Comparing the source code of two page versions is not enough to tell whether the documentation changed in meaning. Save both versions, remove known technical noise from the comparison, then compare the visible text, headings, and links. A person should review every potentially important difference. This article explains how to run the experiment; it does not claim numeric results before the experiment has actually been run.
Public documentation for seven interfaces and modes describes different access and search modes but no shared source-selection formula. Measure accessibility, discovery, links, and source link as separate events.
A web-server log proves that a resource was requested by a particular request. It does not prove source selection, use of a fact, a displayed source link, or a user click. Each claim requires its own chain of preserved artifacts.
Here GEO means “generative search optimization.” 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 interface or mode. These are measurement boundaries, not universal promotion recipes.
Five signals this week show that AI visibility is several different events: Google may separate third-party content from a domain’s reputation, Alice can receive products through a commerce system, Grok Bot searches inside X, and large and field measurements separate source links from referrals. None is a universal recipe.
Here GEO means “generative search optimization.” Google Preferred Sources is a personalized lever for an already interested reader, not a general ranking factor. Grok 4.6, GEO-text detection, Reddit decline, and connector errors show why measurements must be versioned and checked.
Here GEO means “generative search optimization.” The week did not reveal a universal growth factor, but it established four practical limits: conversation context changes answers, one run does not cover all sources, more source links can reduce evidential quality, and knowing where a page came from is not a recipe for producing similar text.
Being included in an answer, ranking among named named subjects, evidential correctness, and a user action cannot be reduced to one metric. This week’s studies support measuring the stages separately and comparing page types within a specific intent.
Site-access rules should name the specific crawler and the purpose of its visit. Allowing or blocking one crawler says nothing about every use at once: search, model training, opening a page for a user, and advertising are separate cases.
A row labelled only “Perplexity” or “Google” is not a reproducible experiment. Record the exact interface and mode, the answer model, the device, whether personalization was active, the repeat number, and what was observed. A mention, source link, ad placement, and user visit remain different outcomes.
A visit from an AI-system bot does not prove that the system used the page in an answer, linked to it, or sent a visitor. Each event needs its own evidence.
GeoAeoAle