Search Signal #06
AI Overviews now appear on 82.91% of branded searches, up from 61% in July. Plus: the spam update's second wave, and new rules on AI content.
Vikas Kalwani
Founder, SAASY LINKS
01. AI Overviews now appear on 83% of branded searches
Ahrefs analysed 232.8M US desktop SERP crawls from July to September, 92.5M of them branded keywords. AI Overviews appeared on 82.91% of branded results in late September against 61.3% in July. On a tracked panel of 100 major brands, 76.2% of those AI Overviews rendered in position one, above the brand's own homepage. Most-cited sources: Wikipedia 71.9%, YouTube 38.6%, LinkedIn 22.8%. DemandSphere logged the same inflection independently.
Why it matters: Your own brand-name SERP is now, four times in five, an AI summary assembled mostly from Wikipedia, YouTube and LinkedIn rather than from your site. Both vendors measured presence, not citation, and neither measured clicks.
Action: Run your brand-name query today and check three things: whether you are cited at all, whether the description is accurate, and whether a competitor appears. Misrepresentation on a branded query is the fastest thing to escalate internally.
02. The September spam update ran a second wave
Announced 24 September with a two-week rollout window, the update landed in two distinct waves. Barry Schwartz dated the first to 25 to 27 September and the second to 30 September. Glenn Gabe: "I received emails from several site owners explaining they saw a huge drop right on 9/30." As of today the Search Status Dashboard still carries no end time. Lily Ray reports early signs of highly templated, programmatic pages being hit.
Why it matters: A site that looked untouched on 28 September may have been hit on the 30th. Anyone who already diagnosed and declared themselves clean needs a second look. The targeting read is practitioner inference, not a Google statement.
Action: Annotate 25 to 27 September and 30 September to 1 October as separate events, not one blurred fortnight. Keep content and technical changes frozen until Google posts an end time.
03. Google rewrote its content documentation three times in two days
Three unannounced edits. On 1 October, the generative AI guidance gained a hard line: "it is critical to manually factcheck and review all AI-generated content." The same day, the authorship section added that "fabricating creator profiles (such as by using AI-generated headshots, made-up names, or false credentials) is a form of deception". On 2 October, Google defined main content and named the four attributes raters assess: effort, originality, talent or skill, and accuracy.
Why it matters: The fake-author clause creates immediate audit work. Any site running AI-generated bylines, generated headshots or invented credentials is now in documented deception territory. Rater guidance describes evaluation, not a ranking factor.
Action: Open your author pages this week and verify every byline is a real, reachable person with real credentials. Where one is not, put a real author behind it or remove the byline. Two hours per site.
04. Nine of 40 UK B2B software firms were never named in any AI shortlist
Recrawled asked 1,608 questions across ChatGPT, Gemini, Google AI Overviews and AI Mode, covering 40 UK B2B software companies in 33 buyer categories. Nine were never recommended on any platform. The naming rate falls as intent sharpens: 63% on UK-specific questions, 43% with no country, 13% on "alternatives to [market leader]", and 0% on "[leader] versus [rival]". Correlation between review-site presence and being named was minus 0.06.
Why it matters: The 0% on head-to-head comparisons is bottom of funnel, where a buyer already has a shortlist and a budget. The minus 0.06 review correlation is the cleanest counter to anyone who thinks G2 volume alone buys AI visibility. All data was collected on a single day, which the authors concede is a snapshot.
Action: Build your comparison-prompt set and run it manually across ChatGPT, Gemini and AI Mode. Where you are absent and a competitor is present, read the cited sources. That is your placement target list.
05. Gemini started appending UTM parameters to outbound links
Spotted around 30 September and acknowledged by John Mueller: "Nice. I see it too." The reason it matters sits in a community comment: UTM survives the in-app webview and shows up in raw server logs, where referrer headers largely do not. That is why a large share of assistant referrals has been landing in GA4 as Direct. Google has published nothing: no parameter values, no trigger conditions, no word on AI Overviews or AI Mode.
Why it matters: The first genuinely new measurement surface in months. If the UTMs are consistent, Gemini becomes separable, which also means your Gemini number is about to jump for reasons unrelated to performance.
Action: Grep this week's server or Cloudflare logs for utm_source values containing "gemini", read the actual values, then build the GA4 segment from what you find rather than from the article. Annotate 30 September.
06. Two million AI citations: the AEO checklist collapses under domain controls
Discovered Labs analysed roughly 2.1M prompt, engine and cited-URL tuples across ChatGPT, Claude, Google AI Overviews and Gemini, collected across 19 B2B SaaS workspaces. The dominant page-level predictor is prompt-content alignment, the token overlap between page and prompt corpus, at beta +0.37. With domain fixed effects applied, FAQ sections came in at +0.07, TLDR blocks +0.05, author bio +0.02 and non-significant, and Core Web Vitals showed no effect at all.
Why it matters: Bolting an FAQ block and a TLDR onto an existing page does roughly nothing once you compare pages on the same domain. What survives is domain authority and whether the page's language matches the language buyers actually use. Vendor authors, observational, and the alignment effect halves out of sample.
Action: Take your tracked prompt set, extract the vocabulary, and compare it against the body copy of pages you want cited. Where the page uses internal product naming and prompts use category language, that gap is the rewrite brief.
07. Agent accessibility beat every AEO proxy across 37,927 journeys
A matched-pairs study of 1,056 live businesses tested whether AI agents could actually read the site: JavaScript-free content survival, structured data, reachable pricing and API docs, bot-control posture. Agent-ready sites against the rest: grounded-answer rate 0.778 against 0.555, clear recommendation 0.204 against 0.106, on-site block rate 0.158 against 0.335. The finding that should reorder a roadmap: the standard AEO proxies went statistically inert once accessibility was controlled.
