How Akrux measures AI visibility

Prompt generation and classification

A live scan first researches the business using public web evidence when provider access is available. Prompt generation uses that research plus the submitted name, category, city, market, aliases and competitors. If generation fails, deterministic templates preserve the same intent structure.

The default 25-prompt intent mix is 12% branded, 12% comparison, 20% category, 20% best-of, 20% purchase and 16% informational. Largest-remainder allocation turns that into whole prompts. A prompt is branded only when its normalized text contains the target name or a known alias; the flag is stored with the prompt.

Mention and position detection

Known entities are detected with normalized aliases, Russian–English transliteration and conservative fuzzy matching. The target's recommendation position is the order in which recognized business entities appear in the answer. Deterministic positive matches take precedence over the language-model extraction pass.

The batched extraction pass can add sentiment and unknown company names. It does not erase a deterministic target mention. Repeated mentions of the same entity in one answer count once for Share of Voice.

Visibility Score formula

For every model family actually queried, Akrux calculates a category mention rate over category, best-of and purchase prompts, and a comparison mention rate over unbranded comparison prompts. The category rate has weight 0.60 and the comparison rate 0.40. Missing groups are omitted and the remaining weights are normalized.

Cross-model results use weights of 0.30 ChatGPT, 0.25 Gemini, 0.20 Perplexity, 0.15 Claude and 0.10 Grok, normalized over the models present in that scan. If the average organic mention position is 2 or better, the score receives a 1.15 multiplier. The result is capped at 100, and the bonus cannot turn an imperfect baseline into 100.

Branded prompts and informational prompts do not enter the primary category/comparison formula. Branded answers produce a separate recognition percentage. Informational unbranded answers can still contribute to overall evidence counts and Share of Voice.

Share of Voice formula

Akrux uses successful unbranded answers. It counts each qualifying business entity at most once per answer, sums those entity mentions, then divides each entity's count by the total qualifying mentions. Configured competitors remain visible with zero mentions; newly detected entities must pass competitor classification.

Known directories, maps, marketplaces, social or source platforms and generic descriptive phrases are not competitor rows. This filtering reduces category noise, but classification can still be imperfect and should be reviewed against the underlying answers.

Failed requests, freshness and versioning

A failed model request is stored as failed and excluded from metric denominators. A partial scan is labeled partial. Extraction failure falls back to deterministic detection for the affected answers; it does not turn missing sentiment or unknown competitors into invented values.

Every scan issues fresh provider requests; answers are not cached across scans. Each score snapshot stores a metric version. The current implementation is version 2, so future formula changes can be separated from historical results.

Important limitations

• Generated answers are probabilistic and can change between otherwise similar runs.
• Provider-hosted model behavior may differ from consumer chat interfaces, personalization and geography.
• A prompt set samples likely discovery behavior; it does not represent every customer question.
• Entity extraction and competitor classification require human review in ambiguous categories.
• Visibility measures appearance in collected answers, not revenue, customer intent or guaranteed future recommendations.