Weekly GEO research: four gaps between access and influence

This issue asks which part of the path from a web page to a user-facing AI answer a site owner can actually observe. We use fetch, index signal, retrieval, mention, citation, and referral as a working chain. This is not a claim about operators’ private architecture: retrieval and source selection are usually hidden. The page therefore does not promise visibility growth or turn a technical signal into a causal effect. It combines official documents and published studies to show which conclusion each artifact can support.

Section sources:[1] OpenAI[2] Perplexity[3] Google

An HTTP log records a request, time, URL, status, and client signals. With a verified User-Agent and official IP, the request can be cautiously classified as a visit by a particular operator. The log still does not show whether the page entered an index, was selected for a specific answer, whether its text was used, whether a link was shown to a person, or whether anyone clicked it. Even timing overlap between a crawl and an answer does not establish causality. Fetch should remain a standalone access metric.

Section sources:[1] OpenAI[2] Perplexity[3] Google

OpenAI’s description of ChatGPT Search confirms that an initial question may be rewritten into one or more narrower queries. General location and enabled Memory may also affect the result. A measurement row containing only engine and prompt is therefore incomplete. A reproducible baseline should preserve the original wording, language, country or region, Memory state, surface, and run date. A personalized run is useful as a separate layer, but it must not be mixed with a clean baseline: otherwise a response change may look like a content effect when it was caused by user context.

Section sources:[1] OpenAI

Perplexity’s documentation separates blocking full or partial text access from retaining minimal URL information such as a domain, headline, or short description. This matters when interpreting robots.txt. It is incorrect to claim that a domain has completely disappeared merely because PerplexityBot was denied access. Crawl access, retained metadata, and use of page text in an answer require separate checks. An old user case about summarizing a blocked page also cannot automatically describe current behavior without checking the documentation version and running a current test.

Section sources:[2] Perplexity

A citation is a visible link or attribution to a specific source. A mention is the name of a brand, page, or entity in answer text and may appear without a link. A referral is an observable user action after the answer. Clickstream research and commercial observations show that these events diverge, but their numbers depend on market, device, period, prompts, and event definitions. Low referral therefore does not prove no influence, and higher citation does not prove a business result. A report should show the denominator and collection method for every metric.

Section sources:[4] arXiv[5] Semrush

An editorial and analytics pipeline should store operator, surface, language, region, mode, original and refined queries, URL, date, observed event, and preserved artifact as separate fields. Web logs can support fetch; Search Console and consoles can support index signals; saved answers can support mention and citation; analytics can support referral. Retrieval should be marked unknown when no reproducible observation exists. This issue has no matched cross-system sample, Russian replication, or causal evidence. Numbers from external studies cannot be transferred to Russia, mobile devices, or a particular brand without a new experiment.

Section sources:[1] OpenAI[2] Perplexity[3] Google[4] arXiv

The central GEO lesson is to stop using one composite score that hides where the process failed. A site can be accessible but not retrieved; retrieved but not cited; cited but not named; named but not generate a referral. These states require different checks and editorial decisions. The next step is a small public-safe Russian baseline with fixed prompts, two surfaces, and repeated runs. That design will produce more knowledge than multiplying pages without preserving evidence.

Section sources:[1] OpenAI[2] Perplexity[3] Google[4] arXiv

Public articles often place different denominators inside one conclusion. For example, the share of sessions with an outbound referral is a session measure, while citation share may be calculated over answers or sources. Before moving a number into your own dashboard, record unit of analysis, period, market, device, query set, inclusion rule, and collection method. When a study uses a commercial measurement stack, separate its published observation from the author’s interpretation. A percentage without a denominator looks precise but cannot support comparison between studies. In editorial copy, the limitation should sit next to the conclusion.

Section sources:[5] Semrush

The next practical test should connect observable events without pretending they form one funnel. For a preselected set of Russian and English questions, we will fix two search modes, the same dates, and several repeats. We will then compare page access in logs, the source set in preserved answers, visible mentions, citations, and referrals from tagged links. Each discrepancy will retain an answer snapshot and a classification rule. If a page does not appear, the result is unknown rather than zero: non-observation may reflect sampling, geography, or inaccessible internal retrieval. This protocol can produce one honest editorial article with an observation table and a separate account of what the experiment cannot measure.

Section sources:[1] OpenAI[2] Perplexity[3] Google[4] arXiv