Tuesday resolves three things practitioners were bracing for: the Search Console dashboard drop turns out to be a Google logging bug, a recipe-image complaint gets an on-record reply and a stopped test, and the ad bidding reset everyone argued about since June actually landed. None of the three says whether the numbers can be trusted cleanly from here.
The Drop That Wasn’t: A Logging Bug, Not A Lost Audience
Search Console’s sudden fall in impressions and clicks since August 12 looked like a real visibility loss. Google says it is a logging error, and it named the exact report where the corrupted numbers live.
- Search Console Drop Since Aug. 12 Is a Logging Bug, Not Lost Traffic The widespread drop in Search Console impressions and clicks that started around August 12 or 13 is a logging bug, not a real loss of visibility, Google has confirmed. John Mueller replied to a user on Bluesky that the team is aware of the issue and working on a fix, describing it as strictly a logging problem rather than a reflection of actual Search visibility changes. Google logged the incident on its Data Anomalies page on August 13 under Generative AI in Search, attributing the drop to a logging error and listing the issue as ongoing, not resolved. The affected report includes the Generative AI performance report, which only finished rolling out to every Search Console property on August 12, so any baseline a team built from that report’s first week has to be thrown out.
The Wait Is Over: A Test Confirmed Stopped, A Bidding Change Confirmed Live
One story from yesterday gets an on-record close. One change practitioners have argued about since June actually started this week.
- Google Confirms AI Overview Recipe Image Test Has Ended Google’s Robby Stein said on X that the AI-generated recipe-step images spotted inside AI Overviews were a small experiment the company is no longer running, and drew a line between that and the July Nano Banana feature, which only triggers when a user explicitly asks to generate an image. It resolves the story we published yesterday, when recipe site Inspired Taste documented an AI Overview generating illustrated cooking steps for its own spaghetti squash recipe with no image request from the user. Stein’s statement does not say how widely the test ran, how long it had been active, or whether Google intends to bring a similar automatic illustration format back in a different design. Four days passed between Inspired Taste’s August 13 complaint and Google’s August 17 confirmation.
- Google’s Ad Bidding Reset Begins, Two More Changes Follow in September Google started pulling budget-limited Target CPA and Target ROAS campaigns toward their stated cost targets this week, the first move in a bidding change the company disclosed back on June 15. Shopping, Travel, Demand Gen, Performance Max and Search campaigns tagged Limited by budget are in scope, including targets set at the ad group level. Two more platform shifts follow inside the same month: campaigns using automatically created assets or campaign-level broad match move onto AI Max for Search starting September 1, and Google separately removes the campaign-level language-targeting setting. An advertiser watching cost per acquisition move next month will have two or three separate platform-level causes to weigh, with no unaffected account left to serve as a clean baseline.
Inside The Black Box, Measured Two Ways: A Shrinking Fetch, A Recall Failure
One study reverse-engineers what ChatGPT actually fetches and finds it reaching fewer sites than it did in July. Another finds today’s models storing nearly every fact they are trained on and still failing to say a third of them back.
- How ChatGPT Actually Fetches, Caches, and Reads the Web RESONEO’s teardown of ChatGPT’s retrieval stack, built from canary pages and a reidentification classifier accurate about 98% of the time by its own account, found unique domains reached per paid Thinking-mode conversation dropped sharply after a late July break: from 21.9 to 15.5 at medium reasoning effort, and from 29.5 to 14.7 at the highest effort setting. Retrieved links fell in step and search fan-outs per conversation nearly halved, from 3.56 to 1.90. RESONEO’s own capture logs place the shift between July 26 and 31, before the August 6 model update, which rules that change out as the cause, though the firm does not say what did cause it. This is one researcher’s reverse engineering of an undocumented system, not an OpenAI disclosure, and it shows the same visibility now dividing among fewer sites, a concentration effect more consequential than any single ranking tactic.
- Google Study: Entity Order Trips Up AI Fact Recall A Google research paper finds frontier models encode 95 to 98 percent of the facts they are tested on, yet cannot produce a direct answer for 26 to 34 percent of those same facts once a question reverses the order in which the two entities first appeared in training. Prompting a model to reason longer before answering recovered 40 to 65 percent of the otherwise unrecoverable facts, at added computational cost, and further scaling of model training did not close the gap. This is research about model recall, not a stated ranking or citation factor, and the paper does not test how any search system selects or cites a page.
Self-Reported, Still Steep: What Six UK Charities Say AI Answers Cost Them
Six charities gave a UK publication their own traffic numbers, and the pattern lines up with independent AI Overview click-loss research, even where the same numbers show a citation count climbing.
- Blood Cancer UK Says Its Leukaemia Page Lost 53% Of Views Blood Cancer UK told Press Gazette its leukaemia information page lost 53% of its page views over the past twelve months, alongside declines of 45% for MGUS, 32% for myeloma, 28% for lymphoma, 20% for AML and MDS, and 17% for MPN, while the charity’s general blood cancer information hub held roughly steady. Over the same window, the charity’s peer support forum saw page views rise 86% and average engagement per visit rise 143%, a shift Blood Cancer UK describes as two distinct audiences rather than a like-for-like offset. Save the Children logged 303% more AI Overview appearances against July 2025 even as its clicks and impressions fell year over year, and its senior digital experience manager declined to isolate AI Overviews as the sole cause. All six charities’ figures are self-reported rather than independently audited, and the decline coincides with the rise of AI answers rather than being proven, on this evidence, to be caused by them.
Today’s Quick Hits
- Apple Adds 4,656 New IPs For Applebot Crawler Apple added 4,656 new IP addresses, 18 /24 blocks plus 3 /28 blocks, for Applebot, the crawler behind Siri and Apple Intelligence, after the published range went years without a documented update.
- llms.txt V2 Adds Formal Links To Markdown Pages Jeremy Howard’s llms.txt v2 spec defines formal link relations so agents can locate a page’s Markdown counterpart, the format’s first update since 2024. No major AI company has committed to honoring llms.txt, and Google’s Search team says the file affects neither visibility nor rankings.
- Google Lets Users Hide Gemini’s Visible AI Watermark Google added a toggle letting Gemini and Flow users switch off the visible watermark on generated images, video and music. The invisible SynthID mark and C2PA metadata stay embedded regardless of the toggle, so provenance survives even where the at-a-glance visible signal does not.