How ceeme works

What exactly does ceeme measure?

ceeme asks AI engines real customer questions and measures three things separately: are you mentioned, is your source cited, are you chosen. Adding those three into one number hides which of the three is broken, so we do not.

Why a prompt and not a keyword?

Because nobody says a keyword to an assistant. People type a situation: "who do you recommend in my area". The engine splits that question into sub-questions itself and answers them separately. Measure on a keyword and you measure a question nobody asks.

One trap we walked into ourselves: put the company name in the prompt and you plant the brand in the model's head, then ask whether the brand is there. That is not a measurement, that is a leading question. So the name only comes in after the answer.

How does ceeme know why AI chooses?

By looking at what underpins the answer. Decision Drivers are the factors that recur for those who do get chosen: source quality, freshness, whether you are recognisable as an entity, citations, local context and human signals. Not a black box — factors you can act on.

Who does the work?

Thirteen agents, each with one task. They hand off to each other, and between writing and publishing sits a gate only you open.

REWORK If Judge rejects, the work goes back to Scribe. That loop runs until it is right or a human decides.

APPROVAL GATE Between Judge and Launcher sits the approval gate. Nothing is published without your approval.

What does that look like in code?

Like this. This is the shape of what goes back and forth — the prompt without a brand name, the answer with its sources, the weighting we publish instead of hide, and the machine-readable profile an AI agent can request. The example company sits on an .example domain: by internet convention, that can belong to nobody.

{
  "prompt": "Which company do you recommend in my area?",
  "markt":  "BE-en",
  "persona": "local buyer",
  "engine": "claude",
  "merk_in_prompt": false
}

{
  "genoemd":   ["Average Ltd", "Example plc"],
  "volgorde":  [1, 2],
  "bronnen":   ["example.com/about", "wikipedia.org/..."],
  "live_gezocht": true,
  "uitkomst":  "cited",
  "gemeten_op": "2026-08-07T04:12:00Z"
}

// A composite score hides which lever is broken.
// That is why we publish the weighting, with source and month.
export const WEGING = {
  mentioned: 1,   // your name appears
  cited:     2,   // your source is cited
  chosen:    3,   // you are recommended
};
export function scoor(runs) {
  return runs.reduce((s, r) => s + (WEGING[r.uitkomst] || 0), 0);
}

curl -s "https://ceeme.be/api/entity?site=average.example" \
  -H "accept: application/ld+json"

# Every ceeme profile can be queried as a machine-readable
# record. That is what an AI agent needs: not a page to
# read, but a fact to verify.

What happens to what is wrong?

You get the exact fix, with the reason and safe steps to apply it. Where you gave access, ceeme executes it — never without your approval, and always with a way back. What we cannot execute, we deliver prepared.

How do you prove it worked?

By asking the same prompts again. Not similar ones — the same, in the same market and the same language. If mentioned shifts to cited, or cited to chosen, that is the proof. If nothing shifts, we say that too.

Frequently asked questions

Read on