This page explains what RunPR measures, how evidence is collected, what each number means, and what the data cannot support. The IAB framework is independent industry guidance. RunPR applies its vocabulary and quality criteria on its own.
Framework edition: August 2026 · RunPR methodology published August 28, 2026
Each scan freezes the engines it actually tests and the model identifier returned by RunPR’s configuration. The saved run, not this general page, is the record for a specific report. Missing historical identifiers appear as Unknown.
| Platform | Current configured model | Collection path | Scan mode |
|---|---|---|---|
| ChatGPT | gpt-5.6-terra | OpenAI Responses API with web search requested | Standard and Deep |
| Perplexity | sonar | Perplexity Sonar API with native source records | Standard and Deep |
| Exa | exa-answer | Exa Answer API with returned source records | Standard and Deep |
| Claude | claude-sonnet-5 | Anthropic Messages API with web search requested | Deep |
| Gemini | gemini-2.5-flash | Google Generative AI API with search grounding requested | Deep |
| Grok | grok-4.5 | xAI Responses API with web search requested | Deep |
Engines are included only when the deployment has the required provider access. A run discloses tested engines, errors, and exact captured identifiers. RunPR does not use browser scraping to collect AI answers. Provider interfaces, search behavior, model routing, limits, and access restrictions may change outside RunPR’s control.
A generation model drafts up to 24 prompts grounded in the tracked homepage, an operator description, and/or configured competitors. Prompts span RunPR’s brand, founder, discovery, competitor, and topical categories. Operators can edit, add, remove, order, and activate prompts for later scans.
The default library is synthetic and provider-generated from client-specific grounding. It is not derived from observed consumer behavior, search-volume panels, demographic panels, or platform-native query logs. Client edits are first-party choices, not evidence of population frequency.
Active prompts are weighted equally. Provider fan-out does not make a prompt more representative. Libraries are created during setup and refresh when an operator edits or regenerates them. Prompt changes break comparability and create a rebaseline event.
RunPR does not currently persist the IAB informational, comparison, recommendation, and transactional intent taxonomy. It shows its own categories and marks IAB intent coverage Unknown rather than reconstructing it from prompt wording. Demographic and cultural framing tests are not part of the current library.
RunPR sends saved prompts to official provider APIs and retains answer text, emitted source records, provider and engine identity, model identifier, timestamps, errors, mention flags, sentiment, and configured competitor detections. Different provider outputs remain visible per engine.
A normal scan retains one answer per prompt-engine cell. An opt-in repeat mode can ask the identical cell up to five times and retain its hit distribution. One response is not enough to characterize non-determinism, so the quality classifier fails closed when repeat evidence is absent.
Plan cadence ranges from every 14 days to weekly, with a shorter Agency priority interval. Actual start and finish times are saved per run. Collection may be delayed by plan capacity or provider availability; a scheduled eligibility date is not represented as an exact collection time.
RunPR does not currently persist query geography, locale, user demographic context, or time-of-day controls. Those values appear as Unknown in run provenance. Results must not be generalized across markets. Search or retrieval is requested where each provider offers it, but the exact run-level retrieval configuration is not historically retained; it appears as Unknown unless direct evidence exists.
Every number below is derived from saved answer cells. Where the evidence stops, the metric stops with it.
Mention Rate is successful answer cells naming the tracked brand divided by successful cells. Citation Rate uses cells that emit the tracked domain. A partial monitored Share of Voice uses the tracked brand and configured competitors from the same answers. Momentum uses saved 95% Wilson intervals and never treats overlapping ranges as meaningful movement.
RunPR can report first answer-text occurrence, ordering among tracked entities, and rank only when an explicit text list supports it. These signals are partial because provider APIs do not preserve the rendered consumer viewport, cards, scroll depth, or collapsed source surfaces. Source emission is not called substantive contribution without answer-to-source evidence.
Sentiment uses a narrow deterministic lexicon around an exact brand mention. General framing is not measured. Narrow identity-confusion and unsupported legal-claim checks are shown as risk flags, not a complete hallucination rate. Factual contradictions require an exact AI quote and an exact conflicting quote from a bounded rendered crawl of the client site; silence remains unverified.
Recommendation Strength is partial and counts only explicit deterministic recommendation language in eligible unaided answers. Ambiguous language and a citation alone do not pass. Post-Citation Click-Through Rate is not measured because RunPR does not join real platform or first-party click streams to these answers. Drafts, reviews, sends, coverage, and workflow activity are never used as click proxies.
Exact brand, domain, founder, and configured competitor patterns are evaluated against saved answer text. Sentiment looks only at a bounded window around a detected brand mention. Missing text, provider errors, and unsupported labels stay missing.
Provider-emitted links remain first-class evidence. For answers that mention a brand without an emitted source, a gated matching system can compare exact answer spans with a frozen, tenant-scoped source corpus. Confidence bands and human review are preserved; a likely source is never rewritten as an emitted citation.
