Weekly GEO research: five boundaries of AI search in September

The September cycle found no secret growth button. It clarified three boundaries: Google eligibility is not page selection, AI surfaces do not share one universal result set, and reputation signals can affect scientific-paper choice even when text is unchanged. A separate official Google Play rule concerns app availability, not AI recommendation. Perplexity also explains that a third-party model operates within Perplexity’s own search and rules. The database should therefore describe a specific surface, version, role, and event rather than AI in general.

On August 31 Google reported a global rollout of a Search Console control. An owner can allow or disallow use of the site’s material and links in AI Overviews, AI Mode, and AI Overviews in Discover. Opting out removes impressions and clicks from those features, while ordinary web search remains separate. Opting in guarantees neither a link, position, nor citation. Research records should add an eligibility_status field and keep it separate from impressions, retrieval, mentions, or clicks. Record date, domain, and surface so a later interface change can be reproduced.

The Answer Index sent the same 480 questions to Google AI Overviews, ChatGPT, Perplexity, and Gemini through an API intermediary. Its 1,919 measurable answers contained 16,069 links to 7,775 domains. The highest average pair overlap was 18.1%. This shows divergence in observed candidate or citation sets, but not the algorithm or which system is better. ChatGPT links appeared in only 72 answers, with question type affecting the result. The figures cannot be transferred to consumer apps, other dates, or Russian without new collection.

Jinadu and coauthors, in arXiv:2609.00248, tested eight language models across 250 paper-selection tasks. Titles and abstracts stayed the same while reputation information changed: conference prominence, citations, and author and organization indicators. After swapping those signals, 39.2% of choices changed; an instruction to use content only reduced but did not remove the effect. This is causal within a fixed candidate list. It does not prove that domain authority universally ranks web pages and does not show that models found the sources on the open web.

The Google Play rule effective August 31 requires a current target API for new apps and updates, while published apps must meet the minimum for newer devices. An app that is not updated may stop appearing to a new user on a newer phone; a limited extension is available. This is a loss of availability and potential Google Play installs, not evidence of a ranking drop or a new recommendation factor for ChatGPT, Gemini, or Perplexity. Reports should name Google Play precisely and keep compatibility separate from AI visibility.

On September 4 Perplexity clarified that third-party models are connected to its own search, links, tools, instructions, and limits. GPT inside Perplexity should therefore not be treated as ChatGPT: the language model and search surface are different entities. Comparisons should store provider, answer_model, search_mode, and displayed sources. This does not show an advantage for any site, but it prevents falsely merging measurements under a model name. The same principle applies to different modes of one product and intermediary APIs.

No source this week revealed a universal ranking factor. The Answer Index is limited to one date and an intermediary; the reputation experiment to scientific selection; Google Play to availability; and Perplexity to an architectural boundary rather than ranking. The next cycle should measure the same query panel across modes, preserve versions and full answers, and count mentions, citations, referrals, and installs separately. For the site, publish clear pages and evidence without promising inclusion in answers. Every new conclusion needs a date, scope, and direct source.

Even a detailed weekly article remains a snapshot: platforms change models, interfaces, sources, and access rules. Conclusions must therefore be date-bound and never become a promise that a page will appear in an answer. A strong knowledge base is valuable because it preserves applicability limits. It records not only positive cases, but also missing evidence, source conflict, intermediary error, and the need for replication. Such an honest archive is more useful to editors and researchers than smooth copy without a data passport.