ORIGINAL RESEARCH · 18 AUGUST 2026

We scanned a hundred of the web's most-used sites. The median scored 8 out of 21.

Everyone has an opinion about whether the web is ready for AI agents. Nobody had measured it. So we did: 100 domains across ten sectors, 21 checks each, 2,100 individual measurements, all taken from the outside over plain HTTP exactly as an agent would see them.

The short version: no site in the sample earned an A, and none earned a B. Sixty-two were graded F, and the best score anyone reached was 15 of 21. Every single one of the hundred failed all three commerce checks. And the single most agent-specific surface on the list — a valid agent card — was published by exactly one domain out of a hundred.

8 / 21
median score across 100 domains
15
highest score anyone reached
62%
graded F, scoring under 9
1%
publish a valid agent card

Correction — 18 August 2026

This page was first published earlier today with figures produced by a version of our scanner that contained three defects. A code review found them the same day, we fixed them, re-ran the entire hundred-site sample against the corrected scanner, and replaced every number below.

What was wrong, and what it did to the results:

  • Soft-404s counted as documents. A site whose catch-all route returns its HTML shell with HTTP 200 was credited with serving llms.txt, ucp, pricing, x402 and api-catalog. This is what produced the original “3% publish commerce surfaces” figure. The true number is zero.
  • The size check could not fail. An oversize homepage returned early in a way that left the measured length at zero, so it passed. The pass rate for that check was reported as 81%; it is actually 48%.
  • HSTS accepted max-age=0, the header that switches HSTS off.

The headline median is unchanged at 8 of 21. What changed is the top of the distribution — the best score falls from 17 to 15, and the one B grade disappears — and the commerce finding, which gets stronger rather than weaker. The earlier figures overstated how ready the web is. The scanner is the same one anyone can run, so the corrected numbers are reproducible today.

The distribution

Scores cluster hard between 6 and 10 and then stop. The right-hand third of the scale is empty: not one site in the sample scored 16 or above, and only five reached 14. There is no established top of this market yet, which is the whole point — the bar is currently low enough to clear with a week of unglamorous work.

1035336715691314984412314151821median 8moltlinestudio.com 21agent-readiness score, out of 21
For calibration: moltlinestudio.com scores 21 of 21, which is where the green line sits. We are not claiming that is impressive engineering — it is a few configuration files and some JSON. We are pointing out that a hundred much larger organisations have not done it.

Six things the data says

Transport is solved. Meaning is not.

98% redirect HTTP to HTTPS and 76% set HSTS — the security basics have won. But only 39% publish any structured data at all, and 36% identify their own organisation in it. The web has learned to be secure and has not yet learned to be legible.

Agent commerce is a rounding error

Zero of a hundred publish a machine-readable catalogue. Zero publish machine-readable pricing. Zero expose any payment challenge. This is not a rounding error or a small sample artefact — it is every site we measured. If an agent is told to compare vendors and buy, almost nothing on the public web can complete that sentence.

llms.txt has quietly arrived

46% now serve one — far higher than we expected for a convention barely two years old, and higher than the 39% who manage JSON-LD. It is the fastest-adopted agent-facing surface we measured, presumably because it is a text file you can write in five minutes.

The AI-native sites are only slightly better

AI developer tools score best of any sector at a median of 9.5, and the frameworks selling agent infrastructure still average under 10 out of 21. Companies whose product is agents are, on average, not much more readable by agents than a mid-market SaaS.

Documentation sites score worst

A median of 5.5 — the lowest of all ten sectors. The corpus agents most want to read is the corpus least prepared to be read programmatically. Reference sites optimise for the human reader and stop there.

Serving JSON is not the same as declaring it

Only 11% send a correct JSON content type on their discovery documents and only 14% send CORS headers. Plenty of sites publish machine-readable files and then serve them in a way a strict client refuses to parse. These are one-line fixes producing measurable movement.

