Not every page that reads well to a person reads well to a model trying to lift one clean, quotable passage from it. Crawl Cove's ai-extractability check scores every indexable page against five structural signals and flags the ones that fall short.
What this finding means
Each page earns points for five signals, weighted out of 100: answer-first (25, an H1 followed by a usable answer, not buried under a long unstructured wall of text), answer-blocks (25, content chunked into question-shaped H2 or H3 sections, or FAQ schema), a comparison table (15), FAQPage or QAPage schema (20), and a visible date plus author (15). A page scoring under 50 gets one finding naming exactly which signals it is missing. Pages that returned anything other than a 200 or carry a noindex directive are skipped.
A real finding looks like this:
AI-extractability score 30/100 (below 50). Missing: no answer-first opening (clear H1 + concise lead answer); no comparison table in structured data (schema.org Table); no author and date in structured data (Article schema).
The last three signals are read from your page's JSON-LD, not from the rendered page. A real HTML <table>, a genuine byline in the page's design, or FAQ content written into the body all score zero for their matching signal if that same information is not also expressed as schema.org Table, FAQPage/QAPage, or Article markup carrying author and datePublished. This was measured directly: on a September 2026 sweep of 312 real homepages, 4.7% of flagged pages actually contained a real HTML table their owners never marked up as one, and 11.9% showed a visible author and date that was never in the page's structured data.
Why it matters
AI answer engines quote passages, not whole pages, so content that cannot be lifted out as a self-contained chunk tends to get summarised from a competitor's page instead, even when your page is the better source. The two heaviest-weighted signals, answer-first and answer-blocks, cost nothing to add and do not require any structured data at all: they are a writing and heading-structure choice. The three schema-based signals reward information that many pages already show visually but never expose in a form a model can parse without guessing.
How to fix it
- Open with a direct answer. Put a clear H1 at the top, and have the page state its core answer promptly rather than building up to it. If the page runs long, sub-headings count as the structure that satisfies this signal too.
- Chunk the body into question-shaped sections, or add FAQPage JSON-LD. Either alone satisfies answer-blocks; both together are stronger for coverage but not required.
- If the page has a genuine comparison table, mark it up as schema.org
Tablein your structured data, in addition to the visible HTML table. The visible table alone will not register. - If the page shows a publish date and an author, add both to an Article-family JSON-LD block (
datePublishedordateModified, plusauthor), not only in the visible byline. - Re-crawl the page to confirm the score clears 50 once the changes ship, and check the finding's evidence for which specific signals still need work if it does not.
False positives and edge cases
- A real table or byline that exists only in the HTML still counts as missing. This is a real limit of the check, not a bug: it reads structured data because that is the only render-stable place a crawl can check these signals reliably, so visible-but-unmarked content is genuinely invisible to it.
- A short or transactional page is not what this is built to flag usefully. The check still scores every indexable page, but a login screen or a cart page is not content an AI answer engine would try to quote regardless of its score, so a low score there is much lower-priority than the same score on an explainer, comparison or guide page.
- This is a structural heuristic, not a content-quality judgement. A high score means a page is well-shaped for extraction; it says nothing about whether the answer inside it is actually correct or useful.
Related reading
For the related check that looks specifically at whether a page has any explicit question-and-answer structure at all, one of the two signals that make up half of this score, see No Q&A Structure for AI Answer Engines. For the two dedicated checks behind the visible-date-and-author signal, see No Author Identity in Article Schema and No dateModified in Article Schema.
Google documentation: AI features and your website
Frequently asked questions
- Is this an SEO ranking score?
- No. It is a structural heuristic for how easily an AI answer engine can lift a self-contained quote from the page, not a ranking factor, and Google has not said it uses anything like it. Treat it as GEO (generative engine optimisation) work, separate from technical SEO.
- Why do I need structured data if the table or byline is already visible on the page?
- Because three of the five signals, a comparison table, an FAQ block and a visible date plus author, are read from the page's JSON-LD only, not from the rendered HTML. A real <table> element or a printed byline that never made it into structured data scores zero for that signal even though a visitor can see it.
- What is the minimum I need to do to pass?
- Answer-first (25 points) and answer-blocks (25 points) alone reach the 50-point threshold with no structured data at all: an H1 followed promptly by a direct answer, plus content chunked into question-shaped H2 or H3 sections. The schema-based signals are reinforcing, not required.
- Does every page need to score 100?
- No. 50 is the flagging threshold, not a target. A transactional or navigation-only page is not the kind of content an AI answer engine would try to quote in the first place, so a low score there matters far less than on an explainer or comparison page you want cited.