Improve how a brand or page is retrieved, cited, mentioned and recommended in AI-generated answers. Classic SEO, crawlability, useful content and verifiable authority form the foundation; AI-specific structure is only an additional layer.
Evidence rules
AI search behavior changes quickly, and many studies are observational. Before current percentage figures, crawler names, platform backends, visibility increases or mechanisms:
- check current primary documentation or the original study;
- state source, date, platform/model, region and method;
- do not present correlation as a ranking factor or causality;
- do not promise citations, rankings, recommendations or traffic impact;
- for Google AI features, base your work on current Google guidance: normal search eligibility and user-oriented SEO; do not claim special markup or an AI text file is mandatory.
No invented ranking factors. Do not describe structural patterns, llms.txt, schema, mentions, author profiles or crawler permissions as a confirmed AI ranking factor unless the respective platform currently and unambiguously documents this.
Context and visibility levels
Record domain, priority pages, brand, offering, target audience, region, language, important prompts, platforms, competitors, baseline visibility and goal. Do not search private directories, credentials or personal data.
Distinguish:
- Retrieved: The system reads the page while forming the answer.
- Cited: The page appears as a source.
- Mentioned: The brand appears in the answer text.
- Recommended: The brand makes it into a selection list.
Useful content can be cited without the provider being recommended. Recommendations can depend on verifiable, web-wide confirmation through reviews, communities, trade media, videos, podcasts or analysts; this is not a guarantee.
Audit process
1. Reproducible baseline
| Prompt | Platform/model | Date/region/language | Answer | Retrieved/cited/mentioned/recommended | Sources | Assessment |
|---|
Use repeated checks to make volatility visible. Do not derive stable visibility from a single answer or account.
2. Technical suitability
Check reachability, status code, canonical, robots, sitemap, indexability, rendered HTML, semantic structure, login/consent barriers, title, headings, authorship, date, internal links and structured data. Structured data must accurately reflect visible content and comply with current platform rules. Take visual, DOM and accessibility tree usability into account.
3. Check crawler guidelines precisely
Do not make a blanket recommendation to "allow all AI bots". Distinguish current user-agent tokens and purposes using primary sources:
- Search/indexing crawlers may affect discoverability;
- user-triggered fetchers retrieve content in response to a specific request;
- training crawlers govern model training and do not necessarily control search citations.
For example, current OpenAI documentation distinguishes a search bot, a user fetcher and a training bot; Google separates Google-Extended from normal Search crawling. Check Anthropic, Perplexity and others against their current behavior and documentation. Weigh privacy, licensing, server load, discoverability and citation. Never change robots.txt without approval and validation.
4. Extractability and Usefulness
Lead with direct, correct answers where this is natural; use descriptive H2/H3 headings, focused paragraphs, genuine comparison tables, process lists, self-contained statements, definitions, limits, dates, units and examples. Set landmarks, headings and alt text correctly. Do not create fragmented "AI bait" pages, doorway variants or large volumes of thin content.
5. Authority and Evidence
Prefer primary sources, give statistics a date and methodology, show practical experience, identify real authors and relevant qualifications, distinguish facts/estimates/opinions/vendor statements, and remove exaggerated or outdated claims. Do not invent statistics, quotations, customers, reviews, awards, qualifications or sources.
6. External Consensus
For recommendation goals, examine reputable review platforms, professional communities, industry reporting, earned editorial coverage, YouTube, podcasts, transcripts and consistent organization data. Recommend only authentic participation; no fake reviews, undisclosed placements, Wikipedia manipulation, forum spam or fabricated evidence.
7. Machine-readable Supplements
llms.txt, public Markdown documentation, transparent pricing/specification files or feeds should be treated, at most, as measurable accessibility/discovery experiments. Keep them consistent with visible canonical content, introduce them only when maintainable, disclose no confidential data or protected material, and never describe them as a Google requirement or confirmed ranking factor. Emerging protocols remain experimental until first-party adoption is substantiated.
Content Patterns
- Definition: direct definition, scope, boundaries and benefit.
- Process: numbered, concrete actions with completion criteria.
- Comparison: verifiable criteria, suitability, limits, date, methodology and sources; do not automatically put your own offering first.
- Evidence Block: claim, primary source, result with date/scope, limitation and practical implication.
- FAQ: real user questions; FAQ markup only for visible content and uses currently permitted by the platform.
Prioritization and Measurement
Assess business benefit, strength of evidence, effort, deception/policy risk and measurability. Order: crawl/index/rendering errors; false claims; missing core information; weak structure/accessibility; missing evidence/authorship; external consensus; experimental files.
Three Measurement Methods:
- prompt tracking with prompt, date, model, region, language, sources and assessment;
- voluntary self-reporting of how users found the offering;
- sales/conversation evidence only where lawful and with the necessary consent.
Also monitor branded searches, qualified visits, conversions, bot logs and cited landing pages. Direct traffic or branded searches can obscure influence; do not automatically infer causation from them.
German/DACH Context
The default assumption is Germanyif no target market is specified. Prompts, sources, models, language variants and response region must be reproducibly documented for Germany. GDPR/TDDDG, copyright, trademark and personality rights are specific review points for logs, tracking, training/crawler decisions, source reuse and personal content. No legal advice or legal guarantees.
- Germany: consider German specialist terminology, search intent, trust in providers/authors and primary sources without inventing ranking factors from them.
- Austria: test Austrian terminology, source landscape, market and legal context separately.
- Switzerland: investigate German, French and Italian prompts, Swiss sources, the market and the revised Swiss Data Protection Act context separately.
A result for Germany does not automatically transfer to Austria/Switzerland. Platform answers may vary by account, model, time, language and location.
Output Format
- Scope, target market and methodology
- Evidence inventory with source versions and dates
- Baseline table for each prompt/platform/region
- Findings at each visibility level
- Findings on structure, authority and presence
- Platform-specific limitations
- Prioritized actions with owner, effort, impact hypothesis and validation
- 30/60/90-day plan
- Measurement table
- Uncertainties, experimental assumptions and claims requiring a fresh check
Checklist
- Current primary sources and their dates are stated
- Retrieval, citation, mention and recommendation are distinguished
- Repeated measurements and regional/language parameters are documented
- No invented ranking factors or causal relationships
- Crawlers are assessed by purpose rather than as a single group
- Technical suitability and rendered content are checked
- Claims, authors, statistics and sources are verified
- No doorway pages, fake reviews or spam measures
- Experimental files are clearly labeled as experiments
- Germany is identified as the default assumption, with Austria/Switzerland treated separately
- No website, crawler or analytics changes without approval
Origin and Differences
German/DACH adaptation of the MIT-licensed original skill by Corey Haines; the homepage and commit remain unchanged in the metadata. Compared with commit 7868cb9251fad80a73d26e488a5ad5f6c4a9f335 , the main text and description were written entirely in German, Germany was added as the default assumption, and Austria and Switzerland were addressed separately. Unsubstantiated ranking factors were explicitly excluded; crawler purposes, volatile prompt measurements, experimental machine-readable files, rights and DACH localization were clarified. The original external tool-register reference remains removed, and the distinction between training crawlers, search crawlers and user fetchers is preserved. Central evidence: docs/UPSTREAM-AENDERUNGEN.md.