How the AI Readiness score is measured
A score nobody can check is just an opinion with a number on it. This page says what we measure, why each part is weighted the way it is, what the score does not mean, and what we refuse to measure at all.
The AI Readiness Framework
Four layers, one question each, and five rules about how any of it may be reported. It is versioned, it is published in full on this page, and every audit we run is measured against exactly this and nothing else.
Discover
Can an AI system reach your information at all?
Read
Once reached, can the page be parsed into something meaningful?
Understand
Can an AI work out who you are and how your information fits together?
Answer
Can the site actually support an answer to what people ask about a business?
The five rules
Every score has evidence behind it
No number appears without the pages, quotes and properties it came from. If we cannot show you where a finding came from, it does not go in the report.
Every problem has an explanation in plain words
A finding names what is wrong and why it matters to a business, not just which property is absent. The technical restatement sits behind a switch for whoever needs it.
Every fixable problem has a route to a fix
Each item says who does the work: us, a coding assistant, or a person. A report that lists problems nobody can act on is a complaint, not a diagnosis.
Every applied fix gets verified
When a fix is published we re-read the pages it touched and confirm the change is actually live before anything is marked done. Nothing is credited on our say-so.
Every change can be monitored
Readiness is re-measured over time, and a drop names what got worse rather than only that the total moved.
Measuring is a fifth of it
A score on its own changes nothing. The framework is a loop, and the last two steps are the ones most tools leave to you.
- 1Measure. Four questions, one score, published weights.
- 2Explain. Where every point went, and the evidence for each finding.
- 3Fix. The problems that can be fixed, fixed.
- 4Verify. The pages re-read, and the change confirmed live.
- 5Monitor. Re-measured over time, with what changed named.
The principle behind all of it
A website is not AI-ready simply because it passes technical checks.
A site can be perfectly crawlable, perfectly valid, and still leave an AI system with no idea who the business is, what it sells, or where it works. That is why the weight sits where it does. Being reachable and parseable are necessary and cheap. Being understood is the hard part, and it is the part that decides whether anything can be said about you.
The four questions, and their weights
These are the weights we actually score against. They are published because a reader should be able to work out roughly where their points went without asking us.
Discover
15% of your scoreCan an AI system reach your information at all?
- Whether a sitemap exists and can be read
- What robots.txt allows, including the named AI crawlers
- The HTTP status of every page we open
- Whether the page needs JavaScript before there is anything to read
- Whether an llms.txt file is served (reported, never scored)
Read
15% of your scoreOnce reached, can the page be parsed into something meaningful?
- Page title, canonical tag, meta description, language declaration
- Heading structure, and whether levels are skipped
- How much of the page is its own content rather than navigation and footer
- Whether indexing is blocked
- Image descriptions
Understand
45% of your scoreCan an AI work out who you are and how your information fits together?
- Whether an Organization or Person entity identifies the business at all
- How much of the site carries structured data with real substance, not just page furniture
- How cleanly that data is wired together: broken references, orphans, duplicates, missing identifiers
- Whether your name, phone, address and website agree everywhere they appear
- Whether about, contact, privacy and terms pages exist
- Whether anything outside your site has heard of the business (reported, never scored)
Answer
25% of your scoreCan the site actually support an answer to what people ask about a business?
- Six fixed questions, asked of a language model given only your page text
- Whether each answer was possible at all, and how confidently
- Whether your structured data independently backs up each answer
How the evidence is collected
We open your pages the way a crawler does
Pages are discovered from your sitemap where one exists, and by following links where it does not. We identify ourselves as AISchemaGenAuditBot and respect robots.txt. A free check opens up to 100 pages, spread evenly across the site so the sample is representative rather than whatever happened to be first. Paid checks go deeper.
Every finding names the pages it came from
A finding that affects 18 of 19 pages says so, and lists examples. Where a claim comes from your own wording, the report quotes the sentence back to you. Where it comes from your structured data, it names the property.
The AI Answer Test uses only your page text
Six fixed questions are put to a language model given nothing but the readable text of one page. It has no access to your structured data, so its answers are what a system reading your prose would conclude. We then check separately whether your structured data independently declares the same thing. The gap between those two is the most useful number in the report.
Points come back exactly as they were taken
Each item on the fix list is worth a stated number of points. That figure is not an estimate: it is calculated from the same formula that produced the score, so fixing the item returns precisely that much.
What the score does not mean
It does not predict whether AI will mention you
Readiness is a property of your website: measurable, in your control, and provably improvable. Whether ChatGPT or Google actually cites you is a property of their systems, and it depends on your reputation, your coverage elsewhere, and decisions nobody outside those companies can see. We measure the input. We report the outcome separately. We will never tell you that adding a property causes a citation, because that is not a claim anyone can honestly make.
It is not a probability
73 out of 100 does not mean a 73% chance of anything. It means you have 73% of the signals we measure, weighted as above. Any tool presenting a readiness figure as a likelihood is inventing it.
Some things we can see, but cannot prove
A public check can see that you declared a link to a profile elsewhere. It cannot confirm that the profile is really yours, so those count as declared rather than verified, and the report says which. The same applies to anything asserted only in your own markup: we can tell you what your site claims, not whether the claim is true.
What we refuse to measure
More checks is not a better tool. Fifty meaningful measurements beat five hundred padded ones, and a score is only comparable if every part of it is defensible. These are deliberately out of scope.
Page speed, Core Web Vitals, mobile friendliness
Real concerns, thoroughly covered by Lighthouse, and not what decides whether an AI system understands your business. Folding them in would inflate the score with things that do not belong to the question being asked.
Content quality
Whether writing is accurate, original or well sourced is not machine-checkable, and a number attached to it would be a guess wearing a uniform.
Broken-link crawling and redirect chains
Screaming Frog territory. We report the HTTP status of the pages we open and stop there.
Anything we did not actually check
When a check did not run, the report says “not checked” rather than scoring it as a failure. A run that skipped a test and a site that failed one are different things, and every percentage we publish respects that difference.
When the methodology changes
The scoring formula carries a version number, shown on every report. Scores produced by different versions are not directly comparable, and we say so on the report rather than letting a change in our maths look like a change in your website.
Every audit stores the raw measurements alongside the score they produced, not just the final number. That means a change to the formula can be applied backwards across the whole history rather than stranding it, so your trend line stays honest through a version change.
Current scoring methodology
Version 7
What this page does not include
The per-check point values and the exact formulas are not published. Everything needed to understand, question or challenge a score is here: what is measured, why, how the evidence is gathered, how it is weighted, and what is deliberately excluded. The arithmetic itself stays ours. If you think a specific finding on your report is wrong, tell us which one and we will show you the evidence behind it.
See it applied to your site
The check is free and needs no account. Every finding shows the evidence behind it.
Check your AI readiness