Free, no accountExperimental
A free allowance per visitor, no account needed.
Awaiting a URL
Five numbers decide whether a page is legible to a retrieval system. Here is what each one means when it is bad.
A retrieval system reads a fixed number of passages from a page and ignores the rest, so anything below that cut is invisible no matter how good it is. Vectorscope applies the same cap the product applies and names the passages that fell outside it.
If your answer, your price, your differentiator or your FAQ sits below the cut, the fix is not more words. It is moving that passage up, or splitting the page so it has a passage budget of its own.
The extraction ratio is how much of the page turned into readable text — a low one means most of what you shipped was navigation, boilerplate or markup that no model will ever see. A long page with a low ratio is a short page as far as retrieval is concerned.
Common causes, in the order we see them: text rendered only by JavaScript, an article wrapped in so much chrome that the body is a minority of the document, and images carrying words that exist nowhere in the HTML.
Groups are passages the embedding thinks are about the same thing, and one enormous group usually means the page says one thing many times. That reads as thin to a retriever even when it reads as thorough to an editor.
Several small groups with one passage each is the opposite problem: a page that changes subject constantly gives a retriever nothing complete to quote.
Asking the page a question returns the passages that answer it, ranked, using only that page — which is the closest thing to watching a model decide what your page is for. If the returned passage is not the one you would have chosen, the page is not organised the way you think it is.
This is also the fastest way to test a heading rewrite: change the heading, run the page again, ask the same question.
Semantic chunking is splitting a page into passages along the lines where its meaning changes, instead of every fixed number of characters. It is the step between reading a page and embedding it, and it decides what a model can quote.
Fixed-length chunking cuts every N characters and is the default in most pipelines because it is trivial to implement. It also cuts through the middle of a sentence, separates a heading from the answer under it, and produces passages that begin with half a word.
It is fast, cheap and repeatable, and it is why so many pages retrieve badly for questions they answer perfectly.
Heading-aware chunking uses the page's own H2 and H3 structure as the cut lines, so a passage arrives with the heading that names it. This is what Vectorscope does first, and it is the single largest improvement available to most pages.
It also means your heading text is doing retrieval work. A section called "Overview" contributes nothing to the passage; a section called "How much does it cost?" contributes the question itself.
Semantic chunking compares neighbouring blocks and cuts where the meaning shifts, which produces passages that stand alone but costs an embedding pass over the page before any retrieval happens. It is the reason a chunker can be slow and the reason the results are worth it.
Nothing here is free of judgement: a cut is a decision about where one idea ends. Vectorscope shows you the decisions rather than a score, because a score would hide them.
This page answers that question for one URL at a time, honestly, which means naming what it cannot see as clearly as what it can.
If you build retrieval rather than content, this page is a reference implementation you can point at any URL in about twenty seconds.
The chunker, the embedding model and the cap here are the ones the product runs in production. A demo using a simplified pipeline would flatter the pipeline it exists to demonstrate.
Four terms, defined once, because a page that uses them without defining them is a page an assistant cannot explain to anybody.
Nobody outside Google can state the mechanism, and this page will not pretend to. What is observable is that cited pages tend to contain a passage that answers the query completely on its own, and that pages whose answer is spread across several sections are cited less often than their content deserves. Vectorscope shows you whether your page has such a passage.
Give the answer its own passage, under a heading that names the question, high enough on the page to survive the retrieval cap. That is not a trick — it is the shape retrieval rewards, and you can check whether your page has it by running it here.
Yes, for one page at a time. It embeds every passage of a URL and groups them by meaning, so you can see which parts of the page are near each other and which stand alone. It does not plot a corpus of thousands of documents.
Heading-aware chunking cuts at the page's own headings; semantic chunking cuts where the meaning changes, whether or not there is a heading there. This tool does both, in that order, which is what a production pipeline should do.
The reading is kept briefly so it can be shared with a link, and the fetched page itself is not retained after the reading is produced. There is no account, so there is nothing to attach it to.
Yes, with a rate limit rather than a paywall. The limit exists because every reading costs an embedding call, and it resets on a fixed window that the page tells you about before you hit it.
Longer answers to the questions this page can only summarise.
One page on its own. Whether that page answers the questions people actually ask an assistant, how it stands against a rival page on the same subject, and which topics your site has no page for at all — those need a corpus of the whole site, measured over time.
Ask for an invite