Methodology
How each number on this site is produced
One operating system and an open learning surface. The product measures AI visibility over time; free tools make individual readings checkable; the glossary defines the shared language; and skills turn a signal into a next move. What follows states what each surface measures and what its figures are allowed to claim.
Counted from something stored. A measured figure can be traced back to the rows it was counted from.
Arithmetic on measured figures, where one input is a choice somebody made — a competitor set, a window.
A rule that reads text and is right most of the time. Directional, and never presented as a fact.
An estimate from assumptions you can see. It stays modeled until it is measured, and does not become measured by ageing.
Every figure carries one of these four wherever it appears. The tag is part of the number, not a footnote about it.
Vectorscope: what an assistant reads on a page
Vectorscope runs one URL through the pipeline a retrieval system runs, and reports what each step produced. It is the same chunker, the same embedding model and the same passage ceiling the product uses — a demo on a simplified pipeline would flatter the pipeline it exists to demonstrate.
1 · Fetch
One request, as a browser makes it. A noindex directive is reported rather than ignored, because a page asking to be left out of an index is the answer to the question being asked.
2 · Extract
Readable text only: navigation, scripts and boilerplate are discarded, and the share of the document that survived is reported as a ratio.
3 · Chunk
Cut at the page's own H2 and H3 first, then where the meaning shifts. Capped at 26 passages, which is the ceiling the product applies.
4 · Embed
Every passage becomes a vector of 1,024 dimensions. The model is named on the reading, so the number is not anonymous.
5 · Group
Passages are grouped by cosine similarity above a floor computed from the page itself — see the floor below, which is the one judgement in the pipeline.
The five numbers, stated rather than described
- passages received
- = min(chunks produced, 26)
- text never read
- = characters in chunks beyond the ceiling ÷ characters extracted
- extraction ratio
- = characters extracted ÷ characters in the fetched document
- self-similarity
- = mean cosine similarity across every pair of the page's own passages
- group cohesion
- = mean cosine similarity within one group
All five are measured. Nothing here is estimated and nothing is compared against another site: Vectorscope reads one page and says only what that page did.
The grouping floor, because it is the one judgement
Two passages join a group when their similarity is above a floor, and the floor is not a constant. It is computed from the distribution of the page's own pairs, because cosine similarity on a template-heavy page sits far higher than on a varied one — a fixed floor would report every form page as a single idea. The floor used is printed with the reading, so the judgement is visible rather than buried.
What Vectorscope cannot tell you
- Whether any specific assistant has your page in its index. Nobody outside those companies can see that.
- How your page ranks against a competitor's. That needs a corpus of both, which is the product rather than the demo.
- Whether you were cited in an answer. That is a measurement over time on prompts you choose, not a property of one page.
Voicescope: the questions a keyword is hiding
Somebody typing a keyword into a search box has stripped their own context out — they know the box wants two words. Speaking to an assistant, the context comes back. Voicescope puts it back: it crosses one keyword with a circumstance and writes the sentences that result. Every question it returns is a hypothesis, and it is tagged Modeled for that reason.
The four circumstances a keyword is crossed with
1 · Urgency
Needed right now, or on the move, or against a deadline. The same need at three speeds is three different sentences.
2 · Who is asking, and on what
A phone, nothing installed, hands busy, somebody standing over them. The device and the company decide how much a person is willing to say out loud.
3 · The thing in front of them
The specific document, page, account, report or person. A question about a thing in your hand does not read like a question about a category.
4 · What actually hurts
The mistake already made, the thing they cannot find, the thing they are afraid of getting wrong. This is the axis that produces the questions nobody thinks to write down.
What one run is, in numbers
- scenarios per run
- = exactly 4
- questions per run
- = 12, aimed for
- questions returned
- = printed beside the number asked for
- runs per address
- = 10 in 6 hours
- an answer is replayed for
- = 72 hours
Both counts are printed because a run that returned nine of twelve has to say nine: a page that states its target and shows a result reads as complete when it is not. A scenario is a job somebody is doing, never a cluster of similar wordings — two scenarios that could share one answer are one scenario.
What Voicescope cannot say
- How many people ask any of these. There is no volume here, and a model asked for one would supply a number it made up.
- That anybody has asked them at all. They are written from a circumstance, not read out of a log.
- Which of them you are found for, or cited in. That is a measurement over time against prompts you chose, and it is the product rather than the tool.
The glossary
A dictionary of the words this product is built out of. Search, retrieval, AI visibility, trust — each term gets one page that says what it means in plain language, and carries the sources it was written from.
It is on this page because a methodology is only as clear as its vocabulary. When a figure on a screen is tagged Derived, or a report says a passage was never retrieved, the words behind those sentences are defined once and in one place rather than explained a little differently every time they appear.
Read the glossaryAI visibility: whether assistants name you
Inside the product, the same buyer prompts are asked of the assistants a workspace has chosen, on a schedule, and every answer is stored in full as evidence. A check is one prompt, asked of one assistant, at one time, in one locale — and every figure below is counted from that unit.
The check
Prompt, assistant, locale, timestamp and the answer in full. Kept, so a number can always be traced back to the words behind it.
The reading
Whether the brand was named, whether competitors were, and which sources were cited. Read from the stored answer, not from a summary of it.
The window
Rates are computed over a stated window of recent checks, and the window is shown beside the rate. A rate without its window is not a rate.
What each figure is allowed to claim
| figure | tag | what it counts |
|---|---|---|
| mention rate | Measured | Checks naming the brand ÷ checks in the window. |
| per-assistant rate | Measured | The same, split by assistant. Assistants differ from each other more than prompts do. |
| citations | Measured | Sources an answer named, resolved to a domain. |
| coverage | Measured | Prompts with at least one check in the window ÷ prompts being watched. |
| share against rivals | Derived | Arithmetic on measured mentions across a competitor set you chose. Derived, because the set is a choice. |
| sentiment near a mention | Heuristic | A rule reading the words around the brand. Directional, and labelled as such. |
| revenue impact | Modeled | A model over your own funnel figures. It stays modeled until analytics are connected, and does not become measured by getting older. |
What this page does not publish
- Which prompts a workspace watches, or anybody's prompt set. They belong to the customer.
- The thresholds and the model routing behind a check. Published, they become a target.
- Anything about one customer's numbers. Aggregates only, and only where an aggregate cannot be reversed.
The product and the open surface
The workspace turns recurring checks into owned actions and measures the result over time. The open tools below expose useful parts of that work without an account; they are not shortened product demos and do not make claims the reading cannot support.
Vectorscope
Read one URL as a retrieval system reads it.
Open tool 02Voicescope
Turn a keyword into the questions a person may actually ask.
Open tool 03BrandScope
Compare the language assistants use for a brand.
Open tool 04PatternScope
Find where content starts to read like a template.
Open tool 05ReviewScope
Inspect the review signals a brand publishes about itself.
Open tool ALL TOOLSChoose the right starting point
Start with a page, a keyword, a brand or a review signal.
Explore free toolsHow to hold us to this
Every screen in the product links here, and every figure on this page names the step that produced it. If a number on a screen has no line on this page, treat the number as unfinished and tell us — that is a bug in the same way a wrong total is.
Read it as data
/api/trust/methodology returns the trust layer as JSON, and /llms.txt describes the open tool for an assistant rather than for a browser.
Check the open half yourself
Vectorscope, Voicescope and the glossary are open. Everything this page says about them can be checked without an account.