Mechanism and Evidence Capabilities

How many tokens represent the brand name in cl100k_base and o200k_base, with optional competitor context. The count is reported directly rather than converted into a universal mention-probability claim.

Uses js-tiktoken to encode the brand name plus up to 3 competitors and 10 common vocabulary terms in both cl100k and o200k. The score is derived from the ratio of your brand's token count to the category average. Single-token brands score near 100; brands exceeding the average token count score proportionally lower.

Tokenization is stable for a particular vocabulary and useful as one structural input. It should not be treated as proof that a model will or will not mention a brand.

How the configured provider probes describe the brand, plus whether the entity-source waterfall resolves a structured entity record.

Runs 3 response probes across up to 5 LLM providers and classifies each response with a keyword rubric. Separately, the entity waterfall tries configured knowledge-graph sources and records source confidence. A schema post-processor inspects the submitted page's JSON-LD.

Confused or inconsistent responses can reveal an identity problem worth investigating. A missing Wikidata match is one absent corroborating signal, not proof that every provider fails to recognize the brand.

Which chunks rank highest inside the corpus assembled by this audit for category-derived prompts.

Collects owned content from the crawled site and relevant Wikipedia articles. Content is chunked, then embedded locally via MiniLM-L6-v2 (384 dimensions) and stored in pgvector. Five category-specific prompts generated from the Stage 1.5 classification are embedded and compared with the stored chunks using top-20 cosine similarity. The score reflects the proportion of top-20 results that come from owned content versus competitor or third-party sources.

This local retrieval simulation can expose weak or missing owned-page coverage. It does not reproduce a commercial provider's private index and cannot guarantee citation behavior.

The top-k probability assigned to the brand token in the recorded category prompts on providers that expose log probabilities.

Uses the active scale's prompt, sample, and provider configuration. It scans the first generated positions for the brand prefix and bootstraps an interval over completed samples. Failed or mock-only rows are marked degraded.

This is a direct observation for specific prompts and provider endpoints. It does not expose training data and should not be generalized to every prompt, model, or answer format.

The statistical association between your brand name and page-derived anchor terms, with optional context from up to three classified competitors whose pages fetched successfully. Pointwise mutual information (PMI) quantifies co-occurrence inside this audit's corpus.

Builds a per-audit corpus from the submitted page and successfully fetched competitor pages. It does not use the bundled demo corpus for customer scores. If the corpus has fewer than 200 tokens or no page-derived anchors, the dimension returns an error quality state and no score contribution.

PMI describes co-occurrence inside the corpus used for the calculation. It does not reveal a provider's training corpus or guarantee generated wording.

How recorded provider responses vary across seven prompt conditions: no-web, retrieval-requested, completion, balanced, system-constrained, safety, and model-comparison.

The database retains historical layer-shaped field names, but the implementation changes prompts; it cannot isolate pretraining, instruction tuning, RLHF, system prompts, safety filters, or fine-tuning. Unsupported and failed cells remain explicit.

Prompt-condition differences can identify responses worth inspecting. They must not be presented as access to a provider's hidden training or alignment stack.

Whether configured public-source checks find the brand in selected aggregator or roundup results, plus citations parsed from captured provider responses.

Two components. First: search against a registry of known aggregators to check whether the brand appears. Second: citation extraction from provider responses that were actually captured during the run. The source waterfall can try configured fallbacks when a source is empty or unavailable.

Source consistency and captured citations are useful public-evidence checks. The audit does not know whether a source was in a provider's training data or treated as ground truth.

Topic clusters present in successfully fetched competitor content but not matched in the submitted page. The result is a difference in this audit's corpus, not proof of an external retrieval loss.

Fetches available content from the submitted site and classified competitor pages through the configured page-source waterfall. Content is divided into passages, embedded, and grouped by topic. Gap severity summarizes the distance between successfully fetched competitor topics and the submitted page.

A successfully measured gap shows a topic difference in the fetched corpus. It does not prove which page an external retrieval system will serve or that every gap deserves new content.

Candidate contradictions between a captured no-page model response and facts present on the submitted page.

One call captures a fact response without page content. A second call compares that response with the public page and flags direct contradictions. The page is only an audit reference, so flagged rows require manual verification.

A verified contradiction is actionable evidence for correcting owned facts or relevant profiles. The module does not prove the provider's training source or the prevalence of the claim elsewhere.

Variation in recorded brand-mention outcomes across generated wording variants for the audited category.

Builds prompt variants, sends them through the configured providers, records the returned text and mention outcome, and summarizes observed variance. The paired intent view classifies those variants with rules; it does not infer private user-query demand.

Prompt wording can materially change an observed response. This capability describes the sampled prompts and providers only, and does not estimate a universal mention rate.

How five derived measurements changed from a prior audit and where they sit relative to product-configured heuristic thresholds.

Defines five feedback-loop metrics and compares the current audit with prior snapshots for the same brand. When history is insufficient, the result is marked degraded rather than presented as measured velocity.

Repeated measurements can show direction and support follow-up investigation. The configured thresholds are not validated causal tipping points, so this capability is labelled degraded even with history.

A modeled composite-score change for a simulated issue fix.

For qualifying issues, the processor can call the configured prediction sidecar with the current feature vector and a simulated fix. If the sidecar is unavailable, the client can return a labelled fallback estimate. The stored prediction source identifies which path produced the result.

A labelled estimate can help rank experiments, but it is not a promised causal lift. Estimates should not be added together unless the model evaluated the combined intervention.

How recorded mention rates differ across rule-classified prompt intents: informational, navigational, transactional, and commercial.

A rule-based classifier categorizes prompt variants from the phrasing-sensitivity module. The product weighting gives commercial and transactional classes more score weight; that weighting is a product choice, not measured revenue attribution.

Breaking out the recorded mention rate by intent can reveal prompt-specific variation without claiming that one class caused a business outcome.

Lexicon-based sentiment and attribute terms in the provider responses captured by this audit.

Runs lexicon-based sentiment analysis on each provider's raw text output about the brand. Extracts attribute terms (adjectives, descriptive phrases) per provider. Results are compared across providers to identify consensus attributes (mentioned by 3+ providers) vs. outliers. The score reflects sentiment positivity, attribute consistency, and alignment with the brand's stated positioning.

The recorded wording can reveal a response worth reviewing. It is not an exact measurement of a provider's overall perception or the cause of that wording.

A six-group allocation of a score gap only when a classified competitor has a completed ResourceAI audit with comparable dimensions.

Matches the latest completed audit by competitor brand name, keeps dimensions with usable scores in both audits, and allocates the positive weighted gap across response proxy, retrieval, entity, content, tokenizer, and other groups. Without a match, no decomposition is emitted.

Matched-audit provenance makes a comparison inspectable. It remains a score decomposition, not proof that a component caused market performance.

A time-stamped search-result snapshot of whether an AI Overview was returned for up to five generated category queries and whether its text mentions the brand.

Runs only when the configured search-result source is available. For returned overviews, it records the query, text, cited sources, brand mention position, and sentiment. When the source is unavailable, the capability is marked unavailable rather than reported as an empty result.

When configured and successfully returned, this is direct query-result evidence for that place and time. It is not universal ground truth across users, locations, queries, or later results.