A visualization that shows how a brand's content clusters in the high-dimensional space used by AI search models.
Marketers who are concerned with aligning their messaging with topics returned by AI search models.
01What it does, step by step
Five steps, and the tool names each one as it finishes so a failure is attributable. It fetches the URL and reports the status code it got. It extracts the readable text, discarding navigation, boilerplate and markup, and reports what fraction of the document survived. It chunks that text along the page's own H2 and H3 structure rather than every N characters, so a passage arrives with the heading that names it. It embeds every passage as a vector. Then it groups them by cosine similarity, at the page's own median, so the groups describe this page rather than a fixed threshold.
The chunker, the embedding model and the passage cap are the ones the paid product runs in production. A demo on a simplified pipeline would flatter the pipeline it exists to demonstrate, and the tool's single most useful number — how much of a long page never reaches an index — only exists because the real cap is applied.
It takes one address and shows you the parts of that page a machine will actually read, and the parts it will skip.
02What the five figures mean
The strip is the whole reading in five numbers, and each one fails in a different way.
Fetched is a status code and a host. Extracted is the ratio of readable text to document: a long page with a low ratio shipped mostly navigation, and is a short page as far as any retrieval system is concerned. Chunked is how many passages the page became and how many characters they hold. Embedded is the vector count and their width. Grouped is how many groups the passages fell into, and how many are repeats.
One enormous group means the page says one thing many times — thin to a retriever, however thorough it reads to an editor. Several groups of one passage each is the opposite problem: a page that changes subject constantly and gives a retriever nothing complete to quote.
03How to use it on a page of your own
- Run a page you already rank for, and read the extraction ratio first. If it is low, the page is discarding body text rather than boilerplate, and no amount of writing fixes that.
- Look for the passage your answer lives in. If it sits beyond the cap, the fix is moving it up or splitting the page, not adding words.
- Ask the page a question you know the answer to, and check whether the top passage is the one that contains it. That is the closest thing to watching a model decide what your page is for.
- Change one heading and run it again. Heading text does retrieval work, and this is the fastest way to see how much.
04Common mistakes
- Reading the group count as a score. It is a description of one page, not a mark out of ten, and two groups is right for a page about two things.
- Treating a low extraction ratio as a writing problem. It is usually a rendering problem: text that exists only after JavaScript runs was never in the document.
- Assuming the passages a retriever receives are the passages a reader sees. A sidebar a reader ignores is body text to an extractor.
- Expecting the reading to say whether an assistant cites you. It reads one page in isolation; citation is a measurement over time on prompts you choose.
05What it cannot tell you
It cannot say whether any particular assistant has your page in its index — nobody outside those companies can see that. It cannot rank you against a competitor, which needs a corpus of both rather than one URL. It cannot tell you whether you are cited in an answer today, which is a measurement over weeks on a prompt set, not a property of one page. And it makes no judgement about the writing: it reports the decisions a pipeline made, which is a different thing from whether the page is any good.
06A worked example
A 4,000-word pricing guide reads as authoritative and returns an extraction ratio of 0.31, twelve passages and one group. The ratio says two thirds of the document was chrome. The single group says the twelve passages are near-identical — the page restates its promise twelve times. A retrieval system asked about pricing receives one of those passages and has no reason to prefer this page over any other. The fix is not more words; it is twelve passages that each answer a different question.
Frequently asked questions
How is this different from a normal SEO audit?
An SEO audit checks whether a page can be crawled and ranked. This checks whether it can be quoted: what a retrieval pass extracts, how it splits, and which passages fall outside the cap a model actually reads. The two overlap on the technical foundation and diverge on structure — an audit rewards a page that satisfies a query, this rewards a passage that survives being lifted out of the page.
Is the pipeline the same one the paid product uses?
Yes — the same chunker, the same embedding model and the same passage cap. That matters for one figure in particular: how much of a long page never reaches an index only exists as a number because the real cap is applied, and a simplified demo would report a smaller loss than the truth.
Why does my page split into so few passages?
Usually because it has few headings. The chunker cuts on the page's own H2 and H3 structure, so a long page written as one continuous argument becomes a handful of very large passages — each of which a retriever must take or leave whole. Adding real headings is the cheapest way to change this.
Does a high group count mean the page is good?
No, and there is no good number. Two groups is right for a page about two things and wrong for a page about one. What is diagnostic is the extreme: one enormous group means the page repeats itself, and a group per passage means it never settles on a subject long enough to be quotable on any of them.
What does it cost, and is there a catch?
Nothing, with an allowance per visitor so one person cannot spend the day's budget — a rate limit rather than a paywall, and no account. The limit, such as it is, is that it reads one page at a time and says nothing about whether an assistant answers with you.
Asked out loud
spoken, not typedThe same term in the words somebody uses speaking to an assistant rather than typing into a box — written from the situation, which is why each one carries the situation it came from.
Read the page's extraction ratio first. Most of the time the answer is that the guide's body text is a minority of the document — navigation, related-post blocks and markup make up the rest — so a page that reads as four thousand words arrives as eight hundred. The tool reports the ratio, and no amount of rewriting fixes a rendering problem.
Run the page before and after and keep both readings. The passage count, the extraction ratio and the group structure all change visibly when headings change, and that is a before-and-after somebody can look at — it is evidence about the page's legibility, not about traffic, and it is worth saying which.
Yes. There is no account, no install and no card, and one reading takes about twenty seconds. Paste the address and read the five figures at the top; the detail below them is for when you have a desk.