How we measure agent readability

As AI systems — large language models, autonomous agents, shopping assistants, and search overviews — become primary intermediaries between people and the web, a site's ability to be read and cited by these systems matters as much as its ability to rank in a traditional search index.

A site can rank well in conventional search results and still be effectively unreadable to AI-driven discovery. The reasons are structural: AI systems evaluate content differently from keyword-matching algorithms. They assess whether content is machine-readable, structurally coherent, semantically clear, demonstrably authoritative, and accessible to the crawlers that feed their training and retrieval pipelines.

The CovenAI v2.0.1 score was built to make these signals measurable. A high score indicates that a site is well-positioned to appear in AI-generated responses, be cited in LLM outputs, and be transacted with by purchasing agents. A low score reveals specific, fixable gaps in how the site presents itself to non-human systems. The score is available free via the scan tool and updated continuously in Agent Analytics for monitored sites.

Each site is evaluated across five independent scored dimensions. Scores for each dimension range from 0 to 100. The composite score is the weighted sum of all five and ranges from 0 to 100. Weights reflect the relative influence each dimension has on observed AI citation and transaction behaviour.

A sixth signal, Agent Correlation, is also measured and reported below for transparency, but contributes 0% to the public score. We treat it as a research signal because it can be inflated with synthetic agent traffic; using it in the headline number would reward gaming over genuine agent-readiness.

Agent Correlation · research signal · 0% to public score

We also observe live agent traffic against scanned sites — visit recency, agent-type diversity, dwell signals across the nine AI agent systems we track (GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-SearchBot, Claude-User, PerplexityBot, Google-Extended, Gemini-Deep-Research). This data is reported alongside the score for transparency, but contributes 0% to the public number. Including it would reward sites with synthetic agent traffic over sites with genuine agent-readiness; we treat it as a confidence signal we use to validate the rubric, not as a public scoring input.

The composite score runs from 0 to 100. Scores are normalised against observed performance across the CovenAI corpus, which is updated as new sites are added. The bands below reflect meaningful thresholds in real-world agent readability outcomes.

In April 2026, CovenAI ran the v2.0.1 scoring methodology across 938 sites spanning a range of industries, regions, and site types. Of those, 488 returned a complete score across all five scored dimensions. The remainder had one or more dimensions that could not be evaluated — typically due to crawl access restrictions or insufficient content depth. The corpus continues to grow as sites connect to Agent Analytics monitoring; the figures below reflect the April 2026 snapshot.

The results show that most sites are partially readable by AI systems but fall short on the signals that drive citation: structured data, clear authorship, fresh date markup, and permissive AI crawler access.

The average score of 47.4 places most sites in the Needs Work band. The median of 50 indicates the distribution is relatively even around the midpoint, with no heavy skew towards the extremes. The highest score recorded was 76, placing it at the top of the Good with caveats band — a reminder that even well-optimised sites tend to have meaningful gaps in at least one dimension.

The most consistent low-score drivers across the corpus are absent citability signals (missing structured data, no author attribution, no date metadata) and restrictive robots.txt configurations that block one or more major AI crawlers.

Scores are derived from live page analysis conducted by CovenAI’s scanning infrastructure — including Coven-Citability-Bot, our own web agent — at the time of assessment. Scores reflect the state of a site at scan time and will change as the site evolves. Sites connected to Agent Analytics receive updated scores on a regular cadence.

The analysis draws on signals from the publicly accessible version of each page as seen by a standard web client, structured data validators, the Agent Diagnostic Layer (ADL) for agent identity and traffic classification, crawl behaviour logs (for the Agent Correlation research signal), Transactability surface probes, and heuristic evaluation of content quality signals. No proprietary or authenticated data is used; scores reflect only what AI systems themselves can observe.

Industry benchmark data is aggregated and anonymised. Individual site scores are not disclosed in public reports.

Because scores decompose into five independent dimensions, improvement is systematic rather than speculative. The highest-impact actions are almost always in the lowest-scoring dimensions. Across the corpus, the most consistent low-score dimensions are Citability and Discoverability — both addressable with focused, low-effort changes.

Structure is consistently the most surprising high-impact area. Pages with genuinely good content often score poorly on structure because of invisible markup issues: FAQ sections built with <details> accordions register zero headings to an AI agent; section titles styled as <p> tags are indistinguishable from body copy; comma-separated item lists score nothing where a <ul> would score full points. These are low-effort fixes with outsized score impact. Read the full breakdown →

Run a free scan to see your site's current scores across all five dimensions. Agent Analytics provides continuous monitoring and updated scores as your site evolves.