---
title: Research · a14y agent readability
description: "Empirical findings about agent readability. How much does raising a site's a14y score actually save an AI agent? Measured results from running real coding agents against real websites at different scores."
doc_version: 0.2.0
last_updated: "2026-08-04T23:33:04.381Z"
---
# Research

Measured results from running real AI agents against real websites at different a14y scores.

-   v0.2.0 July 1, 2026
    
    ### [llms.txt saves an agent real tokens. But it won't read it unless you ask.](/research/llms-txt-linking/)
    
    A clean llms.txt discovery layer cuts an AI agent's tokens, but across five arms no in-page link (a head tag or a visible footer link) got the agent to use it. One line in the prompt did, and cut tokens 33% on the same page. The lever is the instruction, not the markup.
    
    [Read the study →](/research/llms-txt-linking/)
    
-   v0.2.0 June 25, 2026
    
    ### [The State of Agent Readability on the Web](/research/state-of-agent-readability/)
    
    We scored the 50,074 most-visited websites for how ready they are for AI agents. The median scores 52 of 100, not one scored "excellent," and the agent-era layer (llms.txt, AGENTS.md, markdown mirrors) is mostly missing or only half-built.
    
    [Read the survey →](/research/state-of-agent-readability/) · [Explore all 50,074 sites →](/research/web/)
    
-   v0.2.0 June 25, 2026
    
    ### [Which agent-readiness features actually pay off](/research/per-feature-ablation/)
    
    We toggled each of a14y's 11 agent-readiness features on and off to measure what each one is worth to an AI agent. Markdown mirrors and meta descriptions do most of the work, and the agent-skills directory actively hurts.
    
    [Read the study →](/research/per-feature-ablation/)
    
-   v0.2.0 May 16, 2026
    
    ### [Raising an a14y score from 37 to 89 halved an AI agent's token use](/research/scorecard-evals/)
    
    Same docs, same answer quality, statistically indistinguishable judge scores. The only thing we changed was running an a14y audit and shipping the top fixes, bringing the site's a14y score from 37 to 89. That roughly halved Claude's token use and tool calls on the same evaluation task.
    
    [Read the case study →](/research/scorecard-evals/)

## Sitemap

Full docs site index: [/sitemap.md](/sitemap.md).