Not a black box. Every score has a source.
The score, the research it comes from, the evidence behind each finding, and the code that fixes it.
- 200+ research sources
- 65% deterministic · 35% AI
- axe-core verified
- 9 evaluation categories
One number. Five pillars. Tied to conversion.
The UX Score is a weighted composite of five pillars, built from 9 evaluation categories. Behaviour captures drop-off, visual quality first impressions, accessibility the people you never reach. The formula is public.
Move a pillar and watch the score. These are the weights and the arithmetic a real run uses.
- 58
- 78
- 61
- 72
- 70
58 × 30% + 78 × 20% + 61 × 20% + 72 × 15% + 70 × 15% = 67
Behaviour carries the most weight on the web, because it is where people give up.
Each login is walked at the width it is set to; a mobile app is its own product. Where one product has logins at both widths, desktop carries 60% and mobile 40%. Each run adds to what Corexi knows about the product, so scores sharpen over time.
With no analytics source connected, behaviour is not guessed from a screenshot: the pillar drops out and the other four are renormalised. Every score in the product opens to its arithmetic.
A few of the 200+ sources behind every score.
The base is not frozen: new peer-reviewed research is folded into the engine with every release.
- Nielsen Norman GroupUSABILITY
10 heuristics, still the usability baseline
- WCAG 2.2 · W3CACCESSIBILITY
Accessibility success criteria · 2023
- ISO 9241-110ERGONOMICS
Interaction principles standard
- Baymard InstituteCONVERSION
E-commerce UX, 150k+ hours of testing
- Fitts's LawMOTOR
1954 law → today's 44px tap-target rule
- Miller's LawCOGNITIVE
7±2 memory limit → cognitive-load scoring
Screens, then flows, then one number.
A product is scored by an Autopilot run: it signs in, walks the product like a new user, and every screen it reaches is captured and judged. The number is built in three steps.
- 1 · Screen
Each captured screen is judged on the five pillars, at the size it was walked at. A screen that rendered nothing is not judged at all.
- 2 · Flow
The screens on one path ("create a request", "approve leave") form a flow. A flow is as weak as its weakest screen, minus the friction met walking it: actions that changed nothing and moments it had to hand over to a person. The agent's own wandering is reported as task numbers, not subtracted.
- 3 · Product
Screen scores roll up into the product score and fuse with connected behavioural data. The next run compares against this one, screen by screen.
What a finding can and cannot do.
- A finding keeps its identity across runs by the rule it cites and the screen it sits on, even when the wording changes. Fixed once, it does not come back as "new".
- A finding is resolved only after two runs in a row looked at its screen and did not find it. One miss marks it; the second closes it.
- A run that could analyse fewer than half of the screens it captured saves no score. It is reported as a failed run, and never becomes a regression.
- A finding seen again takes the latest run's wording, evidence and severity; its history stays.
- Rules about the whole product (unnamed controls, colour as the only signal) are one finding with the screens listed, not one per screen.
- Findings are written in English whatever language the product is in, quoting the product's own labels.
9 core categories.
Every screen Autopilot reaches is read against these. Each has measurable criteria from peer-reviewed sources, and each maps to friction that quietly costs conversions. Every accessibility finding is cross-checked by axe-core.
- Visual hierarchy
Is the most important element the first to catch attention? Are headings differentiated? Is the flow logical?
- CTA visibility
Can you find the primary action in under 2 seconds? Is contrast above WCAG AA 4.5:1? Is it sized right?
- Form design
Are labels present and readable? Are error states designed? Are inputs sized for touch and keyboard?
- Spacing and layout
Is white space consistent? Is there a grid? Are there cramped or scattered areas?
- Typography
Is body text 16px or more? Is line length 45 to 75 characters? Does weight create hierarchy?
- Colour and contrast
Does text pass WCAG AA? Is the palette consistent? Is colour ever the only carrier of meaning?
- Mobile responsiveness
Are touch targets 44 × 44px or more? Is there horizontal scroll? Is text readable on a phone?
