Thirty seconds, six engines in parallel, two original algorithms. Every step in plain English, no fog.
Any public URL. No card, no setup, no account for the first three. Press enter and Fanalis takes over.
Before we score, we figure out what kind of site you are. A government service is not a SaaS landing. A docs site is not a creative portfolio. We classify you across eleven genres and apply the weighting that fits.
Real headless Chromium opens your URL the way a visitor's browser does. We screenshot, parse the post-JS DOM, and collect the public files (robots.txt, sitemap.xml, llms.txt).
Lighthouse (3 runs, median), axe-core for WCAG, computer-vision metrics, security-header scan, structured-data parsing, and our owned visual model. They feed each other so the second engine never re-fetches what the first one already saw.
FVRS combines ten visual channels via a weighted geometric mean with a Bradley-Terry anchor. FCRS combines four funnel stages via a soft-min bottleneck. Both genre-conditioned.
Each pillar (Speed, Accessibility, Search, AI, Design, Conversion, Security) has a 0–100 score. The composite is a genre-weighted average — so a search engine is not judged like a failed SaaS landing.
Every finding gets sorted and SHA-256-hashed into a 16-character signature. Same input today, same hash tomorrow. The signature is on every report — it's how you'd argue with the score later.
Two tabs: Plain English and Technical. Plain shows the verdict and the fix; Technical shows the formulas, hashes, weights, KV evidence. Same number, different audiences.
Most tools run one thing. Fanalis runs six in parallel and cross-references them — that's why it's thirty seconds, not ninety. Here's every engine, what it owns, and what it feeds.
Performance + accessibility + SEO. Median of three runs, because single runs drift.
→ feeds Speed, Accessibility, Search
WCAG 2.2 ruleset, the same engine Chrome ships in DevTools. Plus our landmark/skip-link/lang hardening.
→ feeds Accessibility
Pixel sweep of the screenshot — colour palette, hue harmony, whitespace dominance, edge density.
→ feeds Design, FVRS
A small MLP head on a frozen SigLIP backbone, distilled from a frontier judge. Runs on our infrastructure.
→ feeds Design, FVRS
A frontier multimodal model gives the second opinion on the design axes. Blended 50/50 with the owned model.
→ feeds Design, FVRS
HTTP Observatory rubric — CSP class, HSTS, frame-ancestors, content-type, cookies, mixed content, SRI.
→ feeds Security
How well your site reads as designed. Ten channels — type system, modular scale, palette, harmony, salience, rhythm, polish, cross-strata consistency, brand singularity, judged aesthetic. Combined via a weighted geometric mean so one weak channel can't hide behind nine strong ones, with a Bradley-Terry anchor that pulls the elite tail apart.
How well your page sets up the visitor to act. A four-stage funnel — Arrive, Orient, Consider, Commit — with 26 leaf signals. Stages combine via a soft-min bottleneck so a single broken stage genuinely caps the score (Liebig's law for the funnel).
Both algorithms are documented in full on /method and reproduce identically for the same input — every audit ships a SHA-256 signature you can verify.
Lighthouse's own docs concede 5–10-point drift between runs. We run it three times and median-aggregate. Then we hash the sorted set of findings into a SHA-256 we surface in every report. If you re-run the same site tomorrow and the hash drifts — that's a real change, not algorithm noise.
Three free with an account. ₹5 each after. Below 70 → Catalyst quote on the last page of your report.
