Methodology
Arcanoris measures whether AI answer engines recommend a brand when buyers ask commercial questions. Every report embeds the methodology version, timestamp, models, and sample size used.
Provider sampling
We query official provider APIs only - never scraped consumer interfaces. API samples are not guaranteed to match what an individual user sees in a consumer chat interface, and we label every result accordingly.
ChatGPT
Responses API with the web_search tool. Answers reflect API sampling, not consumer ChatGPT.
Claude
Claude model responses via API. Source URLs are labelled as model-suggested unless independently grounded.
Gemini, Perplexity, Grok
Each queried via its official API and labelled with the exact model that answered.
Llama, Mistral, Nova, DeepSeek, Kimi, Groq, MiniMax, Sarvam, Qwen
Additional model perspectives via their APIs. The exact model is stored with every answer; a provider that cannot answer marks the scan partial rather than vanishing.
Unbiased prompts
Buyer-intent questions are generated from your category and audience - never from your brand name. Mentioning the brand in the question would prime the model and invalidate the measurement. Prompts span discovery, comparison, alternative, and purchase-intent phrasing.
Scoring weights
The AI Visibility Score is a 0–100 weighted composite. Each query's sub-scores are averaged per provider, then aggregated across providers.
Mention
65%Whether your brand appears in the answer at all, matched against your name, aliases, and domain.
Position
30%Where you appear in the ranked recommendations - first mentions earn full credit, later positions decay.
Citation
0%A diagnostic showing whether a search-grounded answer returned verifiable annotations. It does not affect the core visibility score; independently discovered mentions are reported separately.
Data confidence
5%How complete the evidence behind this scan is: whether providers returned usable answers and whether competitor websites could be read. A thin scan scores lower because less was verified.
Limitations
- AI answers are non-deterministic; scores can vary between runs even with identical inputs.
- We never fabricate provider results. If a provider API fails after retries, the scan completes as partial and the failure is disclosed on the report.
- Recommendations are directional guidance tied to observed answers, website evidence, grounded citations when available, and separately verified web mentions. They are not ranking guarantees.
- Provider APIs may use different model snapshots than consumer products; each report records the exact models sampled.