Google spent Friday insisting its AI features send billions of clicks to the open web, then declined to publish the one number that would settle it. Over the weekend rankings moved again with no confirmation, a security screen turned out to be quietly deleting pages from the index, and a copyright complaint put a price on training data Google never disclosed. Here is what moved.
Google Says the Clicks Are Fine, and Still Will Not Show the Data
The company put a figure on AI-driven traffic for the first time, and shipped one feature that might actually send clicks. Neither came with numbers you can check.
- Google Puts a Number on AI Search Clicks, Not a Way to Check It Nick Fox says AI features in Search now send billions of clicks to websites every week, with no baseline, denominator, or methodology attached. Search Console’s AI reports show impressions but not clicks, which is precisely the figure that would confirm or kill the claim. Independent work still points the other way, at 68 percent zero-click and a 42 percent CTR hit where AI Overviews appear.
- Google’s AI Overviews Now Surface Top Stories, and Loyalty Wins The Top Stories carousel is fully live inside AI Overviews on US mobile, one of the few changes built to push clicks out to publishers rather than absorb them. It also leans on Preferred Sources, so visibility is partly pre-earned by readers who already chose you. Treat newsletter signups and app installs as ranking inputs now.
Rankings Moved Over the Weekend, Unconfirmed as Always
More than a dozen trackers lit up on Saturday, but the more useful signal is what the week around it was doing.
- Google’s Weekend Ranking Blip Is the Least Interesting Part of This Story An unconfirmed update landed Saturday July 18, yet the entire prior week ran hotter than the July 11 movement, and some practitioners now say July 11 is reversing rather than settling. Forum reports ranged from traffic up 400 percent to still down 33 percent on the same weekend, which is why that chatter is noise and not measurement. Pull your own query-level data before you touch anything.
Your Own Stack Can Quietly Delete You From Google
Two stories about configuration you already own, both failing silently, both invisible to anyone checking the site by eye.
- A Bot-Check Screen Can Quietly Replace Your Page in Google John Mueller warned that bot-protection interstitials get served to Googlebot and indexed instead of your content. Because the same screen appears across many sites, Google can treat your page as a duplicate and hand the canonical to someone else’s domain. Human visitors never trigger it, so nothing looks broken: check the page indexing report and the rendered HTML in URL Inspection.
- Google’s Old NotebookLM Crawler Token Expires in August The fetcher is now Google-GeminiNotebook, and the old Google-NotebookLM string only works through August 2026, so every hardcoded rule built on it stops matching. The trap: user-triggered fetchers ignore robots.txt entirely, so a disallow rule was never blocking anything. Discover Sources can pull ten sources per query and return no referral at all.
Where the Money and the Credit Actually Land
A lawsuit over what trained Gemini, a bidding change advertisers do not trust, and a booking path that quietly moved.
- Publishers Sue Google, Cite Internal Note Calling Gemini Training Risky Hachette, Cengage, Elsevier and novelist Scott Turow filed in the Southern District of New York, alleging books supplied to Google Books, Play Books and Scholar were repurposed to train Gemini. The complaint quotes an internal document calling it highly problematic with exposure of tens to hundreds of billions. The remedy that matters is the demand for a full accounting of training-data sources.
- Google Ads Bidding Shift Sparks Fight Over What Predictable Means A mid-August change to bidding for budget-limited target campaigns has advertisers arguing about Google’s own wording. Ginny Marvin says it only touches budget-constrained campaigns using a target; Maggie Humphrey reads more predictable performance as trading efficiency for spend. Log cost, conversion value and realized ROAS now, because the pre-change baseline window closes within a month.
- Google Quietly Pulls Its OpenTable-in-Maps Help Page Google removed the help document for booking OpenTable tables through Maps and redirected it rather than 404ing, while stripping the link from related pages. Nothing is confirmed, so this is a documentation removal implying a rollback. For restaurants the stake is attribution: a reservation completed inside Maps and one completed offsite credit very differently.
Today’s Quick Hits
- Search Console’s Validate Fix Speeds Recrawls, Not Verdicts John Mueller says the button samples flagged URLs and, if clean, queues the rest for a faster recrawl. It does not judge whether you fixed the right thing. Validation runs per issue, so one bad sampled page fails the whole check.
- ChatGPT Hits Record 640 Million Monthly Users Similarweb estimates 640.68 million monthly active users in June, up 2.2 percent on May and 24 percent year over year, with only three declines in 37 months. That is the growing denominator behind ChatGPT citation visibility, though it is one vendor’s panel and not audited platform data.