Weekly AI search research: four gaps between access and influence

This issue asks which parts of the path from publishing a page to an AI-generated answer a site owner can actually observe. Owners may see a bot visit, search appearance, mention, displayed source link, or user click, but the system's internal source selection is usually hidden. A technical signal therefore cannot be treated as the cause of visibility growth. We compare official documents and research to show what each observation can and cannot support.

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

A server log records a request, time, page address, response code, and technical visitor details. If the crawler name and IP address match the operator's official data, the visit can be attributed to that operator. The log still cannot show whether the page appeared in search, contributed to a particular answer, was linked to, or received a click. Even matching times do not prove a causal connection. Bot visits should therefore be measured separately from mentions, links, and user visits.

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 starting measurement should preserve the original wording, language, country or region, Memory state, interface or mode, and run date. A personalized run is useful as a separate layer, but it must not be mixed with a clean starting measurement: 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 source link, a brand mention, and a user click are different events. A brand may be named without a link, and a displayed link may receive no click. User-behavior studies show that these measures diverge, but the numbers depend on the market, device, period, prompt set, and counting method. Few clicks do not prove that there was no influence, while more source links do not prove a business result. Every measure needs a clear total count and collection method.

Section sources:[4] arXiv[5] Semrush

For every run, record the system, exact interface and mode, language, region, original and refined queries, page address, date, observed event, and supporting evidence. Server logs can confirm a bot visit; Search Console can show search appearance; saved answers can confirm a mention or displayed source link; web analytics can show a user click. If the system's internal use of a page cannot be observed, say so. This issue has no comparable sample across all systems, Russian replication, or causal evidence. Results from other 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 lesson for generative-search optimization is not to collapse everything into one score. A site may be accessible to a bot but absent from an answer, appear without a source link, receive a link without a brand mention, or be mentioned without receiving a click. Each case needs a different check. The next step is a small public Russian-language study with a fixed prompt set, two interfaces, and repeated runs. It will provide more insight than mass-producing pages without preserving evidence.

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

Public articles often compare percentages calculated from different bases. The share of sessions that produced a click is based on sessions, while the share of answers with a source link may be based on answers or sources. Before adding a number to your dashboard, record what was counted, over which period, in which market and device, with which query set and inclusion rule. If a commercial service collected the data, separate the observation from the author's interpretation. A percentage without a clear total count looks precise but cannot support comparison. Put the limitation next to the conclusion.

Section sources:[5] Semrush

The next test should connect observable events without pretending they form one funnel. For a fixed set of Russian and English prompts, we will use two search modes, the same dates, and several repeats. We will compare bot visits in server logs, sources in saved answers, mentions, displayed links, and visits from tagged links. For every mismatch, we will preserve an answer snapshot and the assessment rule. If a page does not appear, the result will be 'no data' rather than zero: sampling, geography, or hidden source selection may explain the absence. This protocol can support an honest article with an observation table and a clear account of what the experiment does not measure.

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