifier got separated from which claim. That diagnostic is more precise than any audit tool, because it is the actual failure mode reproduced rather than a proxy for it.

Do this for your five most commercially important pages and you will find more than a full-site crawl would surface.

A Context For The Rewrite Pass

A Wrivio Context for citability could say:

Rewrite this so each paragraph is true when read alone. Move every qualifier, condition, and exception into the same sentence as the claim it limits. Keep the source attribution in the same sentence as any figure it supports. Replace pronouns referring to earlier paragraphs with the full name. Make the first sentence after each heading a complete direct answer. Keep every number, name, and date exactly as written, and do not add facts, figures, or sources that are not in the original.

Press Ctrl+Shift+Space, paste one section at a time rather than the whole page, and check the diff. The instruction about not adding figures is the one that matters, because the most damaging possible outcome of a citability pass is a confident invented number that then gets quoted accurately.

What Not To Add

An FAQ block on every page whether or not anyone asks those questions. Schema markup on content types it does not apply to. A statistics line every 150 words because a tool said density matters. Keyword variants inserted into sentences that were already clear.

Each of these makes the page worse for humans and does not make it more retrievable. The research on what actually moves visibility does not support any of them, and Google’s own guidance on optimizing for AI features in Search states that there is no AI-specific markup to add.

Common Questions

How long should a citable page be?

Long enough to carry its facts and no longer. Density beats length, because retrieval lifts small chunks and padding is what gets lifted from a padded page.

Should every page have an FAQ section?

Only where readers genuinely ask distinct follow-up questions. A generic FAQ block appended to hit a structured data target adds noise to the chunks around it.

Does this hurt readability?

Mildly, if taken to extremes, because self-contained sentences repeat context a human already has. Apply it to the claims that matter and the page still reads normally.

How often should I re-audit?

When the underlying facts change, and once a year otherwise. The failure that actually costs you is a dated claim still stated in the present tense two years on.

Download Wrivio for Windows to run the citability rewrite over unpublished pages locally, so drafts never leave your machine.