GeoAeoAleIndependent publication about AI search
Write an article ↗Research, observations and practical experience about visibility in machine-generated answers.
Google described dedicated reporting for AI Overviews and AI Mode. It helps observe impressions and clicks inside Google Search, but it does not reveal a ranking score, the complete source set, or a causal link between a page change and an answer appearance.
Perplexity documents PerplexityBot for search crawling and Perplexity-User for fetching a page in response to a user request. A matching User-Agent and official IP help classify the fetch, but neither role alone proves that the page appeared as a link in an answer.
This week’s central finding is that technical access and ordinary search eligibility remain foundations for AI discovery, while fetch, impression, citation, and referral are different events. New reporting is useful only when every metric has its own definition, period, and data source.
Citability starts not with a special AI tag, but with an accessible page, a clear claim, and nearby verifiable evidence. Name an entity consistently, define its boundaries, and connect it to a date, method, and primary documentation. Then check whether an independent reader can find the claim in HTML and see exactly what the source supports. This improves verifiability, but does not guarantee retrieval or a citation in any particular answer.
GEO, AEO, AIO, and AI SEO are not one official standard. Treat them as labels for different jobs: page access, semantic understanding, evidence, passage retrieval, answer quality, citation, and referral. Google explicitly recommends ordinary search fundamentals for its AI features and does not require special AI markup. The editor’s job is therefore not to guess a model’s secret format, but to make claims verifiable, pages accessible, and every change measurable. This works for a small project as well as a large newsroom: it shows which part of the chain changed, what evidence supports the result, and where the platform operator’s control begins.
An AI answer is not a transparent search log. We can observe a bot request, a published answer, and displayed links, but one citation cannot reveal the complete retrieval trace, source weights, or internal reasoning. Google describes query fan-out as a possible way to split a complex question into additional searches; that is a documented possibility for one system, not a universal rule for every answer engine.
Measure AI visibility as separate observable events: URL accessibility, fetch, mention, citation, citation relevance, and referral. Predefine the prompt panel, language, country, mode, date, and scoring rule; preserve the full answer and URLs. Report numerator and denominator, and label causal conclusions as hypotheses. This shows surface change without promising an outcome controlled by a platform.
Core material should be available in server HTML, URLs should return truthful status codes, and canonical, hreflang, sitemaps, and robots should match the real site structure. This removes barriers for compliant crawlers. It does not guarantee crawling, indexing, retrieval, or citation; private data must be protected with access control, not robots.txt.
A web-server log proves that a resource was fetched by a particular request. It does not prove retrieval, use of a fact, a displayed citation, or a user click. Each claim requires its own chain of preserved artifacts.
Public documentation for seven surfaces describes different access and search modes but no shared source-selection formula. Measure accessibility, discovery, links, and citation as separate events.
To determine whether documentation meaning changed, a raw HTML diff is insufficient. Snapshot an open URL, normalize noisy elements, compare extracted text and structure, then manually review candidate meaningful fragments. This item reports a methodology and pilot protocol, not a quantitative result: numeric observations must be collected by actually running the experiment.
Google separates eligibility for AI answers from actual selection, different systems use different domains, and a scientific experiment shows reputation effects in a narrow task. Google Play adds a separate app-availability boundary, while Perplexity separates the model from the search surface. These are measurement boundaries, not universal promotion recipes.