Agentic Web CensusCheck your site
EN

Census report · August 2026

The web scores 20.1% in August, a move that only just clears its own margin.

This is August 2026, frozen. The index went up against July 2026, by more than the margin allows us to call a wobble.

What moved it

Poor20.1%

AWCS

What the AWCS is, and how it is measured

Measured on 150,327 pages, 95.9% of which carried the agentic category, with a confidence margin of ±0.14.

▲ +0.2
AWCS over 4 months

15.2%20.1%

Read

What Read measures

Lighthouse builds the accessibility tree an agent would traverse and scores the page pass or fail. The denominator is only the pages the audit could evaluate: a notApplicable or an error is excluded rather than counted as a failure, and the fraction that could be evaluated is published beside it as the coverage.

n=158,168 · coverage 95.3% · margin ±0.13 pp.

53% of the score
Read on its full scale

Weak32.2%

Read: 5 months, up▲ +1.3 pp
Guide

What Guide measures

The llms-txt audit passing: a file is served AND Lighthouse considers it valid. The denominator is every sampled page, so this is a fraction of the whole web and not of the sites that have a file. This is the figure the index uses, because a broken file does not guide an agent.

n=158,165 · coverage 95.3% · margin ±0.13 pp.

27% of the score
Guide on its full scale

Poor14.8%

Guide: 5 months, up▲ +0.9 pp
Act

What Act measures

A count, not a percentage: how many sampled sites register WebMCP tools that an agent can call. It is the audit that defines the Act dimension — the other two WebMCP audits measure how good those tools are, not whether they exist.

n=158,176 · coverage 95.3% · margin ±0.13 pp.

20% of the score
Act on its full scale

Very poor4.6%

Act: 5 months, up▲ +0.2 pp
Reachable

What Reachable measures

Read from the robots.txt of the sampled sites: of the AI agents we track, how many are NOT disallowed. Gradual and not a yes or no — turning away one agent out of seventeen is not the same as turning away all of them. A site with no robots.txt blocks nobody, so it counts as fully reachable: that is a measurement, not a gap. This is the only figure that MULTIPLIES the index instead of adding to it, because it is a precondition: if an agent cannot fetch the page, nothing else about the page matters. One caveat we would rather state than hide: a blanket Disallow under User-agent: * is not counted, because the corpus-wide measurement cannot see it either, and we would rather be consistent than clever.

n=166,024 · coverage 100.0% · margin ±0.13 pp.

multiplies the rest
Reachable on its full scale

Excellent95.5%

Reachable: 5 months, no significant change= −0.1 pp

Every figure of the month

Each one with the sample it was measured on and the margin that comes with it.

FigureAugust 2026Pages
AWCS20.1%150,327
Pages an agent can parse30.9%150,334
Agents that can reach the site95.6%156,733
Sites with a valid llms.txt13.9%150,327
Sites exposing tools6,605150,346
Sites serving an llms.txt34.3%150,327
Sites with an AI crawler policy11.0%156,733
Layout stability87.6%150,295
Forms an agent can fill98.5%6,605
Valid tool schemas98.7%6,605
Google's category score57.5%150,263

The two ends

The five highest and the five lowest technologies of the month, by index.

EndTechnologyAWCSPages
HighestVenmo64.1%293
HighestShop Pay63.9%3,653
HighestAmerican Express63.8%2,798
HighestSeal Subscriptions63.7%108
HighestApple Pay63.2%4,061
LowestCafe240.5%101
LowestShopline1.4%115
LowestAlibaba Cloud CDN1.8%705
LowestIntersection Observer1.8%136
LowestBigCommerce1.9%165

The data behind this report

How this report is written, and which parts a model may touch

figures passed to the model: 31
numbers it may write: only those, or their roundings
if it writes any other: the build FAILS and nothing is published
the contractThe rule is executable and not a promise: lib/prose.ts rejects a number nobody passed in, and thirteen tests hold it there.
Everything on this page is calculated — the figures, the deltas, the tables, the records and this note. The headline, the "what moved it" and the advice were written by claude-opus-5 on 2026-09-24, from the figures above and nothing else: no database access, nothing to compute. The text is stored in the repository and validated again at build time, so a number nobody passed in stops the build instead of reaching this page.Every figure here comes from a sample, not from counting every site — so it carries a margin: how much it could move if we had happened to sample different pages. A figure measured on a thousand pages has a wide margin; one measured on a hundred thousand has a narrow one. It matters for one thing: two figures that differ by less than the wider of their margins are not one above the other, they are tied. That is why nothing here has a position number.

Every figure above, as a file. A report you cannot check is an opinion.

curl -O https://agenticwebcensus.com/report/2026-08.json
shellThe aggregate of August 2026, exactly as the site read it.