This translation has not been editorially reviewed yet. The German version is authoritative. German version →
InhaltQuellen und Messmethoden
Where our audit rules come from
An audit report is only as good as its sources. That is why every one of our findings states what it is based on, and why this page says who is behind those addresses. The most important distinction is the one between standard and tool: The standard is binding. The documentation of an audit program only explains how that program checks the standard.
In short: who is actually measuring here?
We measure the accessibility of your pages with axe-core, the most widely used automated WCAG checker. It comes from Deque Systems and is also built into the developer tools of Chrome, among others. What axe finds, we recalculate against the complete WCAG catalogue and add our own measurements: contrast, keyboard operation, focus. What a machine cannot decide remains open and is judged by a person; it counts neither as met nor as an open finding.
The sources in detail
W3CStandard
The World Wide Web Consortium is the international body that sets the standards of the web, among them HTML and the accessibility guidelines WCAG.
This is where the binding wording of the success criterion that was checked stands. It is the standard itself, not its interpretation.
Open w3.org ↗ · also www.w3.org
schema.orgStandard
Joint vocabulary of Google, Microsoft, Yahoo and Yandex for structured data, the machine-readable statements from which search engines and AI systems draw facts about your page.
Defines what the structured data must look like that we rate in the audit area AI visibility.
Amtliche RechtsquelleStandard
Official publication of the legal text (Official Journal of the EU or the Austrian federal legal information system).
The original text of the law: it is authoritative, not our summary.
Open eur-lex.europa.eu ↗ · also ris.bka.gv.at
Deque UniversityTool documentation
Deque Systems is a US company for digital accessibility. It develops axe-core, the audit program we use to measure your pages for accessibility. The same audit library is built into the developer tools of Chrome and into Microsoft Accessibility Insights, among others; it is the most widely used automated checker for WCAG.
Deque University is the documentation of this audit program. For every rule that was triggered on your site, it explains in detail what is behind it and how to fix it. ⚠ It is not a standard: the WCAG is binding (see w3.org); the Deque page only explains how our measuring tool checks it.
Open dequeuniversity.com ↗ · also deque.com, accessibility.deque.com
Google Web-DokumentationVendor statement
Google publishes here the requirements by which its own search and the Chrome browser rate websites, among them the thresholds of the Core Web Vitals.
The loading-time thresholds in this report (2.5 s for the largest element, 0.1 for shifting content) come from there. We did not choose them ourselves. It also states what Google requires for its AI features: a page must be indexed and eligible to be shown with a snippet, there are no additional requirements, no special markup and no AI file such as llms.txt (“will neither harm nor help”); nosnippet and max-snippet take a page out of AI answers, Google-Extended only controls training of other Google systems. That is what we measure the audit area AI visibility against (as of 2026-09-20).
Open web.dev ↗ · also developers.google.com, developer.chrome.com
Bing Webmaster-RichtlinienVendor statement
In 22 points, Microsoft’s Webmaster Guidelines describe how Bing, Copilot and the grounding interface discover, render, index and cite websites; the “AI Performance” report shows site owners which pages are cited in AI answers.
The most detailed vendor source on AI visibility: sitemaps with canonical URLs only, 301 instead of 302 for permanent moves, no content available only through client-side rendering, NOARCHIVE and NOCACHE take a page out of Copilot answers, essential information at the top instead of a long introduction, markup must match the visible content. Our rules on these cite the point (as of 2026-09-20).
Open bing.com ↗ · also www.bing.com, blogs.bing.com
OpenAI — Crawler-HinweiseVendor statement
OpenAI runs ChatGPT and explains which crawlers it uses to read websites: OAI-SearchBot for search inside ChatGPT, GPTBot only for training its models, ChatGPT-User for requests on behalf of a user; each setting works independently.
We check these names in your robots.txt and with a probe against your firewall. OpenAI verbatim: a site that blocks OAI-SearchBot “will not be shown in ChatGPT search answers”; blocking GPTBot is an opt-out from training and costs no visibility. Blocking layers according to OpenAI: robots.txt, firewall (403), human verification (CAPTCHA, JavaScript challenge), rate limiting (429): exactly what our probe checks.
Open help.openai.com ↗ · also platform.openai.com, openai.com
Perplexity — DokumentationVendor statement
Perplexity is an AI answer engine that cites web sources for every answer. PerplexityBot finds and links pages for search and, according to Perplexity, is not used to train models; Perplexity-User fetches pages on behalf of a user and ignores robots.txt.
PerplexityBot is part of our check of access for AI crawlers. Perplexity requires a firewall to allow the crawler by user agent and published IP addresses; a 403 response takes the website out of its answers.
Open docs.perplexity.ai ↗ · also perplexity.ai, www.perplexity.ai
Anthropic — Crawler-HinweiseVendor statement
Anthropic runs Claude and names three accesses: ClaudeBot collects content for training, Claude-SearchBot for search inside Claude, Claude-User for requests on behalf of a user.
Claude-SearchBot is part of our check of access for AI crawlers; Anthropic verbatim: blocking it “may reduce your site’s visibility and accuracy in user search results”. Blocking ClaudeBot is an opt-out from training and costs no visibility.
Open support.claude.com ↗ · also support.anthropic.com, docs.anthropic.com, anthropic.com, www.anthropic.com
MDNReference work
The reference work of the Mozilla Foundation (Firefox) for web technology, the reference most frequently used by developers.
Explains the affected HTML or CSS feature neutrally and independent of any vendor.
arXiv — GEO (Aggarwal u. a., 2023)Research
arXiv is the open preprint server of science. The paper “GEO: Generative Engine Optimization” was the first to measure which changes to a page raise its visibility in AI answers.
Evidence that source references, quotations and figures measurably raise visibility in generative engines (“up to 40%”, varying by domain). An effect, not a requirement: that is why our rules on it are hints, not knock-out criteria.
SEO-Fachpresse (Ahrefs, Search Engine Land, Semrush)Trade press
Providers of SEO tools and trade media that interpret the requirements of Google and Bing and back them with their own measurements.
Experience for orientation, for instance that an llms.txt does nothing for Google. Not a standard: what is written there applies for us only once a vendor page or a measurement confirms it.
Open ahrefs.com ↗ · also searchengineland.com, www.semrush.com, semrush.com
What we measure ourselves
Much of a report we measure ourselves. Our own measurements include: recalculating every colour contrast on the rendered image instead of the source code, checking whether declining the cookie banner actually takes effect, matching the sitemap against the pages that can really be reached, and the questions no single page can answer: duplicate page titles, duplicate main headings, pages that no other page links to.
The same principle applies to every finding: An empty check is an error, not a free pass. Where nothing could be measured, that is what it says, not a tick.
