We pulled the AI Performance report in Bing Webmaster Tools expecting a boring number and got something more interesting. Over a 90-day window, 1,703 AI citations for autoseo.it.com across 36 active days. That is our own Bing Webmaster Tools data, for our own domain.
Then the Pages tab: cited pages was 1. Not 1 on average. 1 on every single day in the window. The entire citation count came from one URL.
That is the finding worth writing about, and it is not the flattering one. Volume of citations and breadth of citation are different metrics, and only the second tells you whether your content strategy is working. A domain with 1,703 citations from one page has one good answer, not a body of work that assistants trust. This is what we learned taking it apart.
What the shape actually means
The first instinct is to read a high citation count as a health signal. It is not, on its own. Split it:
- Citations measure how often assistants reached for your domain.
- Cited pages measure how many distinct answers you have that are worth reaching for.
A high first number with a low second one is unambiguous: the retrieval layer already trusts the domain — it went back repeatedly — but it found exactly one page it considered a usable answer. That is not an authority problem. Authority problems look like zero citations. This is a quotability problem, and it is fixable at the page level.
The uncomfortable corollary is that most of the corpus contributed nothing. Every post we had shipped in that period, and the assistant kept returning to one URL. Publishing more of the same would not have changed the number.
Why assistants concentrate on a single URL
We went looking for what made the cited page different. Three things, in rough order of importance.
It answered the question completely, in one place
The cited page resolved its question without requiring anything around it. Definition, the direct answer, the caveats, the procedure — all present, in sequence, in the same document.
Retrieval works on chunks. A model assembling an answer wants a passage it can lift with the context still attached. A page where the answer is distributed across four sections, a comparison table and a link to a second post is harder to use than one that states the whole thing in two paragraphs under a clear heading. It is not that the distributed version is worse writing. It is that it is worse raw material.
Once a model has found one page that resolves a topic cleanly, there is no incentive to reach for a second source that resolves it partially.
Attribution resolved to something real
The pages that get cited tend to be the ones where the assistant can say who is making the claim. Not a byline string, but a byline that resolves — a named author, linked to a page about that author, with credentials that are checkable, and an organisation the page identifies consistently.
When attribution is thin, a model can still use the content, but it has a weaker basis for presenting it as a source. Given two pages with similar content and only one with resolvable attribution, the choice is not close.
The entity was unambiguous
Our cited page named the product, the category and the problem in consistent terms, matching how the rest of the site and our external profiles describe them. Pages that used looser internal shorthand did not get picked up.
Entity clarity is boring work — the same product name, the same category words, the same one-line description everywhere — and it appears to matter a lot for whether a machine can connect a page to the thing a user asked about.
What we changed to widen it
These are the changes we made. We are deliberately not attaching outcome numbers to them, because the honest position is that we made them recently and a single domain over one window is not evidence of anything generalisable. What follows is reasoning, not a case study.
Question-shaped headings
We rewrote section headings from topic labels to the questions people actually ask, then made sure the sentence immediately underneath answers that question outright. "Pricing considerations" became "How much does this cost per month". The paragraph beneath answers it in the first sentence.
This costs nothing and it directly changes the chunk boundary a retriever sees. A heading that is a question, followed by an answer, is a self-contained unit.
Self-contained answer blocks
We adopted a rule: the first two sentences under any question heading must stand alone if pasted into an answer with no other context. No "as discussed above", no "this" referring back three paragraphs, no pronoun whose antecedent is in a different section.
It reads slightly more repetitive to a human reading top to bottom. It reads dramatically better to a machine reading one section.
Visible bylines that resolve to a real author page
A name in the metadata is not attribution. We made bylines visible in the rendered page, linked to an author page that states who the person is and what they actually do, and kept the author identity consistent across posts. The point is that a machine following the byline arrives somewhere that substantiates it.
If you have no real author to name, do not invent one. An honestly attributed team byline that resolves to a real about page beats a fabricated persona, and fabricated personas are a reputational liability the moment anyone checks.
Real freshness signals
We made sure every article exposes a machine-readable publication and modification date — a time element carrying a datetime attribute in the markup, matching the date shown to readers, matching the dates in structured data.
The critical part is that the three agree. A page showing "Updated last week" with structured data claiming a date from two years ago is worse than no date at all, because the disagreement is itself a quality signal. And a modified date that moves without the content changing is a claim you cannot support.
FAQPage schema with visible-text parity
We added FAQPage structured data derived from the questions and answers that are actually visible on the page — same wording, same answers, no schema-only content.
Parity is not optional here. Structured data that describes content a visitor cannot see is a guidelines violation and it is trivially detectable. Derive the schema from the rendered text, never the other way around.
One page per question, not one page per topic
The structural change with the longest lead time. A single sprawling post covering eight related questions gives a retriever one target and a lot of noise. Eight focused pages, each resolving one question completely and linking to the others, gives it eight candidates.
This is not a licence to produce thin pages. Each one still has to be genuinely complete on its own question, which usually means the total word count goes up, not down.