The web wasn't built
for AI agents.
Here's the proof.
We audited 300 pages across 100 top SaaS products and scored each on agentability — how usable they are for AI agents making decisions, filling forms, and navigating flows on your behalf.
Confidence: Scores are averages across 3 page types (homepage, pricing, docs). LLM-judged checks carry ±2–3 point variance across runs. Data source: Agentable audit engine · Supabase Prady project · June 2026. ↓ Machine-readable JSON
across 3 pages each
across all products
the ceiling no one breaches
Key Findings
n8n tops the chart
The automation platform built for agents is also the most agent-readable. Developer-first products dominate the top 14. Zapier (#2, 50.5) and Heroku (#3, 48.6) follow. If you build with agents in mind, agents can use you.
OpenAI scored below Anthropic
The companies whose products depend on agent-readable content have the least agent-readable sites. OpenAI: 28.0 (#93). Anthropic: 33.0 (#83). Hugging Face: 31.7 (#85). The AI builders don't practice what the ecosystem needs.
Docs score worse than homepages
Avg homepage: 38.3. Avg docs: 36.8. Docs are where agents spend the most time — finding API references, understanding limits, learning flows. Companies invest in polished marketing and let docs slip.
QuickBooks is Agent-Blind
Salesforce (#92), Tableau (#99, 17.2), Adobe Creative Cloud (#98, 18.4), QuickBooks (#100, 14.2). The enterprise giants with the most to lose from agent adoption have the worst agent readiness. The gap is strategic risk.
Where the web fails agents
Average scores across all 100 products, per principle. The web excels at machine readability (HTML parsing) and collapses on transparency and shadow UI.
100 products. One ranking.
Sorted by average agentability score across homepage, pricing, and docs pages. Click any row to see the full 8-principle breakdown.
100 products shown
| # | Product | Score | Score bar | Tier | Expand |
|---|
To fix this: clear the search box using the × button, or select All from the tier filter above. You can also press Escape in the search box to reset.
Methodology
The Agentability Score is a composite of 8 principles drawn from Agent Factors Engineering (AFE) — the agent-era successor to Human Factors Engineering.
3 pages per product
Homepage, pricing page, and docs page. Each page type reveals different failure modes. Docs score lowest overall.
Deterministic + LLM checks
DOM/ARIA analysis via Playwright for structural checks. LLM-judged scoring for content quality, chunking, and transparency.
8 principles, weighted
Each principle contributes to the composite 0–100 score. Scores averaged across all 3 pages to produce the product ranking.
Tier classification
Agent-Ready ≥45 · Developing 35–44 · Lagging 20–34 · Agent-Blind <20. Based on practical agent usability thresholds.
Confidence & Limitations
Score confidence
LLM-judged checks (17 of 29) carry ±2–3 point variance across runs. Deterministic checks (12 of 29) are stable. Treat scores as directionally accurate — a product at 44 is meaningfully different from one at 28, but 44 vs 45 is within noise.
What's not measured
v0 measures static page state only — no login flows, no error injection, no multi-step task completion. A product with an excellent API but a weak marketing site may score lower than its true agentability. These gaps are planned for M3–M5.
Data source & traceability
All audit results are stored in Supabase (Prady project, eu-central-1). Each product score is the mean of 3 individual page audits. Individual page audit reports are available via the Agentable audit tool.
Disputing a score
If a product's score seems wrong, re-audit it directly at apps.icuboid.in/agentable. Enter the homepage, pricing, or docs URL to get a fresh report. Scores update as products improve their pages.
The 8 Principles
| Principle | What it measures | Global avg | Top scorer |
|---|---|---|---|
| machine_readability | Semantic HTML, structured data, parsability for agents | 0.68 | PlanetScale (0.96) |
| chunking | Content organization, navigability, headings structure | 0.51 | n8n (0.62) |
| control | Labeled interactive elements, keyboard + agent navigability | 0.47 | Mailchimp / Confluence (0.75) |
| status | State communication, feedback signals, progress indicators | 0.44 | Linear (0.78) |
| defaults | Sensible defaults, minimal required agent configuration | 0.37 | Amplitude (0.71) |
| handoffs | Clear CTAs, links, escalation paths between states | 0.38 | Lever (0.56) |
| shadow_ui | Absence of hidden/obfuscated patterns that trap agents | 0.19 | Square (0.50) |
| transparency | Rate limits, pricing clarity, agent-specific documentation | 0.05 | Lever (0.29) |