How We Score AI Friendliness: Inside the ClariCard Algorithm

When we set out to build an "AI Friendliness Score," the first question we had to answer was deceptively simple: what does it actually mean for a webpage to be friendly to an AI?

Search engines like Google are optimized by decades of signals — PageRank, backlinks, click-through rates. But large language models (LLMs) like ChatGPT and Gemini work differently. They don't crawl and rank. They read, summarize, and cite. A page that ranks #1 on Google can still be invisible to an AI assistant if its content is ambiguous, poorly structured, or semantically sparse.

We built the ClariCard AI Score to measure exactly that gap.

Why 100 Points?

We chose a 100-point scale deliberately. It maps to familiar mental models (test scores, percentages), makes grade thresholds intuitive, and gives us enough resolution to show meaningful change over time without appearing falsely precise. The score is a composite of five dimensions, each weighted by its observed impact on LLM citation likelihood based on our internal research corpus.

Dimension 1 — Entity Clarity (25 points)

LLMs build knowledge by recognizing named entities: people, organizations, products, locations, concepts. When a page clearly establishes what it is about and who is behind it, LLMs can anchor that information confidently and are far more likely to cite it.

We evaluate: presence of structured schema.org markup (Organization, Person, Product, Article), consistency of the brand name across the page, explicit authorship attribution, and whether the core subject of the page is named within the first 200 words. Pages that bury their identity deep in the footer or rely entirely on navigation context score low here.

Dimension 2 — Answer Readiness (25 points)

AI assistants are answer engines. They synthesize information in response to user questions, and they strongly prefer sources that contain self-contained, direct answers rather than sources that require contextual navigation.

We measure: presence of FAQ or Q&A schema, sentence-level question-answer patterns (a paragraph that opens with a question and resolves it within 2-3 sentences), definition structures ("X is a Y that..."), and numbered step patterns for how-to content. A page full of marketing language but no concrete answers scores low, regardless of how well-written it is.

Dimension 3 — Semantic Density (20 points)

Semantic density is our measure of how much meaningful, topically relevant content a page contains relative to its total word count. LLMs prefer dense, substantive pages over thin or padded ones.

We compute a term frequency distribution across the page, remove stopwords, and measure the ratio of domain-relevant vocabulary to total token count. We also penalize keyword stuffing (unnatural repetition of the same phrase) and reward use of related terms and synonyms, which signals genuine topical depth rather than optimization.

Dimension 4 — Crawl Signals (15 points)

Before an LLM can cite a page, it (or its data provider) has to crawl it. Crawl signals measure how accessible and well-structured a page is to automated systems.

We check: presence of a valid robots.txt that allows common AI crawlers (GPTBot, ClaudeBot, Google-Extended), sitemap availability, page load time under simulated mobile conditions, canonical tag accuracy, and absence of content hidden behind JavaScript-only rendering. Even a perfect page is invisible if it actively blocks AI crawlers.

Dimension 5 — Trust & Authority (15 points)

LLMs are trained on data that implicitly encodes authority signals. Pages with stronger trust indicators are more likely to appear in training data and more likely to be cited in retrieval-augmented generation (RAG) systems.

We evaluate: presence of HTTPS with a valid certificate, external links to authoritative sources (government, academic, major publications), internal linking structure that distributes authority across the site, a visible and specific "About" or author page, and a privacy policy. These signals don't directly affect LLM training, but they correlate strongly with the content quality that training data favors.

How Scores Are Calculated

Each dimension is scored independently by our analysis pipeline, which runs a combination of static HTML analysis, schema validation, network requests (for crawl signals), and Claude-powered semantic evaluation. The five raw scores are normalized to their respective maxima and summed.

We do not use a black-box model to produce the final score. Every point is traceable to a specific, auditable signal. This matters because we want you to be able to act on your score — not just observe it.

What a Good Score Looks Like

Scores above 80 (Excellent) typically belong to pages that have intentionally optimized for clarity: they have complete schema markup, contain explicit Q&A structures, use clean semantic HTML, and don't block AI crawlers. Think well-maintained documentation sites, structured FAQ pages, and company "About" pages written for both humans and machines.

Scores between 60-79 (Good) represent pages that do several things well but have meaningful gaps. The most common gap is Answer Readiness — most marketing pages are written to persuade, not to directly answer questions.

Scores below 40 (Poor) often block crawlers entirely, contain very little semantic structure, or consist primarily of images and JavaScript with no indexable text.

What's Next

We're currently developing a Citation Tracking dimension that measures whether your pages are actually appearing in LLM outputs — moving from predictive scoring to observed measurement. We're also building per-paragraph scoring so you can see exactly which sections of a page are dragging your score down.

The algorithm is version 2.0. We'll keep improving it as we learn more about how LLMs select and cite sources. Every change will be documented here.