Buyers build their shortlist inside AI answers now, before any touchpoint you can track. We ask the real questions, show you what came back, and name who got recommended instead of you.
Free, no account, about a minute. Five real questions with web search on. You get the answer text, not a score.
linear appeared in 3 of 4 answers.
You are in the conversation but not in all of it. The answers where you are missing, and who was named instead, are below.
the receipt
"I need help choosing project management tools for software teams. What should I look for?"
For software teams, look first at a tool that matches your actual workflow. Jira remains the standard for agile and issue tracking, while ClickUp and Asana suit teams that want...
sources it read
share of voice
Weighted by position and by how commercial the question is. Coverage 100%, 4 of 4 answers returned.
Your analytics show nothing
A buyer asks an engine, reads three names, and picks one. None of that touches your site, so nothing in your funnel records that you were considered and skipped.
Asking it yourself lies to you
Your session carries your history and your location, so you see a friendlier answer than a stranger does. A clean call from the market a buyer is in is a different answer.
The answer moves
Ask twice and you get two results. Without a repeatable set and a variance test, you cannot tell a real change from normal drift.
A score does not tell you what to do
Knowing you are at 12% is not an action. Knowing which page the engine read to answer without you is a call you can make on Monday.
Every claim carries the exact question, the engine, the model and the day it was asked. You can read the paragraph that produced the number, and so can the person you forward it to.
A finding has a name attached: a competitor the engine recommended while you went unmentioned, or a page it read to answer without you. That is something a person can act on this week.
The scoring algorithm, the weights, the confidence test and the places it can be wrong are all published. Every figure opens to the answer that produced it. Most tools ask you to trust a number; this one hands you the working.
We ran Linear against the question a buyer would actually ask about project management tools. It is a useful test because the name is also an ordinary word, which is the case that breaks naive matching.
| rank | brand | appeared in | weighted share |
|---|---|---|---|
| 01 | Jira | 4 of 4 | 23.6% |
| 02 | Asana | 4 of 4 | 12.6% |
| 03 | Linear | 3 of 4 | 12.0% |
| 04 | Notion | 3 of 4 | 8.1% |
| 05 | ClickUp | 2 of 4 | 7.7% |
Measured against the live API on 30 July 2026, not read off a price page. Cost tracks the model and the token budget, not the request, and the spread between the cheapest and dearest engine is 3.8 times. Any plan priced per prompt without also fixing the engine is selling a margin nobody controls.
| rank | engine | class | avg per call |
|---|---|---|---|
| 01 | Perplexity | Search-native | $0.0256 |
| 02 | Claude | Frontier | $0.0631 |
| 03 | Gemini | Fast | $0.0702 |
| 04 | ChatGPT | Frontier | $0.0964 |
The reasoning trace is not the answer
Reasoning models return a private draft alongside the reply. That draft names brands the answer then drops. We exclude it and record how many items we dropped, because counting it would report you as visible where no buyer ever sees you.
Whole word matching, never substring
We measured a brand name that is also an ordinary word returning 2,420 mentions from an index while the brand's own domain returned zero on the same query. Three orders of magnitude of noise.
A failed call is not a zero
It is excluded from both sides of the ratio and the coverage is declared. An answer that named nobody is not a loss either; it means the question did not discriminate.
No confidence band under three runs
A single measurement dressed as precision is how credibility gets spent. Two snapshots differ reportably only when their intervals do not overlap.
Unknown competitors count
Every brand the answer names enters the denominator, listed or not. A share of voice that improves because you stopped tracking a rival is a corrupt number.
Nothing is generated by a model guessing
Every result comes from a real call to a real engine, and the raw response is stored. Without a provider connected, the feature stays inert and writes nothing.
A report is a saved artifact, not a live view. It says the same thing two weeks later, which is what lets it travel inside a company with nobody there to explain it.
The verdict
One sentence. Whether you showed up at all, and where.
Who got named
The brands the engine recommended, in the order it recommended them.
The sources it read
Concrete URLs behind the answers. This is your press and content target list.
Share of voice
Weighted by position and by how commercial the question is, with coverage declared.
Our own first full run measured 15 questions across 4 engines for $4.01, stored 51 answers, and collected 460 cited URLs across 190 domains. The ledger reconciled exactly against the provider.
Point Claude or ChatGPT at your account over MCP and ask it things no dashboard anticipated. Compare two runs. Pull every source that cited a competitor and not you. Ask which question you lost this month. The answers are projected to the useful fields, because a raw provider envelope will drown an agent's context before it does anything.
Each comparison opens by saying when to pick them instead. A page that always concludes you should buy from us is one nobody believes, and we would rather lose the deal we were going to lose anyway.
Five real questions, a live engine with web search on, and the full answer text. No account, no card.