How Sensei scoring works
Sensei scores a public page by extracting its structure, evaluating it against three weighted UX layers, and translating those findings into a report a product team can act on.
The framework behind each score is grounded in design psychology, usability heuristics, conversion research, and competitive benchmarking — not ad-hoc rules. Every criterion maps to a documented discipline with real-world weight behind it.
Who built the scoring framework
The criteria behind each score weren't generated by a model. They were developed by a small team with 15+ years of combined experience as design leads and product directors, translated into a framework a language model can apply consistently across thousands of pages.
The AI does the reading and pattern-matching at scale. The judgment of what actually matters — which UX failures are cosmetic and which cost conversions, which heuristics are load-bearing and which are folklore — comes from people who have shipped and led design work professionally, not from the model inventing a rubric on the fly.
Functional UX: 35%
Task clarity, navigation, feedback, accessibility, mobile markup signals, and whether the page helps a visitor understand what to do next.
Aesthetic quality: 35%
Visual confidence, hierarchy, restraint, typography, information density, and whether the page feels calm enough to trust.
UX practices: 30%
Known design psychology and conversion heuristics, including cognitive load, choice pressure, signifiers, social proof, and CTA quality.
Design psychology and heuristics, not opinion
Our framework draws on established design principles, not just stylistic preferences. Hick's Law explains why too many competing calls-to-action increase decision friction. Loss aversion and risk-reversal signals (free trials, guarantees, “cancel anytime”) explain why their absence creates hesitation at the exact moment a visitor is ready to convert. Social proof, cognitive load, anchoring, and recognized usability heuristics work the same way — these principles help explain why a finding matters and how to address it.
This is deliberate. A model can describe what it sees on a page; it can't decide on its own which observations actually matter. That judgment comes from the practitioners who built the framework — the model applies their criteria, it doesn't invent its own.
What we examine
The scanner fetches the public HTML and extracts headings, body copy, CTAs, forms, links, image alt text, accessibility attributes, metadata, and structural landmarks. When a capture includes a screenshot, visual evidence can supplement those markup signals.
What we do not examine
Sensei does not inspect logged-in states, paywalled content, full interaction flows, live analytics, real Core Web Vitals, checkout completion, or private user data.
Mobile assessment
Mobile-responsiveness criteria are assessed from markup signals, viewport metadata, responsive patterns, and screenshots when available - not from a rendered mobile viewport. That keeps the current score honest about what the scanner can observe.
Competitive benchmarking
Sensei's public library puts the framework into context: 200 real websites assessed across fintech, AI, developer tools, e-commerce, healthtech, and other industries. Explore their scores, captured evidence, and UX reviews at getsensei.io/browse.
Use the library to compare examples and see what stronger UX looks like in practice. Scores use the fixed framework and calibration below; they are not percentile rankings against the library. Adding or removing a site does not move your score.
How the final score is calibrated
Applicable criteria are combined with the layer weights above. Sensei keeps the weighted calculation precise, applies a fixed calibration, and rounds once to the whole number shown in the product.
The calibration leaves the lower and middle range unchanged and expands a compressed high end, with a raw 80 mapping to 90. It is fixed and does not depend on which sites are in the public library, so adding or removing another site cannot move your score.
Stability and changes
Same URL + unchanged content + same methodology version should return the same score. Scores change when page content changes, the scoring methodology is versioned, or provider fallback introduces bounded model variance.