On 7 September 2026 we scanned 40 well-known SaaS websites the way an AI agent sees them. Agents could read almost all of them (median 99 out of 100) but act on far fewer (median 47). Only 3% had a form an agent could submit, and 30% offered a direct route to a human.
Published · Updated · Anonymous: no site is named, and there is no league table.
The results
Measure
40 sites
Median readable score
Can an agent get your content and understand it.
99/100
Median actionable score
Can an agent do something on your site without a human.
47/100
Median trusted score
Is the response clean enough to rely on.
88/100
Median overall score
All 19 checks, weighted.
75/100
Form an agent can submit
A form with a real action target that works without the page's JavaScript.
3%
Direct route to a human
A real mailto: or tel: link, not only a widget.
30%
Linkable booking path
A scheduling URL in the HTML that an agent can open or hand back.
57%
MCP endpoint
Tools an agent can call to operate the product.
5%
OpenAPI document
A machine-readable API description at a discoverable path.
25%
Agent policy file
agents.md or an equivalent stating what automated clients may do.
15%
llms.txt
A curated map of the pages worth reading.
83%
Content needs JavaScript to read
The server sends an effectively empty page.
3%
Blocked to agents
A plain HTTP client gets a challenge or an error instead of the page.
3%
Scores are medians across the panel. Percentages are the share of the 40 sites that passed each check. The same numbers as a file: benchmark.csv.
What an agent can do on these sites
Each row links to what the check means and how to fix it.
How this differs from other agent-readiness studies
Cloudflare (17 April 2026) scanned the 200,000 most visited domains for agent protocols: robots.txt rules, markdown content negotiation, MCP server cards, API catalogs. It found 3.9% of sites supported markdown content negotiation. It does not look at lead forms, booking links or reaching a person. Read the Cloudflare study.
Ora (22 April 2026) scanned 84,271 web products across 110 checks and found a median score of 38 out of 100, summed up as "Agents can read. They can't act." It measures the infrastructure that would let an agent act, not whether a form or a booking link actually works. Read the Ora study.
Siteline (18 June 2026) sent a Claude agent to 100 B2B software products, 534 runs in all, to find their pricing. Nearly one run in three hit an error searching or fetching the site, and only 65% of plans surfaced pricing directly. It tests pricing, not lead capture. Read the Siteline study.
This benchmark asks the question those leave open. Once an agent has read a SaaS website, can it turn into a lead? Can it submit the enquiry form, book the meeting, or reach a person?
Method
The panel is 40 well-known SaaS websites, fixed before scanning, from developer tools to payroll. They are not named here: the point is the pattern, not a ranking.
Each site was scanned on 7 September 2026 with the same scanner you can run on this site. It makes around twelve plain HTTP requests, identifies itself honestly, and executes no JavaScript, which is close to how many agent runtimes fetch a page.
Scores are out of 100 across three groups (Readable, Actionable and Trusted) built from 19 weighted checks. A pass means the signal is present, for example a form with a real action target. It does not mean an agent completed the task end to end.
The limits are real. 40 large, well-known companies are not representative of all SaaS, and the scan reads the homepage and a set of standard paths, so a capability deeper in a site can be missed. The benchmark is re-run monthly, and the date at the top of this page is the latest run.
Questions
What share of SaaS websites have a form an AI agent can submit?
In our 7 September 2026 scan of 40 well-known SaaS websites, 3% had a form with a real action target that an agent could submit without running the page's JavaScript.
Can AI agents book a demo on SaaS websites?
57% of the sites put a linkable scheduling URL in their HTML, which an agent can open or hand back to the person it works for. The rest had no scheduling link an agent could find in the page.
Does having an llms.txt make a site usable by agents?
No. 83% of the panel published an llms.txt, but only 3% had a form an agent could submit and 30% a direct route to a human. Being readable and being operable are different problems.
How do I check my own site?
Run the free scan. It uses the same 19 checks as this benchmark and returns a ranked list of fixes, each with a prompt you can hand to a coding agent.