Observation protocol

How the evidence is built.

A transparent method for tracking non-deterministic AI answers without pretending they behave like fixed rankings.

1. Controlled prompts

Each brand is monitored with repeatable buyer-intent prompts tied either to its product category or to its service and US location. mentionpop runs the configured prompt set on a weekly cadence.

2. Two distinct surfaces

GEO observations cover generative answers from ChatGPT and Gemini. Studio AEO observations cover Google AI Overviews, featured snippets, and People Also Ask.

3. Evidence capture

Each successful observation records the engine or surface, timestamp, mention status, position when detectable, cited URLs, and a capped copy of the raw response.

4. Answer framing

Observations that name the brand are also read for how the brand is characterised, not only that it appeared: recommended, listed without detail, hedged with caveats, described inaccurately, or criticised outright. Every verdict has to cite the sentence it came from, and that sentence is stored only when it is found verbatim in the captured answer. A verdict whose quote cannot be traced back to the transcript is kept at reduced confidence with no quote attached. Answers that never name the brand are recorded as not mentioned rather than scored, because an answer that omits a brand contains no framing to read.

5. Claim checking

Each brand's own published description of itself is captured separately from the AI answers. Assertions an answer makes about the brand are set against that description and flagged either as contradicted, where the assertion conflicts with it, or unsupported, where the assertion is absent from it and cannot be checked. Because that description summarises a homepage rather than an entire company, unsupported means unverified and not disproven — it marks a claim for human review rather than declaring it false. Every flagged assertion must quote the answer verbatim, and a flag whose quote cannot be found in the stored answer is discarded rather than reported.

6. Comparative treatment

When an answer names both the brand and its tracked competitors, the treatment each receives is compared on detail, warmth, ordering, and hedging, and recorded as balanced, brand favoured, or brand disadvantaged. The comparison is confined to a single answer from a single prompt on a single engine, because that is the only setting where differing treatment reflects the answer rather than differing questions.

7. Provider dependency

Monitoring prompts and search observations are processed through DataForSEO. The resulting reports are sampled provider observations, not guaranteed replicas of every personalized consumer interface. Third-party outages, model changes, and retrieval differences can affect collection.

8. Accuracy validation

Selected internal checks compare DataForSEO collection with web-grounded Perplexity Sonar output before strong detection claims are made. Perplexity is a validation source; it is not included in the default weekly monitoring allowance.

9. Rates and deltas

Mention rates use successful observations only; failed API calls do not count as brand misses. Reports compare the current observation window with the previous window and display GEO and AEO separately.

Limits

AI output varies across time, wording, model versions, personalization, and device. mentionpop is a change-detection and evidence system, not a universal ranking oracle or a guarantee of traffic, citations, or revenue.

Learn each measurement

Start with the AI visibility tracker guide, then review how mentionpop measures GEO performance, checks AEO visibility, and separates mentions from AI citations.