Guide
Can an AI agent actually use your website?
Figures measured across 5,293 scanned sites, window ending 2026-08-20
The next customer that visits your site may not be a person. Assistants already research purchases, compare providers, fill carts and book appointments on a buyer's behalf, and every platform is racing to make that ordinary. An agent does not see your design. It reads your markup, your text and your forms, and when those are illegible it does not squint. It leaves and completes the task somewhere it can read.
Where agents fail, measured
The failures are not exotic. They are the same mechanical gaps that already hurt search and accessibility, measured here across 5,293 sites we scanned in the 30 days ending 2026-08-20:
| Finding | Share of sites | What it means for an agent |
|---|---|---|
| Content only appears after scripts run | 77%95% CI 75.8% to 78.1% | The page source carries little readable text, so crawlers that do not run scripts see almost nothing. |
| Page regions not marked semantically | 52.4%95% CI 51% to 53.7% | No header, nav, main or footer landmarks, so machines parse a soup of divs. |
| No structured data | 41.5%95% CI 40.2% to 42.8% | Nothing on the page tells machines what the business is in machine terms. |
| No plain summary for machines to lift | 29.7%95% CI 28.5% to 30.9% | Title or description missing, so an assistant has no clean answer to quote. |
| Form fields without labels | 17.9%95% CI 16.9% to 18.9% | Inputs a screen reader announces as nothing at all. |
| No obvious next step | 43.4%95% CI 42.1% to 44.8% | The page never tells a ready visitor what to do next. |
Share of 5,293 scanned sites raising each finding. Intervals are 95% Wilson intervals. Full tables: benchmarks.
The reading problem comes first
77% of sites serve pages whose text barely exists until scripts run. Some agents run scripts. The crawlers and cheap retrieval passes that feed them mostly do not, and our own 41-day test found AI crawlers never followed a link that only exists in JavaScript. A site in that condition is not agent-unfriendly. It is agent-invisible, which settles the question before it starts.
Then the interpretation problem
An agent that can read you still has to work out what things are. A page built from anonymous divs, with no header, main or footer landmarks, makes an agent infer structure statistically. In our population 52.4% of sites give no semantic landmarks and 41.5% carry no structured data at all. Structured data is the difference between an agent guessing your prices and hours and an agent quoting them.
Then the acting problem
Acting means forms. A form field without a label (17.9% of sites) is a field an agent has to guess the purpose of, and agents guess conservatively: they abandon. The same failures that lock out a screen-reader user lock out an agent, which is why agent readiness and accessibility are substantially the same work billed under two names. And 43.4% of sites never state an obvious next step, which leaves both humans and machines unsure what you want them to do.
What to check on your own site
- View source. Is your real content in the page source before scripts run?
- Landmarks. Does the page use header, nav, main and footer elements?
- Structured data. Does machine-readable markup state your business, offers, prices and hours?
- Forms. Does every input have a label that says what it is?
- The next step. Is there one clear action a first-time visitor, human or not, is invited to take?
The free scan runs the first four mechanically and shows you what an agent-shaped reader actually finds on your pages.
Questions we get
Is this worth fixing before agents are mainstream? Every fix above already pays for itself in search visibility and accessibility today. Agent readiness is the same work arriving with a third payer.
Do I need an API for agents? No. Agents are built to use the same web pages people use. A readable, labelled, structured page is the API.
Can I test this without tooling? Partly. The view-source check and the keyboard-only checkout walk catch the two biggest classes. The rest is what scanners are for.