How we set, verify, and change the weights behind the Hidden Layer 100-point agent-readiness score.
·
The grade answers the implementation question—what agent-readiness signals a site has implemented—not a prediction of citation.
How weights are set
A weight is described as evidence-derived only when a named verified corpus run set the value. If the verified corpus cannot set it, we publish a named decision-maker and a one-line rationale for the product-policy decision instead. The current ledger makes that distinction check by check; it does not relabel judgement as data.
Before this work, all 100 weighted checks had no stated basis that named either a qualifying corpus run or an accountable person. That is the honest starting count, retained in the ledger even as the evidence improves.
Every weight change triggers recalibration against the verified corpus and writes the unrounded score and grade movement for every corpus domain, including an explicit “no grade change”. Keeping six decimal places prevents presentation rounding from hiding a real change. A stale report cannot pass because each report identifies the before and after weight vectors and the corpus bytes it consumed. Download the per-weight evidence and calibration measurements →
Grade Scale
A90–100Agent-ReadyMost audited agent-readiness signals are implemented.
B75–89GoodStrong fundamentals. Minor gaps in agent integration or schema.
C60–74Needs WorkKey AI bots blocked or missing. Significant render gap.
D45–59PoorBlocking most AI traffic. No agent integration signals.
F<45InvisibleFew audited agent-readiness signals are implemented.
01Categories
01
Discoverability
Can AI crawlers find and access your site? Foundation layer — without this, no other checks matter.
Which AI bots can crawl your content? Checks 12 canonical AI User-Agents against your robots.txt, inheriting User-agent: * fallbacks per RFC 9309. Highest-weight category because access is binary.
Have you published the machine-readable files LLMs use to understand your site's content scope?
Values in evidence ledger
/llms.txt published/llms.txt format valid (# heading + markdown links)/llms-full.txt published (single-file complete docs dump)
04
AI Visibility
Can AI parse and cite your content correctly? Structured data, entity linking quality, schema type coverage, content legibility, and emerging agent-mode signals.
Values in evidence ledger
JSON-LD structured data quality — all schema types scoredEntity linking quality — @id + Wikipedia/Wikidata sameAs in JSON-LDOpen Graph meta tags (og:title, og:description, og:image)sameAs entity linking — meta-tag levelContent efficiency — text/HTML ratio (JS-heavy penalty)FAQPage schema qualityArticle/BlogPosting schema quality — author + datePublishedContent-Signal directive in robots.txtWebSite schema with name + urlBreadcrumbList schemaVideoObject schema on video pagesTwitter Card meta tagsAccept: text/markdown content negotiationSchema.org speakable markup/pricing.md machine-readable pricing?mode=agent returns structured content
05
Agent Integration (bonus)
Future-ready signals for autonomous agents. Low-weight because adoption is <1% as of May 2026. We track these to reward pioneers without punishing the rest.
Machine-payable endpoint signaling + Google Merchant Center feed readiness. GMC checks only fire on domains with Product schema detected.
Conditional values in evidence ledger
Link: rel="payment" header or Payment-Pointer header (x402)Google Merchant Center feed detected on product pagesGMC required attributes coverage — id, title, price, availability, image, brand, gtin/mpn
07
Product Pages (e-commerce)
For domains with product pages, we auto-discover up to 10 URLs from the sitemap and audit each for Schema.org product schema completeness. Per-page checks plus an aggregate coverage score.
Is your brand visible to AI systems — in training data, in search, in category rankings? Checks cold recall via CF Workers AI, category share of voice, Wikipedia presence, brand search discoverability, plus third-party corpus signals.
Values in evidence ledger
LLM cold recall — does the configured model describe the brand accurately from training data alone?Category share of voice — does the model name you when asked about your category?Wikipedia / Wikidata entity presenceHacker News mentions via hn.algolia.com — proxy for community/corpus presenceBrand-name search returns own domain via DuckDuckGo instantReddit mentions — measured but currently unweighted because the verified corpus could not reach Reddit
Scoring model: Each currently weighted check has the fixed value and accountable basis published in the evidence ledger. Raw scores are summed across applicable, measured checks and normalized to a 100-point scale; a check we could not measure is not charged to the customer. Bonus categories track early or conditional capabilities with limited influence on a general-site grade. Bot rules inherit User-agent: * fallbacks per RFC 9309; sitemap-index files are recursively expanded to count real URLs. Methodology version: 1.4 (September 2026).