Why do AI engines cite
one source over another?

Citation-probability model

The V9.0 cross-encoder scores every (prompt, page) pair and returns a calibrated P(cite) — an honest probability, not a vanity score. Above 70% or below 30% it commits; the middle band is called an explicit toss-up.

Location-aware ranking

A prompt about Warsaw shouldn't be won by a Toruń-only page. Sorontai reads the page's real city from its domain, slug, title and entities, caps clear mismatches and boosts genuine local matches.

Best page per prompt

Given just a domain, it finds the URL that ranks highest on your site in Google for each query (site: search) and scores that page — not the homepage against a pricing question.

SEE Evidence Map

Segment-level ablation removes each part of the page — URL, title, entities, body — and re-scores, showing how many percentage points of P(cite) each part is actually worth.

Measured AI visibility

Real ChatGPT / Perplexity / Gemini / Claude answers are captured for your prompts, so you see where the brand is cited, merely mentioned, or absent — with the competitors that win each answer.

Crawler that gets in

Tiered fetch escalates from plain HTTP to a Googlebot user-agent, TLS-fingerprint impersonation and a stealth browser, with domain-specific Bright Data collectors for the hardest sites.

Gaps → action backlog

Every gap becomes a prioritized, evidence-backed action with acceptance criteria for the content team, exportable to Markdown or JSON.

Training export

New prompts and crawled pages accumulate for the next model round, exportable by date range — so the ranker keeps improving on your own market's data.

A full scored report, end to end

A real local-service audit rendered by the same engine that runs live: the AI-visibility score, per-prompt citation rankings, the SEE Evidence Map, gaps, the action backlog and measured-visibility layout. Nothing is uploaded and no site is crawled — it's a precomputed case so you can explore every tab.