I scored 91 websites against a 40-check design contract. Zero got an A.

We scored 91 websites across 5 industries — fintech, AI/ML, SaaS, design tools, and developer platforms — against a 40-check design system contract. The contract checks tokens, typography, motion, accessibility, and interaction feel against real design-system standards. Not one site scored an A. Not one scored a B. Not one scored a C.

The best score in the entire dataset was 74% — a D.

This is the capstone of a 7-part series. The prior articles scored individual batches. This one synthesizes all of them and asks the question that matters: if the design industry can't pass a design contract, what is the design industry actually selling?

The full leaderboard

Rank	Site	Score	Grade	Industry
1	Stripe	74.0%	D	Fintech
2	Vercel AI SDK	73.9%	D	AI infra
3	Vercel	67.4%	D	AI infra
4	Linear	66.0%	D	SaaS
5	Asana	64.0%	D	SaaS
6	Modal	63.0%	D	AI infra
7	Figma	62.5%	D	Design tools
8	Wise	62.0%	D	Fintech
9	Replicate	60.4%	D	AI infra
10	Coda	60.4%	D	SaaS
11	Anthropic	60.0%	D	AI models
12	ClickUp	60.0%	D	SaaS
13	Monday	60.0%	D	SaaS
14	Plaid	60.0%	D	Fintech
15	Airtable	58.3%	F	SaaS
16	LlamaIndex	58.3%	F	AI models
17	Mistral	56.5%	F	AI models
18	Hugging Face	56.2%	F	AI models
19	Cohere	56.0%	F	AI models
20	Notion	46.0%	F	SaaS
21	Pinecone	46.0%	F	AI infra
22	Brex	45.7%	F	Fintech
23	Miro	44.0%	F	Design tools
24	Whimsical	44.0%	F	Design tools
25	Pitch	43.5%	F	Design tools
26	Cash App	41.3%	F	Fintech
—	Mercury	BLOCKED	—	Fintech (HTTP 429)
—	Coinbase	BLOCKED	—	Fintech (HTTP 403)
—	Revolut	BLOCKED	—	Fintech (HTTP 403)
—	Chime	BLOCKED	—	Fintech (HTTP 403)
—	SoFi	BLOCKED	—	Fintech (HTTP 403)
—	OpenAI	BLOCKED	—	AI models (HTTP 403)
—	Perplexity	BLOCKED	—	AI models (HTTP 403)

Plus 57 sites scored in prior batches (30 general + 16 dev tools + 11 fintech), all scoring D or F. Full data at designesy.org/leaderboard.

The methodology

The Designesy scoring engine fetches each site's live HTML and CSS, extracts design tokens from :root, and runs 40 deterministic checks against a design system contract. The contract covers 7 categories:

• Tokens — semantic color roles (--ink, --paper, --signal), luminance and saturation validation
• Cadence — text-wrap: balance + pretty, tabular-nums, font-synthesis: none, text-decoration-skip-ink: auto, line-height by role, rem-based scale
• Motion — prefers-reduced-motion support, no transition:all, will-change restricted to transform/opacity, duration tokens
• Takt — press scales above 0.95 floor, concentric border-radius sets, stagger enter delays, soften exit transforms
• Poise — fine-pointer hover guards, press settle scales, keyboard-affordance signals
• Accessibility — WCAG AA 4.5:1 button contrast, :focus-visible rings, contrast on paper surfaces
• Performance — Core Web Vitals (LCP, INP, CLS)

Every check is deterministic. No subjective scoring. No human judgment. The engine fetches the CSS, runs the checks, and returns a score. You can run it yourself on any URL.

Finding 1: The design tools paradox

The companies that sell design tools score the same as everyone else.

• Figma (62.5%) — the design tool the industry uses — scores between Asana (64%) and Coda (60%). It's indistinguishable from project management tools. No text-wrap, no tabular-nums, no ::selection rule on its marketing site.
• Webflow (52.0%) — the no-code design platform — scores below Airtable (58%). Press scales of 0.66 and 0.33 — buttons compress to a third of their size on tap.
• Pitch (43.5%) — a company literally named after design presentations — scores below Cash App (41.3%), a fintech site with zero design tokens. No :focus-visible rules. No prefers-reduced-motion. No duration tokens.
• Miro (44.0%) — a visual workspace — has --ink and --paper set to the same color. 0.00:1 contrast. The text and the background are identical.

This is the sharpest finding in the dataset. The companies selling "beautiful design" can't pass a design contract. Their marketing says one thing. Their CSS says another.

Finding 2: Infrastructure beats models

In the AI/ML sector, the infrastructure companies outscore the model companies by a clear margin:

• AI infrastructure average: 62.1% (Vercel AI SDK, Vercel, Modal, Replicate, Pinecone)
• AI model companies average: 56.5% (Anthropic, LlamaIndex, Mistral, Hugging Face, Cohere, Together AI)

The 5.6-point gap isn't accidental. Infrastructure companies build developer tools — dashboards, docs, deployment UIs. Their product is the interface, so they invest in design engineering. Model companies compete on benchmarks and API latency. The website is marketing, not product.

The scores reflect where design investment actually goes.

Finding 3: Token misconfiguration is universal

The most common failure across all 91 sites isn't a missing feature. It's a misconfigured token.

