Google's John Mueller and Martin Splitt pushed back on building separate markdown shadow sites for AI agents, warning that a well-structured HTML page already serves search engines, LLMs, and humans — parallel builds create tech debt with no indexing upside.
Why it matters: GEO isn't a separate site build — optimizing the canonical HTML you already have is more effective and less risky than spinning up parallel markdown experiments.
An IQRush paper published July 11 shows that citation shares in AI tracking tools (Otterly, Promptmonitor, Peec AI) are snapshots of a continuously shifting target — a competitor gap in your dashboard could be real signal or pure measurement variance, with a new stopping rule available to determine when data is actually trustworthy.
Why it matters: GEO performance reporting needs trend lines with statistical rigor, not single-run snapshots — most dashboards overstate certainty and mislead teams acting on the data.
Two independent analyses found ChatGPT routes queries through multiple undisclosed retrieval backends (Labrador, Bright, Oxylabs, SERP) and the pipeline feeding an answer changed on 11.6% of repeated identical runs — meaning single-run citation audits may be measuring pipeline routing, not brand authority.
Why it matters: Single-run ChatGPT citation audits are structurally unreliable — repeated sampling across sessions is the only way to surface durable brand signal on this platform.
Google is experimenting with an AI Mode "Ask Anything" entry point appended to the outside of the free-listings product grid on mobile — the first time the AI prompt has been attached to a commercial shopping surface rather than embedded within it.
Why it matters: AI Mode is entering commercial discovery surfaces — product pages built for shopping SERPs now need to consider AI answer readiness, not just traditional organic ranking.