What the AgentReady experiment showed and did not show
Ora's 190-probe experiment and 1,033 traced runs, read at their real scope.
In short
Ora's lab found that, across 19 configurations of one documentation site, only JavaScript-only answers and 403 responses stopped retrieval. The data is 190 probes on one site plus 1,033 traced runs on 25 sites. It shows what a fetch returns, not whether an agent understands the page or finishes a task.
Two datasets with different scope
The AgentReady launch on 19 October 2026 rests on two public datasets in the agentready-org/standard repository. Ora published them, and Vercel's guide on making sites readable draws on both.
| Dataset | Size | What it can support |
|---|---|---|
| fetchability.csv | 190 probes: one documentation site, 19 configurations, five topics, two clients | Which page set-ups let a fetch return the answer |
| traces.csv | 1,033 runs on 25 sites, four models, two harnesses, June to August 2026 | How agents move around real sites |
The arithmetic is worth doing once. Nineteen configurations, five topics and two clients give 190 probes, so each configuration was tested ten times. Ten fetches on one site is a small sample of one site.
What the experiment found
Ora's lab found (19 Oct 2026) that of 19 configurations, only hiding the answer in JavaScript and returning 403 to agent user agents stopped retrieval. Hidden in JavaScript, plain fetch retrieved the answer 0 of 5 times. Under a 403, both clients got it 0 of 5 times. Ora's lab also found that llms.txt, JSON-LD, sitemaps and markdown mirrors did not change retrieval, and that agents asked for markdown on about 65 percent of fetches without markdown mirrors or negotiation changing success.
These are vendor findings. They are specific and checkable, because the data is public. The practical reading is that an answer present in the first response, and not blocked, is the condition that mattered here. Several practices that sites often adopt did not move this measure.
The trace data is concentrated in a few sites
Vercel's guide notes that six sites account for 691 of the 1,033 traced runs, or 66.9 percent. That is about 115 runs per site for those six, against about 18 for each of the other 19. Any percentage drawn from the traces describes those six sites more than it describes the other 19. Ora's lab found (19 Oct 2026) that 86 percent of llms.txt fetches arrived through a link, not a guessed path. That figure is a share of fetches in a skewed sample.
Ora's lab also found (19 Oct 2026) that agents reached docs pages in 83 percent of traced runs, the homepage in 69 percent, .well-known paths in 23 percent, openapi.json in 21 percent and sitemap.xml in 4 percent. Those shares describe where agents went in these runs. They do not show that a missing file cost a run its answer.
What to do with it, by reader
- Head of web: use the finding as a triage rule. Check that the answer arrives without JavaScript and that agents are not served a 403. Both are cheap to test, and neither needs a platform change to find out.
- Technical marketer: keep your llms.txt and markdown work, but do not promise a retrieval gain from it on this evidence. The spec keeps llms.txt at SHOULD level, and the experiment tested retrieval only.
- Platform engineer: read the CSVs and rerun the probes on your own site. Check what your server returns to agent user agents; Web Bot Auth and Content Signals address agent access in different ways.
What this guide cannot tell you
The probes test whether an answer is present in the returned content. They do not test whether an agent understands it or completes a task. That is the gap a journey run is meant to cover; see what a journey run tells you. The experiment used one documentation site, so it says little about shops, news sites or apps. The trace models and harnesses were those of June to August 2026 and may not match the agent that visits you now. Ora ran the experiment and the scanner, and wrote the spec with Vercel. The data is open for others to check, and the AgentReady.org entry lists its identifiers.
Sources
- Ora, AgentReady launch. https://ora.ai/blog/agentready-launch, read 2026-11-04.
- agentready-org/standard, README. https://github.com/agentready-org/standard/blob/main/README.md, read 2026-11-04.
- Vercel, make your site readable by AI agents. https://vercel.com/kb/guide/make-your-site-readable-by-ai-agents, read 2026-11-04.
- agentready-org/standard, repository. https://github.com/agentready-org/standard, read 2026-11-04.
Last reviewed 2026-11-05