- rewrote my headings, moved the answer up, did the whole thing you recommend. how long until I see anything, and how do I know it was me and not the model just changing under me?
- Two different questions living in one, and separating them is most of the answer.
Whether the page now cuts into better passages: you can see that in twenty seconds. Run it again, ask it the same question, and check whether the top passage is the one you wanted. That's not a forecast — it's the same measurement, before and after.
Whether assistants cite you more often: that's a window, not an event. You need enough checks after the change for a rate to mean anything. On a daily cadence that's weeks, not days.
And the part nobody likes: if you changed one page and the rate went up, we cannot prove it was you. The model may have been updated. A competitor may have dropped out. That's why the unit of work is a loop with a re-check on the same prompt set — it's the closest thing to a control available, given you cannot A/B test somebody else's model.
- asked ChatGPT the same question twice, got two different sets of brands. so what exactly are you measuring? this feels like measuring the weather by asking one person how they feel.
- That comparison is better than you probably intended, and it's the right objection.
We're not measuring one answer. One check means nothing, and we never show the result of a single check as a fact about anybody.
What we do instead: the same unit — one prompt, one assistant, one time, one locale — repeats on a schedule, and the screen shows a rate over a stated window of recent checks, with the window next to it. The instability doesn't get smoothed away, it becomes the thing being measured. An assistant that names you in 40% of checks and one that names you in 90% are different assistants, even though each individual run is a coin toss.
That's also why the per-assistant split is always there: assistants differ from each other far more than prompts do, and an average across all of them hides the only part you can act on.
The limit: we don't know why an answer changed. We know that it changed, and we keep both.
- before I paste a client's URL into someone else's tool: what do you keep, who on your team can see it, and are you training anything on it? sorry to be that person but I have to ask.
- You're right to be that person, and the honest answer has an awkward bit in it.
What we keep: the reading itself, briefly, so it can be shared with a link. The fetched page is not retained after the reading is produced. There's no account, so there's nothing to attach it to. The rate limit counts against a salted hash of your address — that's rate limiting, not identity, and it can't be turned back into an address.
What we don't do: train anything on your pages. Ever.
The awkward bit: the embeddings are produced by a third-party provider, and the model is named on the reading. That means the text of the page leaves our server and reaches them. If the URL is behind a login or under embargo, that matters and you should know it before you paste rather than after.
And one more: a shared reading is a link, not a private room. We've taken those out of the index, but anyone with the URL can open it.
- ok genuine question, not trying to start a fight — how is "optimising for AI" not just SEO with a new hat on? every few years there's a new acronym and a new set of consultants. what does your tool find that a normal technical audit doesn't?
- Fair, and mostly you're right about the hats. Here's the one thing that is actually different, and it's narrow.
A normal audit measures your page. A retrieval system never reads your page — it reads a fixed number of passages from it, picked by similarity to a question, and ignores the rest. So the useful question isn't "is this page good", it's "which passages would a retriever receive, and is the one with my answer among them".
We ran six real pages. On the Wikipedia article about transformers — 14,163 words — the retriever received 26 passages holding 47,275 characters and left 53,903 characters unread. That's 53.3% of the page, invisible to that retrieval pass. No audit tells you that, because to an audit a 14,000-word page is "good depth".
Now the part that argues against me: on the other five pages, the number was zero. Not nearly zero — zero. They weren't long enough to hit the ceiling. So no, the sky isn't falling and your content isn't secretly broken. It's one specific measurement that bites rarely and hard.
- there are a few AI visibility trackers now. what do you do differently, other than a different landing page? and please don't say "we're more accurate"
- I won't, because I can't show you their internals and a comparison built from screenshots isn't a comparison.
So here's ours, limited to things you can verify without us:
The tool is open, free, and needs no account. Every formula behind it is stated on the methodology page, and you can run your own page and check each one.
Answers are stored in full. The unit is one prompt, one assistant, one time, one locale, and every figure traces back to the words it was counted from. A rate never appears without the window it was computed over — a rate without its window isn't a rate.
Every figure carries a tag: measured, derived, heuristic or modeled. Revenue impact stays modeled until your analytics are connected, and it does not become measured by getting older.
And one thing against us: access is invite-only, so you can't try the product today. You can try Vectorscope.
- congrats on shipping. classic problem though: this looks interesting and I can't sign up. why invite-only? from out here it reads like manufactured scarcity.
- It isn't scarcity, it's a bottleneck in one specific place, and I'd rather name it than be mysterious about it.
The bottleneck is the corpus. To compare you against a competitor we have to vectorise both sites, and our crawl currently follows links — which collects marketing pages, while the competition often happens in the generated long tail. We measured that gap rather than guessing at it: of the page pairs that actually split a search query, 0.4% to 1.8% have both halves vectorised.
Opening self-serve today would mean handing people a comparison report standing on about one per cent of the data. It would look convincing. That's the problem with it.
So what's open is what works completely: Vectorscope, this board, and the methodology page. When the crawl is seeded from demand instead of from links, self-serve becomes honest — and then it opens.
- where does 26 come from? looks like a number someone picked on a friday afternoon. openai and anthropic don't publish a passage limit, so what is it actually based on
- It's ours, and you're right that nobody publishes theirs.
26 is the ceiling our pipeline applies, and Vectorscope shows it so the number isn't anonymous. We're not claiming it's what Google does — we don't know what Google does, and neither does anyone outside Google.
The part worth taking from it isn't the 26. It's that a ceiling like it exists in every retrieval system and none of them publishes theirs. So the finding from the test above isn't "26 is the magic number" — it's that over half of a 14,000-word page went unread under one ceiling, and it will go unread under yours too.
If you're building retrieval rather than content, the fastest check: run a page you already index and compare the passage boundaries with your own chunker's. If your extraction ratio is lower on the same URL, one of the two extractors is discarding body text rather than boilerplate — and it isn't ours, because we print the ratio.
