August was the month the retrieval layer stopped being a black box — and the month the measurement layer broke. Both matter more than any ranking you tracked this quarter.
The most consequential piece of GEO research this month instrumented ChatGPT's own data stream and came back with a structural finding: OpenAI runs its own web index, and Bing does not feed it. That index supplies essentially all search citations in Instant mode — the fast path free users get — while classic scraped Google results account for roughly 3.1% of the retrieval set.
Underneath the index sits a reading cache holding full copies of every page ChatGPT has ever fetched, keyed by URL and shared across all users and tiers. Recrawl frequency is governed by one signal: how often users actually ask about that page. Demand sets your crawl schedule, not your publishing cadence.
The number worth pinning above a desk is the gap between retrieval and reading. A page the model opens and reads is cited 74% of the time. A page retrieved but never opened is cited 7%. Being in the candidate pool and being in the answer are two different markets, and most visibility tooling still conflates them.
Switching to Think mode roughly doubles the retrieval pool, from 15.1 to 35.3 URLs per conversation, which means the same brand can be present in one mode and absent in the other for an identical question. Counter-intuitively, URLs surfaced without snippets were cited more often than URLs with them.
Confirmed rolling out on 18 August and already generating volatility chatter across the forums. Expect a multi-week rollout and hold off on drawing conclusions from partial data — particularly given what happened to reporting the week before.
Search Console's Generative AI report, along with standard Search performance, showed widespread impression and click declines starting around 12–13 August. Google acknowledged a data logging issue. If your AI-surface impressions cratered mid-month, check the date range before you write the postmortem.
Google began rolling AI-generated imagery into AI Overviews in the wild, then paused the experiment within days. Worth watching: illustrated answers reduce the pull of a thumbnail as a click driver, which changes what a cited page is actually worth.
ChatGPT draws more paid clicks from Google than any other top destination — a striking inversion for anyone still treating it as a side channel. Meanwhile the ad layer itself is visibly immature: advertisers have been observed swapping on page load, most likely an ad server defect.
A DevTools teardown found the local results inside ChatGPT resolve to Google Maps data — not Bing, not Yelp. For any multi-location brand, that puts Google Business Profile hygiene directly upstream of AI assistant visibility.
New passage-level analysis separates content that gets quoted from content that gets silently absorbed into an answer without attribution. The distinction reframes what "citation-worthy" writing means at the paragraph level rather than the page level.
Anthropic's revenue run rate passed $65 billion ahead of a reported IPO, with Claude the fastest-growing platform in the category while ChatGPT's share of generative traffic continues to compress. Single-platform visibility reporting is getting less defensible every quarter.
A useful reframe gaining traction: stop asking whether you rank and start defining the canonical set of facts you want every model to repeat about you, then audit the drift between that and what the models actually say.