How we measure

1. Measurement protocol

Every engagement starts from a fixed set of buyer questions: the queries your customers actually put to AI engines. The set is fixed on purpose: if the questions change between measurements, the numbers stop being comparable.

Each question is asked 3 times per engine, every week, under the same conditions. AI answers are non-deterministic, and the same question can name different brands an hour apart, so a single run is an anecdote, not a measurement.

• We report frequency, never yes/no: "cited in 2 of 3 runs", not "cited".
• Re-measurement runs weekly on the same question set and the same protocol, so week-over-week movement is attributable to something real.

2. Engines measured

We measure engines we can query reliably and repeatedly. We do not fill in engines we cannot measure with approximations.

• Measuring now: ChatGPT, Gemini, Perplexity.
• In preparation: Google AI Overviews. No official API is available yet. We will add it when we can run the full protocol against it, not before.

3. Two-layer metric: Mentions and Citations

We track two layers instead of a single score, because they move independently and call for different work.

• Mentions: the answer names your brand.
• Citations: the answer links your page as its source.
• A brand can be named without ever being sourced, or sourced without being recommended. Separating the layers tells you which lever to pull.

4. Movability estimation

Not every query can be moved. Some respond to content and entity work within weeks; others are dominated by aggregators or entrenched brand recall and are unlikely to shift in the short term.

We estimate movability per query and label it in the report. Where we do not expect movement, we say so before you spend money on it.

5. Known limitations

Measuring a non-deterministic system comes with error bars, and we would rather show them than hide them.

• Personalization: engines adapt to accounts and context. We measure from clean sessions, but what your customers see may differ from what we record.
• Time and version variance: answers shift with time of day and with silent model updates. Weekly re-measurement catches the drift; it cannot eliminate it.
• Sampling error: 3 runs per query is a sample, not a census. Reported frequencies are estimates.
• Because of all of the above, every figure in a MentionUp report carries a confidence label: high, medium, or low.

6. What we never do

Some tactics can inflate short-term numbers while exposing your brand to engine penalties. We do not use them, at any price.

• No link buying.
• No private blog networks (PBN).
• No mass AI-generated publishing.