How the score works
The 40-point AI visibility audit
Last updated
The AI visibility audit is the set of 40 checks behind every CitableReport score. Each check looks at one thing that affects whether AI assistants such as ChatGPT, Claude and Perplexity can find, read, trust and quote a website. The checks are grouped into five layers. This page explains how the score is worked out and what each layer covers.
How is the AI Visibility Score calculated?
The AI Visibility Score is the share of points a site earns across the 40 checks, out of 100. Each check is worth 3, 2 or 1 points (77 in all). A pass earns all of a check's points, a partial pass earns half, and a fail earns none.
- Site-wide checks, such as robots.txt, are judged once.
- Page checks, such as the opening paragraph, are judged on every page they apply to, and earn points in proportion to the pages that pass. If 70% of your articles open with a direct answer, that check earns 70% of its points.
- Checks that cannot be assessed are left out, so they neither help nor hurt. Five measurement checks stay out until you answer a short question about each.
Scores fall into four bands: Citable (80 and above), Nearly there (60 to 79), Patchy (40 to 59) and Invisible (below 40).
What does the score not measure?
The score does not measure reputation. It covers what you can change on your own site, not how many other sites link to you, how old your domain is, or how well known your brand is. Those matter to AI systems too, and the report shows them for context, but they are not in the score. So a famous site with plain markup can score lower than a small site that is carefully built.
Layer 1: Crawl and access
Can AI crawlers get in? Robots.txt rules, firewall blocks, sitemaps, redirects, indexing directives, and whether your content is in the HTML or hidden behind JavaScript.
We undertake 8 checks in the crawl and access layer.
Layer 2: Content structure
Can an answer be lifted out? One clear headline, ordered subheadings, an opening that answers the question, question-style headings, lists, tables, summaries and structured data.
We undertake 8 checks in the content structure layer.
Layer 3: Fact density
Is there something worth quoting? Specific figures, plain definitions, concrete examples, links to primary sources, statistics with their source beside them, and content that is kept fresh.
We undertake 8 checks in the fact density layer.
Layer 4: Entity and trust
Can they tell who you are? Organization data, named authors with their own pages, about and contact pages, one consistent brand name, linked profiles, and a place in the knowledge graph.
We undertake 8 checks in the entity and trust layer.
Layer 5: Measurement
Will you know if it works? Analytics, search console verification, an llms.txt file, and whether you track AI referrals, crawler visits and your share of citations.
We undertake 8 checks in the measurement layer.
Why do some checks apply only to some pages?
Some checks apply only to some pages because not every page should do everything. A contact page is not expected to cite sources, and a pricing page quoting its own prices is not expected to attribute them. Sourcing, bylines and summaries are judged on articles only; the fact checks need at least 300 words of text to judge.
How sure are these checks?
The checks are evidence-based but not certain. They look for signals that published research and the AI companies' own guidance associate with being cited. Nobody outside those companies knows exactly how sources are chosen, and it changes as models are updated. A higher score improves your chances; it is not a guarantee. The text checks currently assume English.