Answer passage extractor
Find the passages on a page that best match the extractable shape we publish, and see exactly what disqualifies the rest.
One page in. We read the HTML your server returns, split it into runs of text under each heading, and rank them against the passage shape we publish. Free, no sign-up, the whole result is shown.
The shape we look for: 134 to 167 words, self-contained, definition in the opening.
That band is our own editorial guideline, set in our internal GEO analysis of attensira.com and stated in full further down this page. Our homepage is held to it by a test that fails the build when the definition drifts out of range. It is not a measurement of ChatGPT, Claude or Perplexity. Nobody publishes what length of text they lift, or whether they lift contiguous text at all.
For every candidate you will get the word count, whether the first sentence is definition-shaped and where it lands, the heading it sits under, and each phrase or pronoun that makes it lean on the paragraphs around it.
What this tool measures, and what it refuses to claim
It fetches the URL once, strips out scripts, styles and comments, drops anything inside nav, header, footer, aside and form, and splits what is left into runs of text under each heading. A candidate passage is a run of consecutive blocks under one heading, starting at a block of at least twelve words and extended until it reaches 134 words, runs out of section, or has absorbed ten blocks. Candidates overlap on purpose: every block that could open a passage gets one.
For each candidate it reports four things it measured:
- Word count, and how far that is from the 134–167 band.
- Whether the opening sentence is definition-shaped — a named subject followed early by one of a short fixed list of copulas and definition verbs (is, are, was, were, means, refers to, is defined as, stands for, describes, denotes, is short for), with something substantial after it. A sentence opening on it, this or there is not counted, because it defers to something the passage has not named.
- Where the definition lands, as a word offset from the start of the passage.
- Which phrases make it depend on its surroundings — a bare pronoun opening, a phrase like as mentioned above or listed below, or a third-person pronoun in the opening sentence arriving before anything in that sentence is named. Each finding quotes the text it matched.
What it does not claim: that any assistant lifts passages of this length. The band is our editorial guideline, stated in full on this page and enforced against our own homepage by a test in our build. No answer engine publishes the length of text it quotes, whether it quotes contiguous text at all, or how it chooses. A passage ranked first here has not been cited and is not predicted to be cited. It matches a shape we think is worth writing to, and that is the whole claim.
The self-containment findings are a phrase and pronoun heuristic over the passage's own words, not a grammatical parse. A pronoun we do not flag may still have no antecedent inside the passage. Treat a clean result as "we found nothing", not as "there is nothing".
Where the 134 to 167 word band comes from
From section 6 of our own GEO analysis of attensira.com, which sets an optimally extractable passage at roughly 134–167 words, self-contained, with the definition in the first 40–60 words. That document is an internal working analysis and we do not publish it; the band it sets is what we are publishing, here, in full. Our homepage's What is Attensira? block is 161 words, and a test in our build fails if it drifts out of that window.
We state it because a threshold nobody can inspect is worth nothing, and we run this tool against other people's pages with the same number we hold ourselves to. That is the honest version of a threshold: a stated editorial choice, applied consistently, with its origin named. It is not a measurement, and a tool that presented it as one — "ChatGPT prefers 150-word passages" — would be inventing a fact about a system whose retrieval behaviour is not documented anywhere.
If you disagree with the band, the underlying advice survives without it: write a block that answers the question on its own, put the answer first, and do not make the reader scroll up to find out what it refers to.
How to fix the passages this finds
- Nothing on the page reaches the band because every block is short. Two-sentence paragraphs read well and quote badly: each one is a fragment that needs the ones around it. Merge the fragments that answer one question into a single block that answers it.
- The best candidate opens on a pronoun. Replace It is the practice of… with Generative engine optimization is the practice of…. This is the cheapest fix on the list and usually the one with the largest effect, because it turns a fragment into something that stands alone.
- The definition is there but it arrives at word 90. Move the sentence that defines the subject to the front of the block and let the context follow it. The guideline's first-60-words rule exists so a passage still answers the question when it is truncated.
- A candidate carries as mentioned above or listed below. Those phrases are fine in a document read top to bottom and useless in a passage read alone. Say the thing again in four words, or drop the phrase.
- Every candidate is preamble with no heading. A passage that sits under a question-shaped heading gains the question as context. Fix the heading structure first with the content structure analyzer, then run this again.
- The tool reports almost no text and you know the page is full of it. The words arrive from JavaScript. Nothing on this list matters until they are in the HTML your server returns.
Questions people ask about this check
Does a top-ranked passage mean ChatGPT will quote my page?
No, and we will not imply otherwise. This ranks passages against a shape we publish and follow. Whether an assistant ever fetches your page, whether it judges you the best available source, and what it does with the text once it has it are three separate questions, none of which this tool can see. We do not intercept anyone's traffic to ChatGPT or any other assistant, and nothing here is a record of what a model retrieved.
Where does 134 to 167 words actually come from?
From section 6 of our own GEO analysis of attensira.com, an internal working document that sets it as the editorial target for extractable passages. The document is not published; the band is, on this page, and our homepage is tested against it. It is a chosen reference point that we apply to our own writing, not a measured property of any model. We say that on the result, in the payload the API returns and here, because a number that travels without its provenance eventually gets quoted as a fact.
Why do the candidates overlap? The same paragraph appears twice.
Because every block long enough to open a passage gets its own candidate, and a candidate then absorbs the blocks after it until it reaches the band. A section of four paragraphs therefore produces up to four candidates that share text. That is deliberate: which sentence a passage starts on changes whether it opens with a definition, and picking one starting point for you would hide the option you actually wanted.
It flagged a pronoun that clearly has an antecedent. Is that wrong?
Possibly. The check is a heuristic: it flags a third-person pronoun in the opening sentence that arrives before anything in that sentence is named. It cannot resolve reference the way a reader does, so it will sometimes flag a pronoun that is perfectly clear, and it will certainly miss pronouns that are not. Every finding quotes the text it matched so you can overrule it in a second.
Does it read pages that render with JavaScript?
No. It reads the HTML your server returns and does not execute scripts, which is the same position most documented AI crawlers are in. If a page's content arrives from JavaScript, this tool reports very little text — and that thin result is the finding, not a failure of the check. The crawler URL inspector shows the text-to-script split behind it.
Do you store the URLs I check?
Not against you. There is no account, no email field and no history you could be looked up in. But the reading is held for about an hour in a shared cache keyed by the URL itself, so an identical URL in that window is answered from the cache rather than fetching the page again — which is what stops us re-crawling a stranger's site once per visitor. The URL also appears in ordinary server logs, and your address is counted against a per-tool rate limit.
Where to go next
- Rewrite an opening so it answers the question firstThis tool finds the passage closest to the shape. Content optimization starts from an answer that cited someone else, rewrites the page that should have answered it, and opens the edit as a pull request against the repo the page lives in.
- See which of your pages assistants actually citeA passage ranked well here has not been cited. Citation tracking is where you find out whether it ever is.
- Check the heading structure a passage sits underA passage under a question-shaped heading carries the question with it. This shows whether the headings are there at all.
- Estimate how much of the page a retriever would readA perfect passage below the point where a retrieval window cuts is a passage nothing reads. This shows where the cut lands.
- Confirm AI crawlers are allowed to fetch this pageEverything on this page assumes the words can be retrieved at all. Robots.txt is the gate in front of that.