AI Readiness
Whether AI assistants can reach, read and understand your site. Five scored categories, the checks RankX AI actually runs, and why a score can be withheld.
AI Readiness scores whether AI assistants can reach your site, read it, understand what you sell, answer buyer questions from it, and find anything vouching for you. It reads your site rather than your answers, so it needs no runs and produces a result on day one.
Two things make it unusual, and both are deliberate: the overall score is your weakest category, not an average, and below a coverage floor there is no score at all.
AI Readiness is also the one measurement surface that stays open on free access after a trial ends, which is why this page is careful about which of its checks are fetched live and which are read from stored data.
The five categories
| Category | The question |
|---|---|
| Can AI reach you? | Are the assistants allowed to read your site at all? |
| Can AI read you? | When they do read it, can they tell what your business is? |
| Does AI know what you are? | Is what you sell written down in a form a machine can quote? |
| Can AI answer buyer questions? | When a buyer asks what you cost or how you compare, is there an answer to find? |
| Does AI trust you? | Is there anything beyond your own website vouching for you? |
They run cause to effect, which is also the order to fix them in. A site the assistants cannot reach makes every other category a measurement of nothing, so the reach findings come first no matter how tempting the others look.
The score is your weakest category
This surprises people, and it is the more useful number.
An average of every check by weight can read 85 for a site whose machine-readable layer scores zero. The arithmetic is right and the number misleads: most of the denominator is weight nobody ever loses, so the one thing that is actually broken dilutes away.
The weakest category is a claim that stays true: an assistant that can reach your site but cannot tell what you sell is not 85% of the way there. It also names the thing to fix, which a bare number does not, so the score always renders beside the category holding it down.
A useful property falls out of this: you cannot raise the score by passing more easy checks. Only fixing the worst thing moves it.
The coverage rules
A score never renders without its coverage fraction. "72" alone is a claim about a site; "72, from 14 of 19 checks run" is a measurement. If you quote the number elsewhere, carry the fraction.
Below the coverage floor there is no score at all. Not zero, which reads as "your site is broken", and not 100, which reads as "nothing to do". The honest output is the gap and the action that closes it.
A check that could not run moves the score in neither direction. Counting an unrun check as a failure would tell you to fix something nobody looked at; counting it as a pass would claim a verdict nobody reached. It leaves both the numerator and the denominator, and is counted separately so the gap stays visible.
Why a check might not have run
Each unassessed check says which reason applies, because the remedies differ:
| Reason | What it means | What fixes it |
|---|---|---|
| Pages not yet indexed | RankX AI has not crawled your pages | Run a page scan |
| The fetch was refused | Your server declined the request | Nothing on your side is missing; this is not "the file is absent" |
| The fetch was not attempted | There was no address to fetch | It runs as soon as there is one |
| Nothing has been generated | No artefact exists to compare against | Generate one |
| No ground truth | Nothing of that kind exists to look at: no tracked rivals, no articles, no citation history | Different from "we have not looked at your site", and a page scan will not fix it |
That last distinction is load-bearing. Offering a page scan as the remedy for a Website with no competitors recorded is selling a fix that cannot work, so RankX AI counts separately how many unassessed checks a scan would unblock.
Some checks are fetched live, others read stored data
AI Readiness mixes two kinds of evidence, and the page labels which is which rather than rendering them identically.
The site-side checks are live. Your robots.txt, your llms.txt and your
homepage are fetched every time you open the page, so those verdicts are always
about your site as it is now.
The page-level checks read stored crawl data, which only a metered page scan refreshes. They carry a line saying how many pages they were run against and on what date. Past thirty days that line changes tone and offers a re-scan.
The distinction matters most on free access after a trial, where re-crawls stop and stored page content is deleted after seven days. Without the label, checks turning to "not assessed" would look like the product breaking rather than the evidence running out.
The date line extends the coverage rules above to time: a verdict from four months ago never renders exactly like one computed a second ago, so stale evidence is never presented as current.