• Whimsical: --ink luminance 255/255 — the "dark text" token is pure white
• Miro: --ink and --paper are the same color — 0.00:1 contrast
• ClickUp: --ink on --paper = 0.00:1 — same color
• Linear: --ink luminance 182/255 — "dark text" is medium-gray; --paper luminance 19/255 — "light surface" is dark (inverted dark mode with light-mode token names)
• Cohere: --ink luminance 165/255 — same issue as Linear
• Brex: --ink luminance 191/255 — "dark text" is light gray
• Pinecone: --paper: #fff but contract expects #010102 — inverted token system

These are root-level mistakes. A misconfigured --ink token cascades through every component that uses it. Every text color, every button label, every heading inherits the error. And because the token is named --ink, the team assumes it's correct.

Finding 4: 7 sites block verification entirely

Seven sites — 8% of the dataset — have WAF/Cloudflare configurations that reject the scoring engine's HTTP request before the CSS arrives:

• Fintech: Mercury (429), Coinbase (403), Revolut (403), Chime (403), SoFi (403) — 5 of 12 fintech sites (42%)
• AI/ML: OpenAI (403), Perplexity (403) — 2 of 12 AI sites (17%)

We tried root domains, www variants, and blog subdomain paths. All blocked at the WAF level, not the path level.

Transparency is a design quality signal. Sites that let you inspect their CSS are accountable to their design choices. Sites that block inspection aren't — not because their design is necessarily bad, but because they've made it impossible to know.

Finding 5: No-code builders produce lower scores

Sites built on Framer (Pitch, Miro) average 43.75%. Hand-coded sites average 57.3%. That's a 13.6-point gap.

Framer generates CSS from visual designs rather than from a token system. The tell is var(--framer-will-change-override,transform) in the will-change property — a Framer-specific variable. Framer-built sites in this dataset share the same gaps:

• No semantic --ink/--paper/--signal tokens
• No :focus-visible rules (Framer doesn't generate them by default)
• No prefers-reduced-motion query (Framer animations are JS-driven)
• No duration tokens (Framer uses its own animation timing)

This isn't a Framer problem specifically. It's a no-code-visual-builder pattern. When you generate CSS from a visual canvas, you get the canvas's defaults. And those defaults don't include a design contract.

Finding 6: Two CSS properties have near-zero adoption

After scoring 91 sites, only 1 site passed each of these Cadence resolved tensions:

• text-decoration-skip-ink: auto — prevents underlines from crossing through letter descenders. Passed by: Coda. 1 of 84 scored sites. 1.2% adoption.
• text-underline-position: from-font — positions underlines based on the font's own metrics. Passed by: ClickUp. 1 of 84 scored sites. 1.2% adoption.

These are CSS properties that improve typography with zero design cost. One line of CSS each. No performance impact. No layout change. They just make text look better. And 98.8% of sites don't use them.

The AI slop connection

This dataset lands in the middle of a live conversation. In 2026, the design industry is debating what's being called "AI design slop" — the widespread reliance on AI tools to generate UIs that are technically correct but lack specific identity.

A June 2026 Built In article framed it this way: "The baseline for acceptable design has quietly shifted to whatever AI produces. Because AI produces work that is competent enough to not fail any individual review, the shift is invisible from inside any single sprint. You only see it when you look at the market as a whole."

Our 91-site dataset looks at the market as a whole. And what it shows is that the baseline was already low before AI got involved. The design tools companies (Figma, Webflow, Pitch) score the same as everyone else. The infrastructure companies (Vercel, Stripe) score only slightly better. The model companies (Anthropic, Cohere, Mistral) score worse. And 7 sites block verification entirely.

A CHI 2026 study found that LLM agents hit ~95% design-system compliance at best when reading from a live component registry. At machine speed, the 5% that drifts becomes the screen the next agent copies. If the design industry's own websites can't pass a 40-check contract, what happens when AI agents generate UIs against those same standards?

The answer is already in the data. The 91 sites we scored represent the best of the web — successful companies with real engineering teams. And they average 56.7%. If that's the ceiling for human-built design, 95% LLM compliance isn't a problem. It's an improvement.

What this means

The design industry talks about craft. It sells design systems, design tokens, design tools. It markets beautiful product experiences. But when you run the checks — the same checks that a design system contract requires — the industry's own work doesn't pass.

This isn't about shaming individual sites. Stripe at 74% is doing better than most. Linear at 66% has excellent interaction feel. Vercel AI SDK at 73.9% has the strongest Cadence implementation we've seen. These are real engineering achievements.

It's about the gap between what the industry says and what the industry does.

The design contract exists. The checks are deterministic. The engine is open. Any site can be scored in under 60 seconds. The data is right there. And the data says: the design industry is selling compliance it can't pass.

The fix isn't more design tools. It's not more AI-generated UIs. It's the check itself. As the Akhmedov article argued: the design system didn't fail. Every token resolved. Every color was on-brand. The library was fine. The check failed. We didn't have one.

Now you do.

---

Scored with Designesy — a 40-check design system verification engine. The contract, the engine, and all 91 site scores are open. Run your own site.

This is the capstone of a 7-part series: 30 sites scored, 16 dev tools scored, 11 fintech sites scored, 57-site synthesis, 5 blocked fintech sites, 12 AI/ML platforms scored, 12 SaaS platforms scored.