Skip to main content
Share of voice and citation rate tell you whether you won a prompt. This page tells you why — which pages the answer read, how much weight each one carried, and what to do about the ones that are not yours.

​
The grain

One group per prompt, per model, per surface. Surface is in the grouping for the same reason it is everywhere else in the product: a consumer answer and an API answer draw on different pages, and averaging them produces a plan that fits neither. Country is deliberately not in the grouping. A page that decides a prompt in the US decides it in the UK too, and splitting by country here would quarter every sample for no gain in what to do next.

​
Weight, not raw citations

Sources are ranked by citation share × position, not by how many times a page was cited. A page cited once at the top of every answer decided more than a page cited twice near the bottom of one, and a ranking that could not tell them apart would keep sending you after the loudest source instead of the one that actually decides. Position weight is 1 / mean position. A citation the model attached without a clear position is scored as if it sat at the back of the list — stated as such, rather than dropped or quietly treated as first.

​
Classification: yours, a competitor’s, or someone else’s

Every source in a group is classified the same way the sources report classifies it, so a domain is never “ours” on one screen and “third party” on another:
  • Yours — a page on your own domain.
  • A competitor’s — a page on a domain you track as a competitor.
  • Someone else’s — everything else: review sites, forums, documentation hubs, comparison pages, listicles.

​
”You are on it”

A source is marked on it in either of two cases:
  • it is your own domain, or
  • you have a placement there whose execution status is verified.
An unverified or unknown placement is never shown as on it. This is the same discipline as everywhere else in the product: a claim without evidence does not get to look true.

​
The next move

Each source carries a recommendation: fix your own page, get listed, pitch, answer the thread, or write something cite-able. Today, every workspace gets this from one static, per-engine table, because the observations layer this recommendation could instead draw on does not have enough evidence yet: Once your own workspace has read enough answers on a given source — at least 20 successful reads — the recommendation switches from the static table to what your own history shows. A recommendation always says which one it came from, so you know whether you are looking at a general rule about the engine or a pattern proven in your own data.
Every row and every group carries n: the successful answers behind it, in the same window and counted the same way as the rest of the product. A share without its n is noise in a confident font.

​
Where to find it

  • On the home page, as the default view onto why you are winning or losing.
  • As the agent tool deciding_sources, so a chat or automation can pull the same map.
  • Over the REST API, at GET /v1/deciding-sources (and its equivalent under a project, GET /v2/projects/{id}/deciding-sources) — both take ?prompt= to narrow to one prompt and ?days= to set the window.
REST
MCP
See Reading the numbers for how to treat n, null, and significance everywhere else these shapes appear.