Every check, and how many passed it

Ordered by pass rate. The top of this table is the settled web; the bottom is the part nobody has built yet. The gap between 98% and 1% is the entire opportunity.

CheckCategoryPassed
HTTP redirects to HTTPStrust98%
robots.txt does not blanket-block AI agentsdiscovery76%
A client with no User-Agent is not blockedaccess74%
HSTS is settrust70%
Title and meta description are present and sanecontent60%
/sitemap.xml is present and parsesdiscovery52%
/.well-known/security.txt names a contacttrust49%
Homepage answers quickly and is not enormousaccess48%
/llms.txt is publisheddiscovery46%
Homepage carries JSON-LD structured datacontent39%
Baseline security headers are settrust38%
Structured data identifies the organisationcontent36%
A markdown twin of the homepage existscontent21%
Discovery documents send CORS headersaccess14%
/.well-known/api-catalog is serveddiscovery11%
JSON documents are served as JSONaccess11%
An OpenAPI description is reachablediscovery5%
/.well-known/agent-card.json is validdiscovery1%
A machine-readable catalogue is publishedcommerce0%
An agent payment challenge is discoverablecommerce0%
Pricing is machine-readablecommerce0%

By sector

Ten sectors, ten domains each. The named site is the highest scorer in that sector — the only per-domain attribution in this report, and deliberately so. Publishing a league table of failures would be neither useful to the operators nor defensible for us.

SectorMedianMeanBest in sector
AI developer tools and agent frameworks9.59.815 (crewai.com)
Developer platforms9.510.015 (render.com)
MCP ecosystem and automation9.09.415 (apify.com)
SaaS and productivity9.08.814 (monday.com)
Small-business software8.07.110 (typeform.com)
AI model and assistant providers7.57.19 (huggingface.co)
E-commerce and retail6.56.111 (shopify.com)
Cloud and infrastructure6.56.210 (heroku.com)
Documentation and reference5.55.69 (pypi.org)
News and publishing5.04.88 (bbc.co.uk)

Category pass rates tell the same story more bluntly: trust 64%, machine-readable content 39%, access hygiene 37%, discovery 32%, commerce 0%.

Method

All of it is reproducible, and we would rather you checked than believed us.

Limitations, stated plainly. A hundred domains is enough to show a shape and not enough for tight confidence intervals on any single sector of ten. The sample skews technical and Western, so it understates how bad the long tail probably is. JavaScript-rendered content is invisible to this method by design, because it is invisible to most agents too — but that means a site whose structured data is injected client-side is scored more harshly than a human visitor would judge it. And we sell an audit service, which gives us an obvious interest in the answer being “most sites are failing”. That is exactly why the tool is free, the method is above, and the data is below.

Get the data

One row per domain: sector, score, total and grade. Individual domains are not named, for the reason given above. CC BY 4.0 — use it, cite it, argue with it.

Download the dataset (CSV)   Score your own domain

If you re-run part of this and get a different answer, we would like to know: support@moltlinestudio.com. Corrections get published.

What we would do about it

If you run one of the ninety-five sites that scored under 14, the encouraging part is that the cheapest fixes are also the highest-scoring ones.

An afternoon

Write an llms.txt. Add Content-Type: application/json and Access-Control-Allow-Origin: * to whatever machine-readable files you already serve. Add nosniff and a Referrer-Policy. That is four points for four configuration lines and one text file.

A week

Publish JSON-LD that names your organisation, serve an OpenAPI description, add an agent card and an api-catalog, and offer a markdown twin of your key pages. This is where the sample thins out fast — and therefore where the differentiation is.

Ahead of the field

Machine-readable catalogue, machine-readable pricing, a payment challenge. None of the hundred sites we measured has any of this. Whatever you think of agent commerce, being reachable by it costs very little today and cannot be retrofitted quickly once it matters.

We documented every undocumented shape we hit taking our own domain to 21 of 21, so most of this is copyable. If you would rather it were simply done, that is the audit and implementation service.