Weekly GEO research: four boundaries of measuring AI visibility

This week asks which fresh publications actually change how AI visibility should be measured and how answer changes should be explained. The material combines four primary studies published or updated in August 2026. It is not a platform ranking and does not promise source links. Each conclusion is tied to one experiment, its sample, and the material it actually measured. We count five different events separately: a bot request, finding the page, naming the brand, showing a source link, and a person visiting the site. This frame matters because the word visibility often mixes technical access, source selection, and human behavior.

The August 3 study compared multi-turn conversations while keeping the final user message identical. The model received either the full conversation, only the final question, or a short reconstruction of history. In 44.7% of observations, the full conversation and the question without history produced materially different answers; a judge considered some differences capable of changing user action. This does not show that longer history is always better and did not measure web search. It does make comparisons invalid when one run continues a dialogue and another is a clean question. A test record should store conversation_state and the history-transfer method.

In the August 13 medical sample, three production chatbots selected primary studies for 20 clinical questions and were compared with an expert Cochrane set. One answer found 39.2% of included works on average; combining four repeats found 74.2% at least once. A researcher-framed role produced broader coverage than a patient role. These percentages belong only to twenty medical questions and cannot be transferred to brands or commercial search. The transferable methodological lesson is to count the stable core, cumulative coverage, and long tail separately. Missing from one answer does not prove that a system lacks the source.

An August 11 laboratory study repeatedly rewrote documents that had already been found to make a generative system cite them more often. Under repeated interventions, the texts accumulated unsupported claims; a mechanism rewarding verifiable content produced a better balance on the authors’ test set. This was not a real-world audit of ChatGPT, Google, or Perplexity and does not reveal their algorithms. An editorial metric should separate whether a page was among the pages considered, displayed links, support for a particular claim, and absence of new errors. A higher share of answers with source links is not success if the source becomes harder to verify.

A Google AI Overview audit from May 25–29 used 2,597 queries in topics where errors can affect health, safety, or money. In that sample, pages labelled by a commercial service as AI-generated received source links more often. The study does not establish causation: such detection is disputed, other possible causes were only partly controlled, and one author's connection to the detector company calls for caution. The gap mainly appeared among pages outside Google's first 100 ordinary results. The origin of the text can be tested as one possible explanation, but it does not justify generating articles merely to gain links from AI answers.

S01 measured generation without web search; a search-enabled replication is needed to test whether history changes query rewriting and link sets. S02 concerns medical questions, S03 is a simulation, and S04 is an observational audit with sensitive classification. None reveals a universal ranking factor. The next reproducible test should preregister clean versus continued questions, run at least four repeats, preserve full answers, and separately verify source links and supporting passages. Results should include absolute counts, period, language, sample, and confidence level.

Before every comparison, record the exact prompt, language, country, service mode, and conversation state. Preserve the complete answer, every link, run time, and page version. Record five results separately: a bot request, a brand mention, a displayed source link, whether that link supports the claim, and a user visit. If the sources differ, do not call it a ranking change without a matched control. Readers need the underlying count, exclusions, and boundary of applicability, not only an attractive percentage.

The week’s main editorial error is treating a single observation as system knowledge. A useful publication shows the question, condition, source, and limitation. If a conversation continued, that is part of the test; if an answer was repeated, that is part of the method; if a link appeared, check which claim it supports. This format lets readers separate new knowledge from an attractive assumption. It also preserves history: the next cycle can confirm, refine, or challenge the conclusion without rewriting the old date or hiding disagreement.

For monitoring, an answer should be checked step by step: was the page considered, found, named, linked, actually supported the statement, and visited by a person. These events diverge: a page may be considered without being cited; a link may point to a domain without supporting the sentence. The report should preserve a record of each event, the full answer, conversation conditions, and evidence passage rather than one visibility score. A repeatable protocol turns a weekly note into a cumulative database: the next run can be compared with the earlier observation without rewriting it.