Page audit: why a page isn't cited in AI answers
Every page on your website gets a score between 0 and 100 for how easy it is for an AI engine to pull an answer out of. The score is crossed with what the crawlers actually fetched and what has been cited, so you see fetched, cited and why-not in the same table.
The problem the page analysis solves
When an AI engine answers a question it does not read your whole page. It has a limited text budget per answer and evaluates the content section by section. A page can therefore be both well written and well visited without ever making it into an answer, simply because it is shaped so that nothing can be lifted out of it.
That does not show up in ordinary web analytics. Traffic numbers tell you what happens after someone found you, not why an engine passed you over before that.
How the score is set
The scoring model encodes measured patterns, not opinions. Each part weighs as much as it has been shown to affect whether content actually makes it into an answer.
| Part | Score | What is measured |
|---|---|---|
| Answer block | 25 | The first paragraph under the heading, 30-70 words, answering the page's question directly. |
| Length | 20 | 300 to 1,500 words. Zero points above 3,000, where only a fraction gets through. |
| Section structure | 20 | The largest section between two subheadings stays under 300 words. |
| Tables and lists | 15 | Content in table form is cited 2.5 times more often than the same thing in running text. |
| Freshness | 10 | A machine-readable update date within 30 days. |
| Structured data | 10 | Markup that states what the page is, for example a frequently asked question or an article. |
What the weights are based on
Two published studies lie behind the model:
- An experiment on how answers are assembled showed that the text budget per answer is limited and allocated in order, and that pages over 3,000 words get roughly 13 percent of their content through.
- A review of cited pages in March 2026 showed that the median page was around 941 words, that tables were cited 2.5 times more often than running text, and that 76.4 percent of the top citations had been updated within the last month.
The model is therefore not an opinion about what is nicely written, but a translation of what measurably gets through.
An action list instead of a score
A number between 0 and 100 does not tell you what to do. Every page therefore gets a list of what is missing, sorted by how many points the fix recovers, phrased as an instruction rather than a code.
A measurement that could not be read is left blank instead of being set to zero. The difference matters: a page without an update date is not a page that was updated in year zero, and a guessed value would have sent you after the wrong fix.
What is required
The page analysis needs no tracking code, no API keys and no changes to your web server. It reads your public pages, starting from your sitemap where there is one and otherwise from your most visited pages, and analyses up to 100 pages per round. The feature is included in the AI answers add-on and costs nothing beyond it.
See which pages hold up
The page analysis is included in the AI answers add-on, which can be added to any plan.
Read about AI answers