What the checks look at
The full list, with the category and weight of each, is in the AI Readiness check reference, which is generated from the product so it cannot list a check that is not run.
The shape of it:
Reach is about access. Whether AI crawlers are blocked, whether robots.txt disallows everything, whether commercial pages carry a noindex, whether snippet controls cap what may be quoted, and whether your content only exists after JavaScript runs.
Read is about machine-legibility: whether you publish an llms.txt, whether
it says anything useful, and whether facts a buyer needs are locked in PDFs.
If you have no file yet, the free llms.txt Generator
drafts a curated one from your sitemap; it makes the site easier for an agent
to read, and no study has measured it lifting citations.
Know is about entity clarity: whether your organisation is described in structured form, whether it is linked to the profiles that identify it, and whether your services are written down.
Answer is about the questions buyers actually ask: whether you state a price at all, whether the price is parseable, whether comparisons and integrations exist, and whether pages answer their own headings.
Trust is about everything beyond your own site: review footprint, author signals, and whether your dates are present and consistent.
Crawler access is checked per purpose, not per vendor
The reach category checks fourteen crawler tokens across ChatGPT, Claude, Perplexity, Google, Apple, Meta, Amazon, TikTok and Common Crawl, and it distinguishes what each one is for:
| Purpose | What blocking it does |
|---|---|
| Search | Removes you from that assistant's answers. This is the one that costs you |
| Training | Keeps your content out of a future model. Does not affect whether you appear in answers today |
| User fetch | Stops the assistant fetching a page a person explicitly asked it to read |
Blocking only the training crawlers is a legitimate choice and RankX AI scores it as one. A robots.txt that blocks training but allows search costs almost nothing in answer visibility, and the check's weight reflects that rather than charging the full penalty for a deliberate opt-out.
Blocking a search crawler is different, and it is the finding that makes every other number on your account meaningless.
Fixing what it finds
Each finding carries an action, and there are three kinds:
- RankX AI does it, where a connected WordPress site makes the change applicable directly.
- Copy this, where RankX AI generates the artefact and you place it.
- Hand off, where the fix belongs to someone with access to something RankX AI does not have.
Findings can also become tasks on the Tasks board, which is where they belong if the work will not happen in the next ten minutes.
RankX AI watches the site-side signals daily
The reach and read checks answer live, and RankX AI also runs them once a day
in the background on every account, so a change gets noticed whether or not
you open the page. When AI crawler access changes for the worse, or an llms.txt
disappears, a strip at the top of this page says so on your next visit.
The watch compares meaning rather than bytes, treats an improvement differently from a regression, and never reports a failed fetch as a change. See the daily site-signal watch for what it fetches, what it stores and what it deliberately does not do.
Answer accuracy is a separate, metered band
Everything above reads your site. The answer accuracy check reads the assistants' answers instead, and judges what they say about your business against the facts you have published in Brand Book.
It sits beneath the readiness score on this page and costs 5 credits per run, covering one pass over every stored platform answer for one brand prompt. It is metered, so it is not part of the free band.
Sharing a readiness report
An AI Readiness run can be shared as an unauthenticated link, which is the usual way an agency hands one to a client. See share links for every shareable surface and what each exposes.
Where to go next
- The AI Readiness check reference, the generated list of every check.
- The daily site-signal watch, which keeps the reach checks current without you opening anything.
- Answer accuracy, for whether what assistants say about you is true.
- The Website Audit, which is the technical crawl rather than the machine-legibility scorecard.
- Tasks, for turning findings into work.
Last updated
Tracked platforms
The six AI assistants RankX AI tracks, what each is good for, how the per-Website and per-prompt platform choice works, and why Google AI Overviews sits apart.
The daily site-signal watch
RankX AI fetches your robots.txt, llms.txt and homepage daily on every account, compares meaning rather than bytes, and reports when crawler access worsens.