Four free tools, and what each one measures

The measurement problem came first

For twenty years the question is this page working had one answer: look at the clicks. That answer is quietly expiring. In the first four months of 2026, 68.01% of Google searches in the United States ended without a click, up from 60.45% in 2024 — and the Pew Research Center, tracking 68,879 real queries from 900 consenting adults, found that when an AI summary appeared people clicked through on 8% of searches, against 15% when it did not.

The number worth staring at is smaller than either. In that same study, users clicked a source link inside the AI summary about 1% of the time. A page can be read, quoted and relied on by the thing answering the question, and register as nothing at all in the analytics.

So we built the instruments we needed and left them open. Four of them, no account, no card.

Vectorscope: the passages, not the page

An assistant does not read your page. It reads pieces of it — a retriever splits the document into chunks, embeds each one, and pulls back the handful whose vectors sit closest to the question. Whatever did not survive that split was never in the running, however well it was written.

Vectorscope does that split on a URL you give it and shows you the passages that come out, in the order a retriever would rank them. It is the difference between my page explains the pricing and the paragraph that explains the pricing survives chunking as a self-contained passage and scores against a pricing question.

This is not our theory of retrieval. Anthropic's own write-up of contextual retrieval opens on the same failure: a chunk reading the company's revenue grew by 3% over the previous quarter is useless on its own, because it names neither the company nor the quarter. Prepending context before embedding cut their retrieval failures by 49%, and by 67% with reranking. The lesson for a page is the same one: a passage has to carry its own subject.

Brandscope: what assistants say when you are not in the room

Brandscope asks live assistants about a brand and stores what comes back, with the model name and the timestamp on every answer. Not a score derived from a score — the sentences themselves, and which competitor appeared beside you in them.

The reason it stamps everything is that the answers move. We hold 91,365 stored visibility checks across projects, and the same prompt put to the same model a fortnight apart routinely returns a different set of brands. A single reading is an observation, not a rank. Anybody selling you one number for your AI visibility is selling you the average of a moving thing without telling you it moves.

Voicescope: the question, in the words buyers use

The phrasing you would type is not the phrasing your buyer would speak. Voicescope takes one question and rewrites it the way different people actually ask it — the cautious procurement version, the impatient founder version, the one that names a competitor instead of a category — then shows how the answers differ.

It matters because retrieval is triggered by the wording, not the intent behind it. Two phrasings of one need reach two different sets of passages, and a page tuned for the version you would type can be invisible to the version your buyer speaks.

Patternscope: the pages that were printed, not written

Patternscope reads a set of pages from one site and finds the ones built from a single template — the same skeleton with the city, the file format or the job title swapped in. Programmatic pages are not automatically bad, and the tool does not say they are. What it gives you is the count and the shape: how many of your pages are the same page, and what varies between them.

That is a retrieval question, not a taste question. A thousand pages that differ by one noun produce a thousand nearly identical vectors, and a retriever asked a question in that area has no reason to prefer any one of them.

What they have in common is what they refuse to do

Four tools, one rule: say only what the reading supports.

That sounds like a slogan until you see what it costs. It means no single score. It means a reading is stamped with its moment rather than presented as a standing fact. It means a bounded list says what it left out instead of quietly reporting its cap as its contents. It means that where a signal is not measured the tool says not measured rather than printing a zero — because a zero is a measurement and an absent integration is not one.

And it means saying when a number would be misleading. The run counts of these four tools are trimmed by a nightly retention sweep, so the figure in the database is what is currently kept, not how many people have used it. We could have printed it. It reads like a traffic stat and it is not one, so it is not in this article.

How to read a result without fooling yourself

Take two readings, not one. A single Brandscope run tells you what one model said once. Run it again next week before you conclude anything about direction.

Read the passages before the score. If Vectorscope returns a chunk that does not name your product, no amount of keyword work on that page will fix the retrieval problem — the paragraph needs to carry its own subject.

Treat an absence as an absence. The tools distinguish we measured nothing from we could not measure. Those are different sentences and they call for different work.

Where this goes next

The tools are the public half of an instrument that runs continuously inside the product: the same chunking, the same live assistant calls, the same refusal to report a number that is not a measurement. If you want the vocabulary these readings use, the glossary defines each term once, in five languages, and links the terms that depend on each other.

If you want to see the instrument on work of your own, start with Vectorscope and a page you already believe in. It is the fastest way to find out whether the thing you wrote survives being read by a machine.