Propps measures a defined sample of AI answers, not everything an AI system might say. Our goal is to make the sample, evidence and limitations clear enough for you to assess each result.
What is collected
A measurement sends your approved prompts to configured provider APIs without conversation history. Propps does not automatically insert your brand into the measurement prompt. We save the prompt version, provider, model, mode, context, settings, time and original response.
Basic, Medium and Thorough use one, three and five independent answers per configured combination. Plan eligibility applies. Search-enabled and knowledge-only answers form separate series. Enabling search does not guarantee that the provider searched; we retain actual search evidence where the provider exposes it.
API answers versus consumer products
These are API measurements. They do not reproduce every personalised session in ChatGPT, Gemini, Claude or Perplexity. Consumer applications can use different models, tools, prompts, locations and history. Country and language context is not a guarantee of perfectly localised search.
What the metrics mean
| Metric | Interpretation |
|---|---|
| Mention rate | Share of analysed answers that mention the brand |
| Recommendation rate | Share of analysed answers that recommend the brand |
| Top recommendation rate | Share of analysed answers with an explicit first recommendation for the brand |
| Citation rate | Share of valid answers citing an owned page |
| Prompt coverage | Share of measured prompt versions with a brand mention |
Mention order and explicit recommendation rank are different. A mention is not automatically positive sentiment or an endorsement. Source citations remain separate from incidental links and unused search results.
Missing evidence and comparisons
Failed calls are not brand absences. Missing analyses are excluded from metrics that depend on analysis. Reports show the usable answer counts and coverage. A zero score is a measured result; a missing score means there is insufficient usable evidence.
Current visibility includes available answers in the selected period, including the current week. Run reviews compare matching configurations against the preceding eight complete calendar weeks, anchored to the run start in the workspace timezone. Other comparison windows are labelled in the dashboard. Repetitions are pooled within each series and week; comparable series and weeks receive equal weight.
Relative percentage change is unavailable when the baseline is zero. New configurations and limited history need separate interpretation. We do not claim statistical significance or causation from a change in these samples.
AI interpretation and corrections
Analysis and summaries use AI and may be wrong. Evidence links let you check the original answers. Coaching suggests hypotheses and possible actions, not guaranteed ranking improvements. Human corrections remain traceable; affected saved reviews can be marked outdated instead of silently rewritten.
Website-assisted onboarding is separate from answer evidence. It reads a bounded selection of public company pages for setup. Cited sources in scan answers are not crawled by the reporting process.