The evidence behind agent-readiness
Interrupt Index scores what an autonomous agent actually encounters on the way to a completed task. This page is the public record of the research those scores rest on. Every figure is drawn from our July 2026 evidence review; where a statistic rests on a named external study, the source appears beside it. We do not publish a number here we cannot stand behind.
Retrieval is how agents meet the web
The single most load-bearing fact in our model: agents from the major assistants meet the web primarily as retrieval fetchers — server-rendered HTML read once, not a browser session that executes JavaScript. Our July 2026 evidence review found that no major AI retrieval fetcher executes JavaScript (a 500M-fetch GPTBot analysis), so client-rendered content is effectively invisible to them500M-fetch GPTBot request analysis. That is why our Machine-Readability Score measures static, server-readable signals and grades the JavaScript-dependence check proven— and it is the motivation for the new group (d) “discovery / agent access” in machine-readability model v3 (see Methodology).
Retrieval fetchers honor robots.txt — proven
The on-demand retrieval UAs (OAI-SearchBot, ChatGPT-User, Claude-User, PerplexityBot, Perplexity-User) demonstrably fetch robots.txtand comply with it: Cloudflare’s August 2025 compliance test showed ChatGPT-User fetching robots.txt and stopping when disallowedCloudflare, Aug 2025. Blocking them makes a site uncitable in AI answers. This is the evidence behind our scored retrieval_bot_access check (group (d)), graded proven, and it is why our own robots.txt names each of these UAs explicitly with allow: / — a review site for agents must be maximally readable by agents.
llms.txt: adoption growing, consumption near zero — emerging
llms.txt publisher adoption grew roughly 8.8× in twelve months, yet a May 2026 log study across 137,000 domains found 97% of llms.txt files received zero requests137k-domain log study, May 2026. The real consumer today is coding agents on developer docs, not general assistants. We grade the llms.txt check emerging rather than proven for exactly this reason — adoption trajectory is real, production consumption by general agents is not. We still score it: emerging standards are cheap to publish, trajectories can flip fast, and our standards-lifecycle policy is to retire a check only when its standard is dead (deprecated by its own steward), not merely unadopted (see Methodology).
Model Context Protocol: real adoption, operator-configured — plausible
MCP is the one agent-integration standard with real adoption — 10k+ public servers, ~97M SDK downloads, and Linux Foundation stewardshipMCP registry + npm download counts. But MCP connections are operator-configured, not auto-discoveredfrom a vendor domain: a well-known manifest isn’t a thing an agent reads on first contact the way it reads robots.txt. That is exactly what plausible means in our grade model — strong mechanism and adoption, but no measured consumption path an autonomous agent walks on its own. Our mcp_server check (group (a)) rewards both a /.well-known/mcp.json manifest and a match in the official MCP registry.
The evidence-grade model: proven / plausible / emerging
Every Machine-Readability check carries a display-only evidence grade — our honesty label for how strong the real-world evidence is that the signal changes agent behavior. Three values, mirroring the T0/T1/T2/D confidence tiers on the interrupt side:
- ProvenMeasured consumption or controlled-experiment effect (retrieval fetchers honoring robots.txt; AI fetchers not executing JavaScript).
- PlausibleStrong mechanism, adoption, or documented failure mode, but no direct measurement (accessibility-tree semantics for browser agents, MCP servers).
- EmergingPublisher adoption growing; no observed consumption by production agents yet (llms.txt is the canonical case).
Grades never weight the score: the four groups stay equal-weight (owner decision), so the grade is accountability, not a fudge factor. The full table lives in the evidence-alignment review, and the user-facing explanation is on the Methodology page.
We retire only dead standards — ai-plugin.json (deprecated 2024)
The one standard we have removed from scoring is ai-plugin.json, the OpenAI Plugins manifest, deprecated by OpenAI in 2024 (see Methodology for our standards-lifecycle policy). A review site scoring a dead standard would be exactly the failure mode these grades exist to prevent. Emerging standards with growing adoption (llms.txt, AGENTS.md, A2A agent cards) stay scored and graded emerging; the bar for removal is steward-deprecation, not merely low consumption.
For the scoring algorithm those evidence grades sit next to, see Methodology; for the neutral standards and external research behind each, see Resources.