> Bron: https://neuralex.nl/en/blog/score-36-naar-85-agent-leesbaarheid > An online store scored 36 on AI readability and climbed to 85+ without changing a single line in the CMS. What that score measures, which fixes made the difference, and why they shipped through an edge worker instead of the CMS. [Back to insights](/en/blog) GEO ·August 11, 2026 ·7 min read # From a score of 36 to 85+: what actually raises an agent-readability score An online store scored 36 on AI readability and climbed to 85+ without changing a single line in the CMS. What that score measures, which fixes made the difference, and why they shipped through an edge worker instead of the CMS. The short answer: a low agent-readability score almost never comes from the content itself — it comes from the layer around it, the files and rules an AI agent consults before it even reads your pages. In one of our own runs with the [Agentic Search Optimizer](/en/showcase/agent), an online store scored a 36. After the recommended fixes — all shipped through an edge worker, without touching a single line in the CMS — the score climbed to 85+. That gap wasn't rewritten copy. It was three technical layers invisible to a human visitor: access for AI crawlers, machine-readable structure, and a navigation file that tells a model where to look. Below is what a score of 36 actually reveals, and which fixes made the difference. ## What a score of 36 actually means The Agentic Search Optimizer fetches a site the way an AI agent does — not the way a human sees it in a browser — and runs it through 15+ parallel checks: crawler access, schema.org data, agent discovery, robots rules for 19 known AI crawlers, and how machine-readable the content itself is. A score in the thirties almost always points to the same pattern: the site reads fine for humans and for Google, but is invisible or ambiguous to the crawlers and agents searching on behalf of an AI assistant. No `llms.txt`, missing or incomplete schema.org markup, robots rules that say nothing specific about AI crawlers, and content that technically exists but isn't structured predictably. That's a different problem than a classic SEO audit surfaces. A site can score well on speed and indexability — the classic foundation — and still be unreadable to a language model that needs structure and explicit markup to understand *what* a page is, not just that it exists. ## The fixes that made the difference **1\. Robots rules for AI crawlers.** A `robots.txt` written only with Googlebot in mind often locks AI crawlers out by accident — or leaves them unaddressed, which some agents read as caution. Explicit rules for the known AI crawlers were the first and cheapest fix. **2\. Schema.org markup.** Structured data tells a machine explicitly what a page is — product, article, organization — instead of making it guess from running text. For an online store, that's the difference between an agent recognizing a product page and citing it correctly, versus an agent that only sees a block of text. **3\. An llms.txt.** This navigation file points a model straight at the pages that matter, instead of forcing it to crawl the entire site itself. How that file works, and whether your site needs one, is covered in [our piece on llms.txt](/en/blog/llms-txt-uitgelegd). **4\. Agent discovery.** Like our own reference implementation — which uses DNS-based agent discovery and a machine-readable service description — this case also got discovery records that let an agent determine what the site offers immediately, without having to search first. ## Why an edge worker, not the CMS The most notable choice in this case was practical, not technical: every fix ran through an edge worker sitting between the visitor (or agent) and the existing CMS, instead of going through CMS changes directly. That turned a weeks-long implementation into days — no migration, no developer having to learn an unfamiliar CMS, no risk of a content change breaking a template. The edge worker adds the missing layers — headers, robots rules, schema.org, llms.txt — on top of what's already running. For a site that can't simply be rebuilt, that's often the only realistic route to a fast score jump. ## What this score does and doesn't guarantee A high agent-readability score closes the easiest leak: a model that simply can't read a site can't cite it either. It doesn't guarantee a mention — a technically flawless site can still be skipped if a competitor gives a more convincing answer. Technical readability and actually being cited are two separate measurements; how to check the latter yourself, periodically, is covered in [our piece on measuring AI visibility](/en/blog/meet-je-ai-zichtbaarheid). ## Conclusion The gap between a score of 36 and 85+ wasn't new content — it was four invisible technical layers that most sites simply never set up: crawler access, schema.org, an llms.txt and agent discovery, all addable without rebuilding the existing CMS. How that technical foundation combines with the editorial side into a complete GEO picture is covered in [our pillar article on GEO](/en/blog/geo-geciteerd-door-ai). Want to check? ## What does your site score on agent readability? The Agentic Search Optimizer runs a site through the same 15+ checks — crawler access, schema.org, llms.txt, agent discovery — and delivers a score with concrete fixes. [See the Optimizer](/en/showcase/agent) [Ask your question](/en/contact) --- Volledige (opgemaakte) versie: https://neuralex.nl/en/blog/score-36-naar-85-agent-leesbaarheid