Methodology changelog

The AI mentions check now also asks Gemini with Google Search grounding. The mentions check sits beside the score and adds no points to it.

Before a report is delivered, a correction pass removes every sentence our accuracy checks flag, and delivery waits until none remain.

Public benchmark figures now update only from the living benchmark. When the living figures cannot be read, the pages keep their last dated figures.

On Shopify stores, fix steps now name the Shopify admin, Liquid and product.variants exactly. Other platforms keep steps written for their own system.

The scan detects duplicate and conflicting structured data on every site and marks each conflict for resolution. It never recommends a second copy of an entity the theme already prints.

A person's role written two ways, such as Founder on the page and Chief Executive Officer in the markup, is read as one identity, and no conflict is reported.

The scan date is no longer passed to scoring as page context. Freshness is judged only from the dates a page publishes itself.

The access check reads every documented AI crawler token, and it matches robots.txt paths the way the robots.txt standard describes.

A page counts as noindex when either the meta robots tag or the X-Robots-Tag header says so.

An llms.txt or agents.md file counts as found only when the server returns a plausible text file at that address.

Every public page that quotes the living benchmark takes its figures from one reading under one date stamp.

Findings are checked against the structured data observed across the whole scan, so a fact that holds for one page is never reported as a fact about the whole site.

The extractor records image alt coverage, body links, the page language, hreflang and microdata types from the raw HTML, and it merges them with the JSON-LD facts before any finding is written.

Each scan has a time limit for crawling, so a very large site returns a clear result within a bounded time. Very large JSON-LD blocks are trimmed largest first, so the readable text of a page always reaches scoring.

The crawl follows the host that a site's own sitemap declares for its content, and text files are never scored as pages.

The figures in our FAQ are pinned to the cohort of all websites, so each figure names the population it came from.

Before a finding can say that something is missing, the engine checks facts extracted from the page, such as its headings and schema types. A finding that the page disproves is removed.

The living benchmark began. It grows as we scan, refreshes daily and runs beside the frozen 2026 study.

Every scan summary shows where the site stands against the study cohort.

AI crawler access became a measured signal. The scan tests content pages with AI crawler user agents and reads the robots.txt rules that apply to them.

Bot walls, soft 404 pages and empty renders are kept out of scoring, and a site with too few readable pages receives a Crawl Diagnosis that names the cause.

The scan identifies the platform first, so Shopify stores receive Shopify specific fixes and other platforms receive fixes for their own system.

The crawler waits for each page to settle before it captures it, so two scans of an unchanged page read the same content.