State of agent readiness

More than a quarter of the sites turned the agent away

What the readable pages get wrong

Of the 73 home pages that could be read, 4 scored Excellent (85 or more), 36 Good, 21 Fair, and 12 Poor (under 50). The highest score was 100 and the lowest 18. These are the checks that failed most often.

1. Search and training crawlers decided separately

   56 of 73 failed (77%)

   Being cited and being trained on are one decision instead of two, so the site either gives away more than intended or blocks itself out of AI answers by accident.

2. JSON-LD in the server HTML

   30 of 73 failed (41%)

   Every fact about the page has to be inferred from prose. An agent guesses at who wrote it and when, and guesses wrong on pages that do not follow a familiar shape.

3. Structured data says what the page is

   30 of 73 failed (41%)

   The markup describes the site and the organisation but never this page, so nothing carries its headline, its dates or its author.

4. Open Graph tags with an image that resolves

   23 of 73 failed (32%)

   A link to this page becomes a bare URL wherever it is shared or previewed, including inside chat interfaces.

5. One H1 per page

   18 of 73 failed (25%)

   Once markup is stripped, nothing says what the page is called. An agent has to guess the title from the body text, and often guesses the navigation.

6. Sitemap lastmod dates look real

   16 of 73 failed (22%)

   Crawlers cannot tell what changed, so they either recrawl everything or trust none of it. Stamped-looking dates train them to ignore the field.

7. Organization and contact information

   15 of 73 failed (21%)

   Who publishes this and how to reach them is not stated, which is part of what anyone weighs before relying on a page.

8. Semantic landmarks mark the content

   13 of 73 failed (18%)

   Nothing separates the content from the navigation and footer, so an extractor keeps the menu and the boilerplate alongside the text.

By category

A median over three or five sites is a description of those sites, not of an industry. Read the category rows with the “readable” column in view.

All 100 sites

Each score is for the home page as it was served to the scanner on 2026-10-11. Scores change when pages change. Anyone can run the same scan on any of these.

How this was measured, and what it cannot tell you

• The sites. We picked 100, ten in each of ten categories, because people know them. It is not a random sample and not a ranking of the web.
• One page each. Only the home page was scanned, once, on 2026-10-11, from one location. Sites that failed for a transport reason were tried a second time.
• The scanner. The Agent Analyzer, check version 2026.10.1. It fetches a page without running JavaScript and identifies itself as AgentExperiencesScanner. It obeys robots.txt.
• The score. A weighted total of automated checks on access, structure, metadata, structured data, and trust signals. It measures how easily an agent can read and attribute a page. It does not measure whether AI systems cite the site, and a high score does not predict that they will.
• Size limit. Two home pages were larger than the scanner’s 5 MB limit and count as not readable. That is our limit, though an agent with a fixed context budget faces the same problem.