How the score works

Weights as published on 18 September 2026. When a weight changes, this date changes with it.

The AI Website Critic is a deterministic website scorer. It runs ten checks against a homepage, each check subtracts published weights for the problems it finds, and one critical failure caps the whole score. No language model is involved in the number, so the same page gets the same score every time, and every finding points at something you can verify in your own HTML. This page publishes the whole calculation.

What does the AI Website Critic measure?

A website brings in work when three things are true: people and machines can find it, a stranger believes what they read, and that stranger can work out what to do next. The ten checks are grouped under those three questions. Two carry extra weight because they are where most real-world failures live: whether a crawler can read the page at all, and whether a visitor can actually make contact.

How is one category scored?

Every category starts at 100. Each finding carries a weight, and the category score is 100 minus the sum of the weights of everything found, never below 0. A category with nothing to report scores 100, with one exception: Content Quality, Trust Signals, Purpose Clarity and Experience and Mobile are capped at 85 in the automated pass, because counting what is on a page is not the same as reading it. The top of those four scales is reserved for the full audit, where a person reads the site.

What does each check look for?

Every finding the engine can raise, with its severity and the weight it subtracts. If a finding is not in these tables, the tool cannot raise it.

Crawler Comprehension

Does the content exist before JavaScript runs? Most AI crawlers do not execute JavaScript, so a page that only fills itself in through a browser is blank to them. This is the check that most scanners skip, and it is the single most expensive failure a modern site has.

AI Readiness

Are AI crawlers allowed in, and is there a plain-text summary waiting for them? Fourteen crawlers are read from robots.txt, in three tiers. The five that power a live answer engine are also tested by requesting the homepage under their own user agent, because a firewall can refuse a crawler that robots.txt welcomes. A crawler counts as blocked if either says so. Tier matters: blocking a search crawler removes you from answers people read, blocking a training-only crawler does not, and the weights say so.

robots.txt is read the way crawlers read it: a group naming the crawler wins over the wildcard group, and only a root Disallow: / counts as a block.

Schema Markup

Is there structured data telling machines what the page is? Only JSON-LD in the raw HTML counts, because that is what a crawler receives.

AEO Structure

Is the page shaped like an answer? Answer engines extract by heading, and they match a visitor's question to a heading phrased as one. Judged on the rendered page, so a client-rendered site is not charged twice for the same root cause.

Technical SEO

The basics that every engine reads first. The certificate is checked with a direct connection rather than inferred from the response, because an expired certificate still serves pages to scripts while stopping every real visitor at a browser warning.

The redirect checks come from two extra requests: plain http:// for the same host, and the www or bare twin of the address, where one plausibly exists. A deeper subdomain has no twin and is not checked.

Content Quality and E-E-A-T

Is there enough on the page, and does anyone put their name to it? Capped at 85 in the automated pass.

Trust Signals

Would a stranger believe this business exists and is reachable? Capped at 85 in the automated pass.

Purpose Clarity

Can a visitor tell in three seconds what this business does and for whom? Capped at 85 in the automated pass.

Funnel and CTAs

Can the visitor act, on a phone, with one tap? This is where enquiries leak silently, and it carries a weight of 1.5 for that reason.

Experience and Mobile

Does the page work on the device most visitors are holding? Capped at 85 in the automated pass; visual design is judged from a screenshot in the full audit.

How is the headline score computed?

The headline score is the weighted average of the scored categories, then placed inside a band set by the severity of what was found. The weighted average alone was tried first and proved too generous: a site can be completely invisible to AI crawlers and still average a B, because nine healthy categories outvote one fatal one. A site with a critical defect is not a B-grade site with one problem. It is a site that fails at something, and the headline has to say so.

The ceiling rule

The worst findings across the scored categories choose a band. The weighted average then decides where in the band the score sits, using the headroom above the band's floor: a site that scrapes into its band sits at the bottom of it, one that is otherwise excellent sits at the top. The band can only ever pull a score down. If the weighted average is already below the band, the lower number stands.

Formally: position = (weighted - floor) / (100 - floor), clamped between 0 and 1, then score = min(weighted, round(floor + (ceiling - floor) * position)).

Which findings get the "Holding your score" badge?

Only the findings of the severity class that chose the band, and only when the band actually pulled the score below the weighted average. A critical finding on a site whose average was already low is not holding anything, so it carries no badge. The badge says "fixing this is what moves the number", which is a different and more useful fact than severity alone.

What do the letter grades mean?

Because one critical failure caps the score at 55, a site with any critical finding cannot grade above D, and a site with two cannot grade above F. That is deliberate.

Why does the free check score only four of ten?

The free check reads raw HTML, which is exactly what a non-JavaScript crawler receives. Four categories can be judged honestly from that: Crawler Comprehension, AI Readiness, Schema Markup and Technical SEO. The other six describe a human with a browser, and scoring them from raw HTML on a client-rendered site would report "no call to action, no testimonials, no phone number" about a site that visibly has all three. Those categories are marked unscored rather than guessed at, and the free headline is labelled with the question it answers, "Can they find you?", computed only from the four categories that contributed to it. The full audit loads the site in a real browser and scores all ten.

What happens when a site is down or blocks the scanner?

A homepage that returns a server error or a 404 is an error page, not a website, so it is not scored. Scoring it would invent findings about content and calls to action that an error page cannot meaningfully have. The tool says the site could not be reached and stops. A homepage that returns 401, 403 or 429 to a normal browser request is reported as blocked by a firewall or rate limit, which is different information from being down, and the crawler comparison is skipped because it would be meaningless when a normal browser is refused too.

What the tool does not do

• It does not use a language model to score. Every number comes from counting and matching. The paid audit adds a person reading the page; the free check never claims to.
• It reads the homepage only, plus robots.txt, llms.txt and sitemap.xml. Inner pages are not crawled.
• It does not know whether AI engines are currently citing you. It measures readiness to be cited, not citations. Nobody can measure the latter from your HTML.
• It probes five AI crawlers live and reads fourteen from robots.txt. The five are the ones behind ChatGPT, Claude and Perplexity. Blocking at the firewall for the other nine is not tested, because each live probe is another request against a stranger's server.
• It has no field performance data. Speed is one server response time from Johannesburg, not Core Web Vitals from real visitors.
• It cannot read what it is refused. A firewall that blocks the scanner gets reported as such, not scored around.
• Absence of evidence is not excellence. A category with no findings means the counted things are present, which is a floor, not a verdict. That is why four categories are capped at 85 until a person reads the site.

Is the calculation stable?

The weights on this page are the weights in the engine on the date at the top. When they change, the date changes, and a score from before that date is not directly comparable with one after it. The same page scanned twice on the same day gets the same score, apart from things that genuinely changed in between: a slow server response, a certificate nearing expiry, or a firewall that refused one of the five probed crawlers this time.