Why it matters: An agent that cannot fetch your pricing page falls back to secondhand sources and hedges. Read this alongside item 06: on-page formatting dissolves under domain controls, and off-site proxies go inert once an agent tries to read the site. Preprint, and the treatment variable is the authors' own commercial ranker.
Action: Curl your pricing page and docs with a plain user agent and with JavaScript disabled, and check whether prices and plan names are in the HTML. Then check your WAF for blanket AI-agent blocks a security team added without telling marketing.
08. A judge dismissed the leading AI Overviews antitrust suits
Judge Amit Mehta dismissed the antitrust claims brought by Penske Media and by Chegg over AI Overviews. His holding: plaintiffs "failed to plead any actual agreement whereby Defendants promised to 'sell' Plaintiffs any specific amount of traffic", and "an expectation is not an agreement. It is simply how a general search engine works." Reporting is split on whether dismissal was with prejudice, and no appeal has been announced.
Why it matters: There is now a US federal ruling that Google owes nobody a given volume of referral traffic. That kills the quiet hope that litigation will put the clicks back. Note the contrast: the same week, a different judge sent roughly $1.72bn of ad tech damages to a jury. Courts will compensate for ad market conduct and not for lost traffic.
Action: Use this when someone wants to defer AI visibility budget pending "how this all shakes out". It has shaken out. The only live route to enforceable attribution rules is regulatory, which currently means the UK CMA.
09. Ahrefs changed how it models traffic and AI volume
Two changes in one release note move numbers already reported. The Adaptive Organic Traffic Model assigns individual CTR weights to 15 SERP feature types including AI Overviews, replacing uniform weighting, and becomes the default once testing concludes. AI Adjusted Volume replaces Google search volume inside Brand Radar with platform-specific estimates. Separately, the Brand Radar prompt dataset roughly doubled from 14M to 30M prompts.
Why it matters: Q4 numbers will not be comparable to Q3, and the cause will be the tool rather than performance. A prompt database doubling moves share of voice on its own, even when visibility is flat.
Action: Export your current static-model traffic estimate and Brand Radar share of voice before Adaptive becomes the default, so you have a bridge. Then put one sentence in the October report explaining the measurement basis changed and performance did not.
10. ChatGPT Ads got conversion reporting and seven more markets
Five changes shipped on 30 September: bulk product-feed campaign creation from a spreadsheet, a Products tab showing review status and rejection reasons, separate internal and public brand names, launch across Southeast Asia and Taiwan taking the platform to 68 markets, and the public Insights API now reporting conversion events beyond a campaign's primary goal. Ads still appear only on Free and Go plans.
Why it matters: Conversion reporting changes the conversation. Until now the honest answer for B2B SaaS was that you could buy placement inside the answer and not measure what it did, which is an easy no.
Action: The plan tier still decides it. If your buyers sit on Pro or Enterprise seats they never see the ad. If you sell product-led or self-serve, scope a four-week test at a budget you would not miss and judge it against your worst paid search campaign.
11. Google's publisher payment pilot pays most participants almost nothing
Google is paying roughly 100 publishers for content feeding AI Overviews, AI Mode and Gemini, administered inside Search Console. The Information reports that for several small to midsize sites, "the payments amount to less than one-tenth of 1% of their advertising revenue." The spread is wide: under $1,000 over several months at the low end, one participant above $1M annually. Participants say they do not know how Google calculates it.
Why it matters: The number matters less than the mechanism. Google has built a per-source contribution measurement and is surfacing it in Search Console, which is the metric this industry has been approximating from the outside with scrapers and panels.
Action: Check whether your Search Console shows the AI contribution surface, since that is the earliest signal it is widening. Start asking your visibility vendors whether they can align their citation counts to it.
Also worth knowing
- A judge sent roughly $1.72bn of ad tech damages to a jury. Google lost most of its summary judgment motions in the publisher ad tech case on 30 September. Surviving claims involve Enhanced Dynamic Allocation, Project Bernanke and Project Poirot. No trial date confirmed.
- The AI Overview tracking parameter story was wrong, and the retraction is the useful part. Links appearing to carry an attribution parameter turned out to be a Chrome extension rewriting URLs, presumably to sell the data. Far more people saw the original claim than the correction.
- Google asked the EU General Court to suspend the order opening Search data to rivals. The July decision requires Google to share rankings, queries, click-through rates and view counts with competing search and AI services from January 2027. If the suspension fails, rival engines get Google behavioural data.
- A vendor study puts 82% of commercial-intent citations on third parties and 3% on owned pages. NP Digital with Writesonic, across a claimed 9M AI answers. Treat with real scepticism: the time period is never stated, the funder is undisclosed, and the lift multiples have no described control group.
- A randomised publisher experiment found AI search widened topic diversity and raised consumption. Inside The Washington Post, 37,561 readers assigned by cookie: shared information consumption rose 69% and effective topics per reader rose from 4.58 to 5.55, while conventional result clicks fell 21 points. Data is from late 2024, so two years old.
- On the calendar. 9 October: the CMA choice screen consultation closes, the only live route to an enforceable attribution duty in a major market. Around 8 to 10 October: the spam update should post an end time. 1 November: Cloudflare AI Search billing begins. January 2027: EU Search data sharing starts unless Google's suspension succeeds.
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