Search Signal #04
AI Overviews cut CTR 23.1% for the most exposed French sites. Plus: Google pays publishers, Cloudflare splits training from search, rank data breaks.
Vikas Kalwani
Founder, SAASY LINKS
01. Google started paying publishers when AI answers use their content
Google is piloting a Search Console feature called AI contribution. The signup copy reads: "Accumulate earnings when your content contributes significantly to AI-generated responses in participating products." It covers Gemini, AI Overviews and AI Mode, and pays on contribution rather than on clicks. Digiday reports at least dozens of publishers approached, skewing small and mid-sized, and extending beyond news. Google confirmed only that the pilot exists. No rates, no eligibility criteria, no timeline. Participants called the amounts lowball.
Why it matters: For two years the only currency Google offered for AI answer usage was a link. It has now attached a payment to contribution instead, which is a tacit admission that the traffic deal stopped working.
Action: Open Search Console on every client property this week and look for the AI contribution widget. Even at lowball rates, being in the pilot puts you inside whatever measurement Google builds next, and that access is worth more than the cheque.
02. Google broke the rank trackers, and one lost 80% of its data
New blocking against search scrapers and third-party tracking tools began around 13 September. Derek Perkins of Nozzle reported roughly an 80% drop in the data he could retrieve from Google. Sistrix says collection is running at a reduced rate. Glenn Gabe reports Ahrefs still scraping fine, Semrush slower, Sistrix far behind. Barry Schwartz's warning: be careful looking at data from these tools, because it might be wrong. Bing began showing human verification challenges the same week.
Why it matters: The 15 September volatility spike everyone is discussing sits inside the exact window where sampling broke. Movement in your charts may be measurement loss rather than ranking change.
Action: Before you explain any client dip dated 13 September or later, email your rank tracking vendor and ask what their collection rate has been since then, then put that answer in the client note. Anchor the report on Search Console performance data, which Google cannot throttle.
03. Cloudflare split AI training from search crawling, then migrated your setting for you
A new Disallow AI Training setting lets a site refuse training use while staying crawlable for search, which fixes the Googlebot collision Cloudflare created two weeks ago. Operators earn a new Accountable designation on four criteria, and Apple, Google and Microsoft qualified on the mixed-use test. It works through Google-Extended and Applebot-Extended. The sharp edge: existing settings migrated automatically on 15 September, mapping both Block and Block on pages with ads to Disallow AI Training while preserving search access.
Why it matters: Anyone who previously selected Block expecting a hard block has, since 15 September, been allowing search crawling unless they go back and re-select it. The setting changed without anyone touching it.
Action: In the Cloudflare dashboard, open the AI crawler controls on every client domain and record which state each one is actually in now: Disallow AI Training, or a full Block. Note it in the client file so the next person does not have to guess.
04. AI Overviews cut CTR 23.1% in France, and this study has a control group
Ahrefs tracked 963 French domains through aggregated Search Console data, comparing the 28 days before the 22 July AI Overviews launch in France against the nine days after. Domains where more than 20% of queries trigger an Overview lost a median 23.1% of click-through rate, and 82.2% of them declined. Panel median was minus 5.7% CTR. The detail that matters: domains under 2% exposure gained 2.3%, which is a real control group. Vendor research, a nine-day window, one country.
Why it matters: Most AI Overview CTR studies have no counterfactual, which is why they never settle an argument. The low-exposure group moving in the opposite direction is what makes this number hard to dismiss.
Action: Segment your tracked keywords by whether an AI Overview appears, then compare Search Console CTR across the two groups. Budget roughly one point of CTR loss per point of exposure rather than applying one blended decline across a whole account.
05. Two audits say AI visibility tools measure a different internet from your buyers
Researchers at Maastricht University and the University of Zurich ran 2,528 real commercial-advice queries across five surfaces. On the same query, ChatGPT and Gemini interfaces shared only 5.4% of displayed domains, and 76.7% of comparisons shared none at all. Mean domain overlap between an API and its own consumer interface was 12.0% for ChatGPT and 14.8% for Gemini. A separate study of 171,264 real conversations found Reddit and YouTube dominate ChatGPT results but are absent from Claude search entirely.
Why it matters: A blended cross-platform visibility score averages citation sets that barely intersect. Any tool sampling through an API is measuring something roughly 85% different from what a buyer actually sees.
Action: Ask every AI visibility vendor in your stack, in writing, whether they collect through the API or the consumer interface, and keep the answer. Then split every visibility report by platform and stop publishing a single blended number.
06. Six years of oversight for Google ad tech, and no AdX sale
Judge Brinkema's 106-page remedies opinion was unsealed on 16 September. The remedies are behavioural, run six years where the DOJ asked for fifteen, and apply worldwide. Google must connect AdX and DFP to Prebid with functionally equivalent capabilities, submit AdX bids to rival publisher ad servers on identical terms, stop requiring DFP publishers to also use AdX, and share real-time win and loss data with publishers. A court-appointed monitor gets access to systems and source code. AdX divestiture was rejected.
Why it matters: The structural case against Google has now failed twice in a row, in search remedies and in ad tech. Google's shape is not changing this decade, so any client narrative built on a breakup is dead.
Action: No action needed yet, because nothing is in force until final judgment is entered. Proposed final judgments are due 2 October and the remedies take effect 60 days after signing, so put early December in the calendar as the date worth watching.
