Report · scanned 16 August 2026
6 of 274 SEO agency sites pass all nine AI-readability checks.
We ran the nine technical checks an AI assistant needs to read a page on 274 European agencies that sell SEO or GEO services. Most are mostly readable: the median score is 91 out of 100. Very few are fully readable, and the most common gap is the cheapest to close.
By Mike Cecconello. Method and every figure below; no agency is named.
Pass rate per check
The nine checks, the weight each carries in the score, and how many of the 274 sites passed. The weights are not evenly spread, and deliberately so: they follow what search practitioners rate as actually affecting whether you appear in Google AI answers, in Zyppy's Google Ranking Factors Expert Survey 2026. Crawl access is the strongest factor there at +2.20; structured data sits near the bottom at +0.80; an llms.txt file scores +0.05, which is indistinguishable from no effect at all.
Crawlers get no guidance at all and some will not proceed.
GPTBot, ClaudeBot or PerplexityBot is told to leave. The site cannot be cited by an assistant that never read it.
Pages are discovered by luck rather than by list.
The page is prose to a machine, with nothing it can extract as a fact.
The company is a document, not an entity. Assistants recommend entities.
An assistant has to parse the design to find the words.
The two lines every index and every answer engine reads first are missing.
The page has no single stated subject.
Emerging and optional. Included for the trend, not the score.
Three findings
1. Almost nobody fails the basics. Almost nobody clears the bar.
Robots, sitemap, title and structured data each pass on more than 93% of sites. Those are the checks a good CMS does for you. The checks a person has to decide to do are where sites drop: 166 of 274 fail at least two, 64 fail at least three, and 6 pass all nine.
2. The most common gap is a text file.
Machine-readable copy (llms.txt or markdown) is the weakest real check: 124 sites, 45%, have
neither an /llms.txt nor a markdown
copy of the page. It is also the most common single biggest gap where the scan named one
(95 sites). It takes an afternoon.
3. One site in eight is not an entity.
35 sites have structured data but no Organization schema, so an assistant can extract facts from the page and still has nothing to attach them to. Assistants recommend entities. A site that is a document, however well written, is not in the running.
Checks passed, out of nine
No site passed fewer than three.
Site language
The list skews Italian because the search did.
Method
- Sample. 274 agencies in Europe that advertise SEO or GEO services, 160 in Italy and 114 elsewhere, narrowed from a first-party list of 1,758 reachable agency sites in 12 languages to those that market GEO and whose sites were live in August 2026. 2 further sites were excluded because the scan did not complete.
- Fetch. Four public URLs per domain: the homepage,
/robots.txt,/sitemap.xmland/llms.txt. No login, no crawl beyond the homepage, a 25-second deadline per site. - Checks. The nine listed above, each a pass or fail. They are the same checks the free AI-readiness check runs, and the code is in the open-source repository under
lib/audit/agent-readiness.ts. - Score. Weighted share of checks passed: high for crawler access, structured data and Organization schema; medium for robots, sitemap and machine-readable copy; low for title, h1 and content signals. Content signals are excluded from the score.
- Date. All sites scanned on 16 August 2026. Aggregates computed from the result file on 7 September 2026.
Questions about the data
Which sites were scanned?
274 agencies in Europe that advertise SEO or GEO services, 160 in Italy and 114 elsewhere, narrowed from a first-party list of 1,758 reachable agency sites in 12 languages to those that market GEO and whose sites were live. Two more were scanned and excluded because the scan did not complete. Agencies were chosen because they sell readability to their clients, so their own sites are the fairest test. No agency is named here.
What does the scan read?
Four public URLs per domain: the homepage, /robots.txt, /sitemap.xml and /llms.txt. It never goes deeper than the homepage and never logs in. The same nine checks run on the free check page on this site, and the code that runs them is in the open-source repository.
Why does the homepage say six pass, when the median score is 91?
Because the two measure different things. The score weights the checks by how much each one actually affects whether an assistant can read and cite you, so a site that gets the heavy things right scores in the nineties while still failing two or three light ones. Passing all nine is the stricter bar, and six sites clear it. Most sites are mostly readable; very few are fully readable.
Can I get the raw data?
Not the per-site rows: they name the agencies and were gathered for a scan, not for publication. The aggregate counts on this page are the complete set of figures we quote anywhere, and the method is reproducible on any list of domains with the free check.
The median moved from 83 to 91. Was the scan redone?
No. The scan is the same one, run on 2026-08-16, and no site's pass or fail on any check changed. What changed on 2026-09-22 is how the score weights those checks: it now follows measured effect on AI visibility rather than how severe a failure looks, which lifted sites that are fully crawlable but carry no schema or llms.txt. The same 274 rows were rescored under the new weights. Pass counts and the six-of-274 figure are weight-independent and did not move.
Will this be repeated?
Yes, on the same list, so the pass rates can be compared. The date on this page is the date of the scan it reports.
Be the site the assistant names.
Add a domain and the first draft is written while you watch. Every publish is your decision, a click on the draft or a rule you set and can hold, and you can read the source or self-host it free.
Set up your site and read your first drafts before you enter a card.