On 3 October 2026 we ran the scanner over the homepages of 30 well-known sites, one request each, and scored the response against the 38 rules that existed at the time. None of them reached a grade A. The best was 83, the average of the sites that answered was 62, and the largest single group —16 of 30 (53%) — graded F.
That result is easy to over-read in both directions. It does not mean the average website is badly built; these are not average websites. It also does not mean AI search is broken for everyone. What follows is the distribution behind the headline number, including the nine rows that never produced a page at all, because the shape of the spread is the finding and the single average is not.
What did the 30 sites actually score?
Two sites reached a B, four a C, eight a D and sixteen an F. Nothing reached the 90-point band, which is the band this scanner calls an A.
| Grade | Score band | Sites | Share of the 30 |
|---|---|---|---|
| A | 90–100 | 0 | 0% |
| B | 80–89 | 2 | 7% |
| C | 70–79 | 4 | 13% |
| D | 60–69 | 8 | 27% |
| F | Below 60 | 16 | 53% |
The top five, in order: salesforce.com at 83, siemens.com at 80, vercel.com at 79, cloudflare.com at 77 and stripe.com at 75. The bottom of the answered set is sap.com at 41, which is still a page that rendered and a score that measures it.
Why is a B the ceiling rather than a missing A?
Because points are spread across twelve weighted dimensions and an A needs almost all of them. The best page in the sample missed seven to nine points across several unrelated rules, not one big one.
This is the part a single number hides. An engine that cannot read the page costs 16% of the score at once; a missing entity costs 12%; quotation and citation rules together cost another 11%. A site can be excellent on content and lose the A on markup, freshness and the two or three rules nobody has ever looked at.
| Dimension | Weight | What it is asking |
|---|---|---|
| AI crawler access | 16% | Can the crawler fetch the page at all? |
| Machine readability | 12% | Is the markup parseable, and is there an entity? |
| Content depth | 11% | Is there enough text, and is it structured? |
| Citability and evidence | 11% | Are there figures, quotations and outside sources? |
| Answer readiness | 10% | Are headings question-shaped with short answers under them? |
| Trust and authority | 10% | Author, contact routes, sameAs, and a visible date. |
| Semantic structure | 8% | Headings, landmarks and a single H1. |
| Metadata and discoverability | 7% | Title, description, canonical, Open Graph. |
| AI context files | 5% | robots.txt signals and an llms.txt. |
| Freshness | 5% | A machine-readable date that is recent. |
| International readiness | 3% | Language and hreflang declarations. |
| Delivery and mobile | 2% | Viewport, payload, no JavaScript dependency. |
What happens if the nine error responses are counted as zero?
The average falls from 62 to 44, and that second number is arithmetic rather than measurement. A 403 is not a badly built page; it is a site declining to answer, and scoring it zero would describe the decline as a content problem.
So the study excludes those nine rows from the mean and keeps them in the table. Of the 30 requests, 21 (70%) returned a page that could be scored. The other nine (30%) returned 401, 403, 429, 202 or 520. The distinction matters because the two failures need opposite responses: a low score on a page you served is a fix list, and a 403 is a decision about who you let in. That is the subject of its own piece, linked below.
What should a study of 30 homepages be trusted for?
It is a fixed sample, one page per site, read once, and it says nothing about the pages behind the homepage. Treat it as an illustration of how the checks behave on real sites, not as an estimate of the web.
- The sample is not random. 30 recognisable sites across 15 categories were chosen, which is useful for recognition and not for statistics.
- One page per site. A homepage is often the least representative page on a site, and the scanner scores the URL it is given.
- The rule set has moved. The rows were collected under 38 rules; the scanner ships 40 today, and the study page says so rather than quietly re-scoring history.
- Nine rows are error responses. They are published, flagged, and excluded from the average, because deleting them would hide the most interesting finding in the set.
How do I find out where my own site lands?
Run the scan. It takes one request, needs no account, and returns the same checks the study used — with the evidence that produced each result rather than just a number.
The useful comparison is not 83. It is the list of rules your page fails, because that list is actionable and 83 is not. If you want the full rows first, every one of the 30 is published at the study data endpoint, and the method behind the scores is on the methodology page.