VercatusMethodology
How Vercatus measures, and where it stops.
One number, Answer Share, and one test of a change, the verdict. Both are designed so a buyer can check them, and both say what they cannot tell you.
Judged against
topics left alone.
Every play you ship gets a verdict within 45 days. The topics and engines a play aimed at are compared with topics no play touched, from the day the engines could see the change: when their crawler fetched the changed page, or when the page they read first named you.
- Day 14First readToo early, or a clear early result
- Day 30Second readMost verdicts settle here
- Day 45Final readThe verdict is frozen
States
- WonAnswer Share on the play's topics rose clearly more than on the topics left alone.
- Lost groundIt fell clearly more than on the topics left alone.
- No effectThe answers support no move bigger than the smallest one the data could see.
- InconclusiveThe window closed without a clear answer; the verdict says how many more answers would decide it.
- Too earlyNot enough answers yet since the engines could see the change.
Grades
- AOne play, nothing else changed, the controls flat, plenty of answers.
- BClear, but bundled with another play, one confounder, or fewer answers.
- CThe direction only. Never called a win.
- DBefore and after, with no usable control. Never called a win.
A or B
Built-in controls: engines that never saw the change, and answers given without search, should stay flat. When they move, the grade drops. Only grades A and B may say a play caused a gain.
A modelled range, never a count of prompts.
No engine publishes how often it is asked. Vercatus estimates AI conversations a month per topic and market from sourced signals, and calibrates the estimate against first-party AI data from connected sites, such as Search Console's AI impressions and assistant fetches.
- Always a rangewith a grade and the sources behind it
- Not estimatedwhen too little evidence stands behind a topic
- Calibrated across sitesin aggregate, so it sharpens as more sites connect their data
What each Why lens
counts.
Three sources of an answer, and what engines say about you.
- Memory
- What each engine names with search off, from a small share of the poll kept out of Answer Share.
- Sources
- Every page the engines cite, fetched politely with our signed crawler, read and kept in versions with the brands it names; a brand's own site never counts as a source for it.
- Search
- The searches engines expose while answering, looked up in the top ten of the index that engine uses; a results page about something else counts as not checked.
- Engine Briefs
- What an engine says when asked about you, matched to the evidence; never a reason on its own.
SEO Health,
scored by rule.
Every result is a measurement: a status, a tag, a length, a timing. Each rule cites the search engine's own documentation, a problem on a more important page costs more, and every point lost belongs to a named rule. How the score is made
| Question | Weight |
|---|---|
| Can search engines find and index it? | 35% |
| How does it look in search results? | 25% |
| Is the content well built? | 20% |
| Is it fast and safe for visitors? | 20% |
What the numbers
cannot say.
Knowing them is what lets a Won verdict mean something.
- Correlation in the lenses. Why shows what goes with being named; only a verdict says a change caused a move.
- Small effects. Many verdicts read No effect or Too early, and say how many more answers would decide them.
- Bundles. Plays shipped together share one verdict, which says so.
- Engines change. A window with an engine change is graded down.
- Demand is modelled. A range with a grade, never a count of prompts.
- SEO Health is a sample. It reads the pages that carry the site and says which; it is not rank tracking.
One measurement,
and what it can say.
The home page pictures show "cheapest way to send money abroad", asked of 8 engines on 5 October 2026: real answers and the pages and searches behind them, not a customer's data.
What counts as named
A brand counts once per answer that names it as an option. A model reads the names and they are checked against the text; publishers and finder sites are dropped.
Who answered
ChatGPT (gpt-4.1-mini-2025-04-14), Google AI Overviews (google-ai-overview), Perplexity (sonar), Gemini (gemini-3.8-flash), Claude (claude-sonnet-5), Grok (grok-4.3), Le Chat (mistral-medium-latest), Google AI Mode (google-ai-mode), each with web search on, asked 2 times on 5 October 2026.
What the Why counts mean
Pages: third-party pages the answers cited, not Wise's own site. Search: the engines' searches, looked up in Google's first ten (9 of 11 could be). Memory: 5 engines asked with search off.
What it is not
Not Answer Share: one question, 16 answers, no weighting; not a ranking or an endorsement. Wise's 16 of 16 is a range of 81 to 100%.
Evidence, not mentions.
Create your account, add your site, and see what Vercatus already reads on your domain before you choose a plan.
Cancel any time · Data stored in the EU