agentability

Agentability · an open experiment on the agentic web · a new episode every day

Can AI agents actually use the web?

We find out in public, every day. An AI producer reads what the world is searching that morning and turns it into ten real errands — what time is kick-off and on which channel, what magnitude was the quake, what does it cost, what actually happened — and a real AI agent attempts them using nothing but plain web requests: no logins, no JavaScript, no human help. Every transcript is published verbatim, wins and failures alike. Alongside the show, the 1186 most-visited sites on the web are scored on how usable they actually are for an agent.

What we found · 1186 sites audited · episode of November 2, 2026

33%
of the 1186 most-visited sites turn AI agents away

12% say no in robots.txt · 23% just wall them out · only 2% do both · 49/100 average · today the agent finished 9/10 errands and hit 5 bot walls

apple+2tvinsider+1service.gov.uk+9fcandorra+7abc.net.au+7weather.gov+1baltimoreravens+3libero.pe+2cbc.ca+7marketbeat+3
9 done1 gave up honestly
Read today's transcripts → See all 1186 site scores

New here? Watch an agent try and fail · look up a site's score · fix your own site · get yours audited, free · take the raw data

no retriesno editingno cherry-pickingread-only http get — no javascript, no logins, no formsagent: deepseek-flashproducer: deepseek-v4-pro + live searchevery transcript published verbatimno retriesno editingno cherry-pickingread-only http get — no javascript, no logins, no formsagent: deepseek-flashproducer: deepseek-v4-pro + live searchevery transcript published verbatim

Where it got interesting

done Jesús de Miguel: what happened and the club's word fcandorra.com +7 more ⚠ 1 bot wall done Pick the best OSAP single-IP source cbc.ca +7 more ⚠ 1 bot wall done Pick the best Apple earnings source marketbeat.com +3 more ⚠ 1 bot wall

The other half: the 1186 most-visited sites, scored

Every week the 1186 most-visited sites on the web are audited against the conventions real AI agents rely on — llms.txt, crawler policy, content you can read without a browser, structured data, MCP. Every site gets a public report with a score and the exact fix for each failed check. When the agent hits a wall in an episode, the index has usually already predicted it.

49/100average readiness score
12%publish llms.txt
27%block at least one AI crawler
33%turn AI agents away — 12% say so in robots.txt, 23% wall them out
3%wouldn't answer at all

Who is ready, and who is not

19 sites answer every check perfectly, so the top of the ranking is a 19-way tie and tells you nothing. The spread, and the bottom, do.

Perfect 100/100 — 19 sites

alchemer.comapify.comatlassian.comcloudflare.comcursor.comdropbox.comelevenlabs.iogoguardian.comhotmart.comkbag.selinktr.eenotion.compaypal.comrazorpay.comsemrush.comsmartsheet.comstripe.comsurveymonkey.comvercel.com

Where agents get stuck — ranks 1182–1186 of 1186

#SiteScoreGradeSignals
1182 quora.com closed by policy 0 Closed by policy
1183 realestate.com.au closed by policy 0 Closed by policy
1184 reddit.com closed by policy 0 Closed by policy
1185 wordreference.com closed by policy 0 Closed by policy
1186 yelp.com closed by policy 0 Closed by policy

Full ranked index of 1186 sites →

Work on one of these sites? Every failed check on your report page has a concrete fix, and scores refresh weekly. Request an audit of any site — it's free and takes one issue.

Why this exists

Agents are the web's newest audience: assistants that read pages, cite sources, and run errands for people. Whether the web actually works for them is an empirical question — so we test it, in public, every week, with verbatim transcripts, reproducible checks, and open data. History accrues weekly (16 snapshots so far).