Skip to content
FanalisFanalis
METHOD · PAPER·OPEN · FORMULA

No black box.
Every weight, in print.

Three measurement layers, two original algorithms (FVRS, FCRS), genre-aware weights, SHA-256 receipts. The audit you see is the audit your competitor sees — same formula, same rubric.

The thesis

Every other audit tool judges every site as if it were a SaaS landing page.

Google.com isn't a failed funnel — it's a utility. lusion.co isn't broken — it's shipping WebGL on purpose. The first thing Fanalis does is figure out what kind of site it's looking at. Everything after that is weighted to match.

The seven axes

Seven independent measurements, weighted by site genre.

Each axis is scored 0–100 on its own. The composite is a weighted average — and the weights move with the site. A portfolio leans on design. A government service leans on accessibility. A shop leans on conversion. We don't use a one-size profile.

Speed

Lighthouse + 2026 CWV

Core Web Vitals against the official 2026 thresholds — LCP, CLS, INP, FCP, Speed Index, TTFB. Run three times, median taken. Single-run noise doesn't decide your score.

Anchored to Lighthouse

Accessibility

axe-core (WCAG 2.2) + hardening

The same axe-core engine Chrome ships in DevTools, plus skip-link, landmark, language and single-H1 hardening that Lighthouse alone doesn't enforce. Log-decay scoring — penalties saturate, no fake zeros.

Lighthouse score + hardening delta

Search visibility

Lighthouse SEO + hardening

Lighthouse's SEO audits, plus sitemap and robots.txt presence, OG and Twitter completeness, structured-data parsing, internal-link density, and URL hygiene the SEO category doesn't audit.

Lighthouse SEO + hardening delta

AI readiness

Original. Agentic-web rubric

llms.txt presence and quality, robots stance toward GPTBot / ClaudeBot / PerplexityBot / Google-Extended, JSON-LD entity graph, semantic HTML, and whether the page is readable without JavaScript — because most AI crawlers don't run it.

No incumbent — original

Visual design

CV metrics + owned model

A computer-vision sweep (colour discipline, harmony, whitespace, edge density) blended with an owned vision model that scores typography, hierarchy and rhythm. Trained by distilling a frontier multimodal judge into a small head we serve ourselves.

Owned model · no LLM at audit time

Conversion clarity

DOM heuristics + visual model

Headline clarity, CTA dominance, supporting copy, social proof markers, contact path, pricing transparency, primary-action distance from the fold. Conversion readiness — what you control — not promised lift.

Original · no incumbent

Security

HTTP Observatory rubric

CSP class (strong / ok / weak / none), HSTS preload, X-Frame-Options or frame-ancestors, X-Content-Type-Options, Referrer-Policy, Subresource Integrity, cookie flags, mixed content, Permissions-Policy and version-disclosure leaks. Deterministic to the byte.

Deterministic · cited rubric

Architecture

Three layers, in this order.

Each layer hands its evidence to the next. Cheap signals first, expensive judgements last — and the cheaper layers never get overruled silently.

  1. Layer 1

    Fetch + render

    Playwright-driven headless Chromium, a strict 25-second cap, full-page screenshot at 1440×900, post-JS DOM extraction. SSRF-guarded against internal hosts. Bypasses strict CSPs only to inject our auditor scripts — never to scrape.

    Cheap · deterministic

  2. Layer 2

    Lighthouse + CV + DOM checks

    Lighthouse runs 3× sequentially (concurrent Chrome instances would inflate variability — the exact thing the median defeats). Computer-vision metrics, axe-core, structured-data parse, header analysis. Everything that can be measured without judgement.

    Measurable · reproducible

  3. Layer 3

    Owned visual judge

    A small neural head on top of a frozen multimodal backbone, distilled from a frontier judge. One embedding pass per screenshot, served from our own infrastructure. No LLM call at audit time once trained — your screenshot is not shipped to any third party.

    Judgement · owned

Novelty

What no other free audit does.

Genre-aware

Eleven site genres, eleven weight profiles.

Before we score, we classify. Utility, e-commerce, SaaS, marketing, docs, blog, portfolio, government, community, creative (WebGL / 3D), web-app. Each gets its own pillar weighting — a portfolio shouldn't be judged on conversion, a government service shouldn't be judged on hero CTA. Mis-weighting is what made every prior audit tool feel arbitrary.

Owned model

A visual judge we trained ourselves.

Distilled, served on our infrastructure. Other tools that score visual design either skip it or wrap a third-party LLM at request time. Ours runs locally, costs us pennies per audit, returns the same number for the same image, and gets sharper with every batch we train on.

Agentic-readiness

The first audit that grades you for ChatGPT, Claude, Perplexity.

AI crawlers read your site differently. They don't run JavaScript. They favour Schema.org entity graphs. They check robots.txt for opt-in opt-out signals like GPTBot and ClaudeBot. They look for llms.txt. We grade for all of that — most tools don't even know it exists.

Reproducible

A signature on every report.

We hash the sorted set of findings into a SHA-256. Same input — same hash, every time. Drift in our scoring shows up immediately, in our own evals and in yours. Lighthouse's documented run-to-run variability is exactly what we built around.

What we won't do

The audit-tool clichés we refused.

  • ×
    Fake 0s and fake 100s. Scoring saturates with a curve. A genuinely broken site lands in the 30s; a small slip never lands at 100. Inflated extremes are how trust dies.
  • ×
    A different score every run. Same input, same SHA-256 signature, same number — every time. We surface the hash in every report so you can prove it to yourself.
  • ×
    “There are 347 errors!” We report severity-weighted findings, ranked. The first three on the list are the three that move the score most. No panic copy.
  • ×
    A free score, paywalled fixes. Free runs are uncapped on your own deploy. We don't gate the evidence behind an upgrade. If you want help fixing, that conversation is opt-in.
The engine

Lighthouse where it's right. Our own where it isn't.

Google's Lighthouse is the best free measurement of performance and crawler-readiness on the web. We don't re-invent it. Performance, accessibility, and SEO inherit from Lighthouse's category scores — run 3× and median-taken to defeat its documented variability.

Then Fanalis adds the layers Lighthouse doesn't cover: the visual judge, the agentic-readiness audit, conversion clarity, and a hardened security pillar modelled on Mozilla's HTTP Observatory.

See the full scoring rubric →
Stack at a glance
  • RenderPlaywright + headless Chromium, 1440×900, 25 s cap
  • PerfLighthouse 12, 3 runs, median by performance score
  • A11yaxe-core 4.10 (WCAG 2.2) + skip-link / landmark hardening
  • CVPixel sweep: palette, harmony, whitespace, edge density
  • Vision modelFrozen multimodal backbone + distilled head, served on our infra
  • SecurityHTTP Observatory rubric, deterministic
  • SignatureSHA-256 over sorted finding IDs

All inference runs on our own infrastructure. No third-party scoring service. No prompt sent to an LLM at audit time once the model's trained.

Read the method. Then run the audit.
Below 70 — talk to us.

Three free with an account, then ₹5 each. Site below 70? That's where the Catalyst engagement starts.

Run audit →See Catalyst