- Cognitive load
Can a first-time user tell what to do? Is navigation clear? Is there decision paralysis?
- Accessibility
Alt text, form labels, keyboard navigation, ARIA, focus indicators. Every claim checked by axe-core.
Five more lenses, for how people think.
WCAG covers compliance. It does not say whether a neurodivergent person can finish signing up. 15 to 20% of people think differently, and each of them is a conversion.
- ADHD friendliness
Visual clutter, animation control, content chunking, clear next-step actions.
- Dyslexia readability
Line length, letter spacing, justified text, all-caps blocks, contrast below 7:1.
- Autism predictability
Consistent navigation, predictable layouts, clear labels, few unexpected changes.
- Sensory sensitivity
Animation intensity, autoplaying media, contrast extremes, visual overwhelm.
- Colour-vision deficiency
Red and green pairs, colour as the only carrier of meaning, problem colour pairs.
Three layers. One score.
Visual analysis says what looks wrong. Behavioural data says what actually costs you. The reasoning layer puts them together, so a finding comes with both.
- Visual
Every screen the walk reaches, read by AI vision across the 9 categories. Deterministic checks (contrast, touch targets, spacing) run beside the AI's reading. What people see.
- Behavioural
Real signals from your analytics and the Corexi snippet: bounce, rage clicks, scroll depth, form abandonment. The moments people give up.
- Reasoning
Checks one against the other. A contrast problem on a screen people leave scores higher than one nobody trips on. Ties every finding to what it costs, and sharpens with every run.
| Source | Window | Signal |
|---|---|---|
| GA4 | 7 days | Sessions, bounce rate, engagement |
| Clarity | 72 hours | Rage clicks, dead clicks, quick-backs |
| Mixpanel | 7 days | Event volume, active days |
| Amplitude | 7 days | Active users, event cadence |
| PostHog | 7 days | Events, active users (deduplicated) |
| Firebase | Auth | Mobile app analytics |
| Hotjar | Auth | Survey and feedback data |
| Corexi snippet | 1 hour | Scroll depth, rage clicks, dead clicks, time on page |
Visual-only runs work from day one. Every source you connect makes the diagnosis sharper.
- Visual AI only~40%
- +1 provider~65%
- +2 or 3 providers85–95%
Verified, not guessed.
Every accessibility finding is cross-checked against axe-core at runtime. Verified findings carry a badge; findings only the AI saw stay flagged. You always know which is which.
The AI found it and axe-core confirmed it against WCAG 2.2: 27 rule IDs checked. The standard for a compliance claim.
The AI saw a likely problem but no deterministic rule confirmed it. Flagged for a person to review, and open about its confidence.
How far an agent gets on its own.
Products are used by agents now, and where an agent stalls, a person stalls too. Autopilot is such an agent, so every run already holds the evidence. The score is arithmetic over the walk, shown beside PX, not folded into it.
- 40Flows reached
Flows the agent finished without help count in full, with a stumble in half, blocked not at all.
- 25Steps against the fewest
Steps taken against the fewest a person who knows the product needs. Wandering costs; straight lines do not.
- 20Actions that worked
Actions that did what their label said, across the whole walk. A click that changed nothing counts against.
- 15No hand-off
Logins, codes and captchas handed to a person, per flow. A saved session avoids most of them.
85+ agent-ready · 70 to 84 mostly · 50 to 69 agents hit friction · below 50 not agent-ready yet. A run that named no flows shows no number.
Privacy first. Always.
The Corexi snippet is optional and cookieless. Anonymous behavioural signals only: no personal data, no session replay, 24-hour retention.
We connect to your providers through read-only APIs and pull aggregated metrics only.
Bounce rates, session depth, engagement patterns. Never identities, emails or browsing histories.
Processing on EU infrastructure. Credentials encrypted at rest (AES-256). Disconnect and delete at any time.
The whole method, step by step: how a run is scored.
See what your UX is costing you.
Give it your product's URL. We set up the login with you once; from then on it goes round on its own, after every release.