A page is extractable when one of its passages still makes sense after it is lifted away from the page. That property does not make the page rank, and it does not make an answer engine cite it.
I built the checker on this site to keep those claims separate. It tests nine structural properties, names the source behind each one, and labels the difference between a documented requirement and an editorial heuristic.
Extractability belongs to the passage, not the ranking system
An extractable passage names its subject, makes one claim, and carries the support needed to understand that claim. A reader can paste it into a blank document without repairing pronouns or hunting for the evidence above it.
That is an editorial property. Google's systems still decide whether to crawl the page, index it, rank it for a query, and show a supporting link.
Google's guidance for AI features says there are no additional technical requirements for AI Overviews or AI Mode beyond ordinary Search eligibility. Meeting the requirements does not guarantee crawling, indexing, or serving.
The AI Overviews checker therefore reports what its checks found rather than predicting placement. A clean result means the tested passages avoid the structural problems the tool knows how to detect.
Snippet eligibility is the one documented gate
Google states that a supporting page in AI Overviews or AI Mode must be indexed and eligible to appear in Search with a snippet. That makes noindex, nosnippet, and a restrictive max-snippet directive relevant controls.
The checker can inspect those directives when it receives HTML and the response's X-Robots-Tag header. Plain pasted prose cannot establish snippet eligibility, so the check reports that it was skipped.
Only the snippet check restates a platform requirement. The remaining eight are heuristics for whether a passage is clear enough to lift and verify.
Nine checks separate access, structure, and support
The checks answer different questions, so the result is a count rather than a weighted visibility score.
| Check | What it asks | Basis |
|---|---|---|
| Snippet eligibility | Can Google show the page with a snippet? | Google's documented technical requirement |
| Direct answer first | Does the opening answer before it introduces the topic? | NN/g's inverted-pyramid guidance and Google's preference for clear organization |
| Claim-bearing headings | Do headings state a useful point instead of naming a topic? | NN/g's layer-cake scanning research |
| Self-contained sections | Does each section name its subject before using pronouns? | Passage-level extraction heuristic |
| Liftable definition | Can a definitional sentence stand alone? | Observed answer shape, reported as a heuristic |
| Dated evidence | Can a reader tell when a claim or page was checked? | Freshness and trust heuristic |
| Sources beside quantities | Does a measured claim carry attribution where it appears? | Google's guidance on original work and clear sourcing |
| Units beside numbers | Does each number say what it counts? | Extractability heuristic for quantified claims |
| Stated limits | Does the page say where its answer stops applying? | Trust and qualification heuristic |
The table is not a recipe for AI citations. It is a review order for finding passages that lose their meaning when removed from their original layout.
The opening should answer before it explains
NN/g's inverted-pyramid guidance puts the conclusion before background. That helps a hurried reader, and it gives an extractor a useful first passage instead of a paragraph explaining that the topic matters.
The heuristic cannot tell whether the answer is correct. It only detects common preamble patterns and checks whether the opening asserts enough to be an answer.
An opening can pass while making a false claim. Accuracy still comes from sources, direct observation, or a reproducible artifact.
Headings should carry the page when the body disappears
NN/g calls reading headings alone the layer-cake scanning pattern. Its eyetracking summary describes that pattern as more effective than the F-pattern readers fall into when useful subheadings are absent.
A heading such as "Limitations" tells the reader where they are but not what the section concludes. "Robots.txt cannot prove that access works" carries the useful boundary even when the paragraph below it is hidden.
The checker uses a text heuristic to flag short topic labels. It cannot judge whether a technically assertive heading is true.
A section should survive being quoted alone
Section openings that begin with "this" or "it" often depend on a previous paragraph. Naming the subject again costs a few words but keeps the passage intact when a reader lands on an anchor or an engine retrieves only that section.
Definitions need the same independence. A useful definition says what the term is before it adds history or comparison.
The checker looks for that sentence shape near the top of a page. It will miss good definitions written in another form, so the finding remains a prompt for review rather than a verdict.
Evidence should travel with the claim
A date elsewhere on the page does not date a changing claim once the claim is quoted alone. Put the date in the same passage when a product rule, audit, or measurement can drift.
Quantities need the same local support. The sentence should say what the number counts and name its source or method close enough that the attribution survives extraction.
Google's people-first content guidance asks whether a page provides original information and whether its sourcing creates trust. The checker can spot missing units or attribution words, but it cannot verify the underlying source.
A stated limit makes the answer more useful
A passage that claims universal reach is easy to quote and hard to trust. A useful answer names the exception that would change the decision.
The tool looks for language that marks a limitation or trade-off. That test is deliberately broad because a regex cannot decide whether the stated boundary is the important one.
The article on AI crawlers and robots.txt provides a concrete example: an allowed crawler in robots.txt can still be denied by a WAF. The limitation changes what evidence an auditor must inspect next.
Schema can describe a page but cannot earn an AI citation
Google says no special schema.org markup is required for AI Overviews or AI Mode. Existing structured data can still support eligible rich results when it matches the visible page.
The checker reports JSON-LD as information only. It does not count schema as a pass, a failure, or evidence that the page is more likely to be cited.
llms.txt does not make Google use the page
Google also says sites do not need new machine-readable files or AI text files to appear in its generative Search features. The audit of four public llms.txt files found useful indexing patterns, but validation could not show that an assistant fetched or acted on any file.
An llms.txt file may still serve a tool that chooses to request it. That utility is separate from ranking and citation.
Crawler access and extractability are separate tests
A clear passage cannot be cited by a system that cannot fetch it. The AI crawler access checker reads robots.txt by agent and separates search access from training access.
The reverse is also true. An allowed crawler can reach a page whose opening buries the answer or whose evidence loses context when quoted.
Run the access test first, then review the page itself. Neither result predicts placement, but together they remove two failures a publisher can actually control.
Use the findings as an editing queue, not a score
Fix a failed snippet directive before rewriting headings because it is a documented eligibility problem. After that, use the heuristic findings to locate passages that need a clearer subject or closer support.
Stop when the answer is accurate and self-contained. Passing every text pattern is not the reader's task, and a page written to satisfy a checker can become less useful than the draft it replaced.