07. OpenAI turned ChatGPT ads into sponsored agent conversations
After clicking an ad, OpenAI says a user can start a clearly labeled conversation with a business-sponsored agent inside ChatGPT, separate from normal answers. Buyers can ask about features, sizing and compatibility before ever visiting the advertiser's site. An Ads Manager plugin builds campaigns from natural language, generating copy and imagery from landing pages, with HubSpot and Shopify integrations. US Shopify merchants have it now, international from 23 September. Sponsored Agents is a test with selected US advertisers.
Why it matters: Amazon started buying space beneath the answer two weeks ago. A sponsored agent sits inside the conversation, where it can resolve the evaluation questions that used to send a B2B buyer to a comparison page or a docs site.
Action: List the five questions prospects ask before they book a demo, usually integrations, pricing tiers, migration effort, security posture and support model. Make sure each has a specific, current, crawlable page, because a competitor's agent answering them first means your content never gets retrieved.
08. Structure moves AI citations, and source order moves them far less than claimed
The CITECHOICE audit replayed 103 verified source pairs drawn from real agentic search transcripts, crossing source order against structured versus prose renderings. Structured rendering, meaning headings, lists and tables, added 0.50 citations per answer with no rise in total citations, so it redistributes credit rather than creating it. The observational rank 1 versus rank 5 citation gap of 42.3 points collapsed to 7.9 under controlled reordering. And 15% of cited or not-cited decisions flipped on a fresh run.
Why it matters: Most of the position effect in circulation is confounded, so rank-position citation correlations should come out of your pitch decks. The 15% flip rate means a single-run visibility reading sits inside the noise.
Action: Set your prompt tracking to at least five runs per prompt and report the median rather than one reading. When a client asks why a mention disappeared this week, check the re-run before you write the explanation.
09. Google stripped free Shopping listings out of the EEA in two days
Free product listings and the popular products carousel disappeared from EEA results on 17 and 18 September, with practitioners measuring 90 to 100% declines in Germany, France, Belgium, Sweden and the Netherlands. The space went to comparison shopping services. Ginny Marvin, Google Ads Liaison, said the removal followed from Digital Markets Act requirements. The 60-day compliance clock from the EUR 890 million fine in July ran out on 21 September. Detected through rank trackers, not through any Google announcement.
Why it matters: DMA compliance is now removing surfaces outright rather than reordering them, and doing it without notice. If it can happen to free Shopping listings, it can happen to any unit Google owns in Europe.
Action: If a client sells into the EEA, read Google's regional differences in Search experience documentation and check the eligibility criteria for the aggregator and supplier units. That placement is live, documented and almost entirely uncontested.
10. Google Discover is testing an AI summary in front of the publisher link
Tapping Dive deeper in Discover returns, in Google's Robby Stein's words, a short topic overview with prominent links to related stories, community reactions and original reporting, rather than opening an article. The summaries carry a disclaimer that they can make mistakes. Google says it will test the experience on videos to start and experiment with multiple designs in the weeks ahead. It was announced late on a Friday through a post on X rather than a blog.
Why it matters: Discover is the last large Google surface still sending publishers real volume, and an AI layer has now appeared in front of it. Videos first means there is time, but not much.
Action: No action needed yet, this is a test on one format. But if any client gets meaningful Discover traffic, split it into its own view in Looker Studio this week so you have a clean pre-period when this rolls out properly.
11. A legacy robots rule may now block more than you think
Google changed one sentence of crawler documentation on 17 September. Crawling preferences addressed to the Mediapartners-Google user agent used to affect Google AdSense. They now affect various ad-related Google products, with AdSense and Google Ad Manager given as the two examples. Google does not say whether Google Marketing Platform, Local Services Ads or AdMob fall inside the new scope, so the list is open-ended by Google's own wording. Roger Montti spotted the change at Search Engine Journal.
Why it matters: A Mediapartners-Google disallow was a common copy-paste on sites that never ran AdSense. It may now be blocking ad products the site actually uses, and it will surface as a mystery six months from now.
Action: Search every client robots.txt for Mediapartners-Google and remove the disallow on any site that is not deliberately excluding ad crawling. It takes ten minutes across a book of clients.
Also worth knowing
- Volatility twice, confirmed neither time. Movement from 4 September appeared to revert on 13 September, then spiked again on 15 September, and the Google Search Status Dashboard still carries no September 2026 entry at all.
- Search Console lost a day of crawl stats. The 15 September data is missing across all profiles, and while Google has backfilled similar gaps before, it has not said it will this time.
- AI brand prominence tracks search interest, not market share. A Northwestern paper built 1,200 recommendation lists across six models and found prominence correlated with Google Trends interest and online conversation rather than category share.
- Review volume moves AI local picks, star rating barely does. An NYU Abu Dhabi audit found a 3 to 5 times review-count premium among recommended restaurants but a rating premium of at most a tenth of a star.
- Chrome now measures ad load on every site. Four experimental CrUX metrics cover ad count, density and weight at the 75th percentile, and none of them are part of Core Web Vitals.
- Two dates for the diary. Google's reply brief in the search remedies appeal is due 29 September, and proposed final judgments in the ad tech case are due 2 October.
SAASY LINKS
We build the visibility this newsletter reports on
Editorial links, AI citations, and organic pipeline for B2B SaaS. Senior-led, 90-day sprints, no long-term contracts.
Book a Free Strategy Call