The client’s rendered site is the bounded first-party reference. Explicit contradictions use deterministic gates after a comparison pass. Missing pages, vague copy, different scope, or no statement produce Unverified, not an error claim. Core visibility is not validated against a passive consumer panel or platform-native behavioral reporting.
Repeat-mode cells retain attempted runs, successful runs, mention frequency, errors, and volatile cells. Single-shot and legacy scans show Unknown. RunPR does not invent a universal acceptable-variation threshold because the IAB framework leaves that threshold open by metric, platform, and query type.
Exact model identifiers are compared between runs. A model change is annotated. Prompt-set changes are also detected. Either event prevents a like-for-like trend and creates a rebaseline-required event in the measurement projection.
A movement is called platform-driven only with evidence across multiple brands or the full category on one platform after brand, competitor, and query-set causes are ruled out. Current RunPR data usually cannot prove that pattern, so cause remains Unknown. Market-driven classification likewise requires corroborating first-party or competitive evidence.
The classifier evaluates sample size, query volume, intent coverage, cadence, reproducibility, validation, documentation, platform coverage, and aggregation independently. Fewer than 50 unique prompts is treated as exploratory. Cross-platform fan-out does not inflate that count. Decision-grade requires every required criterion to be supported by saved evidence. Missing or unknown evidence fails the criterion. Today’s normal 24-prompt, single-response scans therefore show Not enough evidence even when their directional signals are useful for investigation.
| Area | Capability | Status | Evidence | Limitation |
|---|---|---|---|---|
| Presence | Mention Rate | Supported | Exact persisted answer cells with deterministic brand detection and a disclosed denominator. | Results describe the tested prompt library and engines, not all consumer AI use. |
| Presence | Citation Rate | Supported | Provider-emitted links and named source records are retained with the answer. | Implied use without an emitted or named source is excluded. |
| Presence | Share of Voice | Partial | Tracked-brand and configured-competitor appearances are counted from the same answer set. | The competitive set is client-configured and answer-level counts do not represent the full category universe. |
| Presence | Visibility Momentum | Partial | Compatible runs use saved 95% Wilson intervals and a conservative non-overlap movement rule. | Movement is not causal and becomes unavailable across an incompatible or reset baseline. |
| Prominence | Position | Partial | RunPR can report first occurrence, tracked-entity ordering, and explicit list rank from saved answer text. | RunPR does not capture the consumer interface, viewport, cards, collapsed sections, or below-fold placement, so this is not rendered-surface Position. |
| Prominence | Publisher Content Utilization | Partial | Exact answer spans can be matched to a bounded, immutable source corpus when the source-attribution gate permits. | An emitted source alone does not prove substantive contribution; absent matching evidence is reported as not measured. |
| Portrayal | Sentiment | Partial | Positive, neutral, or negative labels are saved only for detected brand mentions. | The current deterministic local-context lexicon is narrow and has no published accuracy benchmark. |
| Portrayal | Framing | Not measured | No general framing taxonomy is persisted. | RunPR does not infer leadership, value, age, or category framing from silence. |
| Portrayal | Hallucination Rate | Partial | Narrow identity-confusion and unsupported legal-claim checks retain the exact answer evidence. | These checks are not a complete hallucination-rate classifier and are not promoted to one. |
| Portrayal | Factual Inaccuracy Rate | Partial | High-precision contradictions require exact answer and client-site quotes from a bounded rendered crawl. | The client site is a limited first-party reference, and silence or ambiguous evidence remains unverified. |
| Persuasion | Recommendation Strength | Partial | Explicit recommendation or rejection wording is retained for eligible unaided answers. | Ambiguous recommendations are not scored and a citation is not treated as a recommendation. |
| Persuasion | Post-Citation Click-Through Rate | Not measured | No platform or first-party click stream is joined to visibility answers. | PR drafts, approvals, sends, and placements are not substituted for user click-through. |
| Quality & disclosure | Measurement quality tier | Supported | A deterministic fail-closed classifier evaluates every framework criterion from persisted evidence. | Current RunPR scans usually remain below Directional because the prompt library and repeat evidence are intentionally small. |
| Quality & disclosure | Run provenance | Supported | Scan time, engines, captured model IDs, prompt/response counts, categories, prior baseline, and unavailable values are projected per run. | Geography and some provider retrieval state are not captured historically. |
| Stability | Variability baseline | Partial | Opt-in repeated cells retain per-query and per-engine hit distributions. | Single-shot and legacy scans have no within-window variability baseline. |
| Stability | Drift and baseline events | Partial | Model-ID and prompt-set changes produce explicit annotations and rebaseline events. | RunPR cannot call a shift platform-driven or market-driven without category-wide or first-party corroboration; it reports Unknown. |