AI Overviews Jumped From 26% to 80% of Branded Searches

The Big Picture

Google AI Overviews moved onto branded searches at scale in the last days of September: DemandSphere saw presence on tracked brand queries jump from about 26% to over 80%, and a separate test found AI summaries for 93 of 100 enterprise brands, including Salesforce and Mailchimp. For B2B, that means the first thing many buyers now read about you on Google is a generated paragraph, not your homepage. In the same week Google told sites to fact-check AI content by hand (metadata and structured data included), its two-week spam update hit a second wave, and the measurement debate sharpened: Search Engine Land published the case for sampling AI answers 60 to 100 times, while the week's own evidence showed why a bigger sample still does not produce a number you can manage against.

Google AI Mode & AI Overviews

AI Overviews now sit on top of most branded searches

DemandSphere tracked branded keywords daily across markets and devices. Through most of September, AI Overviews appeared on roughly 27% to 35% of them. Presence then jumped to 69.21% on 26 September, peaked at 90.48% on 27 September, and held above 80% on 28 and 29 September. Google has not announced a change. Chris Long of Nectiv separately tested 100 enterprise brands via SerpApi and found AI Overviews for 93 of them, including Reddit, Salesforce, Mailchimp and Facebook, plus smaller companies.

Neither dataset measured clicks, so the effect on branded click-through is unknown. DemandSphere counted an AI Overview whenever one appeared, whether or not it cited the brand.

Citation layout keeps shifting

Google is again testing citation cards at the bottom of AI Overviews rather than on the right, and testing a "Loading..." state in place of the "Show more" button. Earlier in the month, Gemini 3.8 Flash launched in AI Mode without links or citations before Google fixed it. Layout tests like these change where, and whether, a cited source is seen without any change to whether it is cited.

AI Mode info monitoring rolls out to everyone

Robby Stein announced on 28 September that AI Mode's info monitoring is rolling out globally in the Google app. Users tell AI Mode what to watch and Search keeps checking sites, forums, social posts and real-time data. Every example Google gave was local or shopping. Google has not explained how it picks the sources behind an update, or whether updates carry links.

Key stat: AI Overviews appeared on 90.48% of DemandSphere's tracked branded queries on 27 September, up from about 26% at the start of the month.

• Search your own brand name, your brand plus "pricing", "reviews", "alternatives" and "vs [competitor]" today, and read what the AI Overview says. Is it accurate? That is a yes-or-no question with a fixable answer.
• If the summary is wrong, trace the claim to its source page (yours or a third party's) and fix or update that page. The summary is only as good as what it retrieves.
• Watch branded click-through in Search Console from 26 September onward, and annotate the date.

Sources: Google AI Overviews jump from 26% to 80% of branded queries (Search Engine Land), Google AI Overviews Return For Many Large Brand Names (Search Engine Roundtable), AI Overviews testing citation cards at bottom (Search Engine Roundtable), Google Rolls Out AI Mode Info Monitoring To All Users (SEJ), October 2026 Google Webmaster Report (Search Engine Roundtable)

GEO / AEO / AIO

The case for "sample more" got its fullest statement this week

Search Engine Land ran a piece by Greg Jarboe arguing that AI visibility is unstable and the fix is longitudinal measurement. It pulls together the research well: SE Ranking ran 10,000 keywords through AI Mode three times in one day and found just 9.2% average URL overlap; SparkToro and Gumshoe found under a one-in-100 chance of getting the same brand list twice; Ahrefs found AI Overview content changed about 70% of the time between checks; a University of St. Gallen study found about 65% of cited sources changed day to day. The recommendation: build a prompt portfolio, run each prompt 60 to 100 times (SparkToro) or at least seven times a day (St. Gallen), and report visibility as a rate with a range. (Search Engine Land is owned by Semrush, which sells an AI visibility product; the article carries Semrush AI visibility calls to action.)

Why a bigger sample does not fix it

ViMi's position is that AI visibility tools are not effective, because LLM output is probabilistic. The research in that article is the strongest evidence for the position, and the recommendation is the vendors' own defence of it.

An AI visibility score is a draw from a distribution, not a measurement of a stable quantity. Running 100 samples instead of three narrows the interval around the average. It does not make the average describe anything a buyer experiences. Each buyer gets one draw, shaped by their own wording, history, location and the model version that day. A 42% visibility rate with a tight confidence band is a precise description of a population of answers nobody reads as a population. And it cannot carry the before-and-after claims the dashboards are sold on, because the distribution itself moves underneath you: the same article cites Promptwatch data (labelled provisional) showing Reddit's share of ChatGPT Search citations falling 86% in a few days in August after a platform-side change. When the thing you are measuring is redrawn by the platform, a change in your rate does not tell you whether your work did anything. The article concedes this: "Correlation in this data isn't proof of cause."

What the contrary evidence says, honestly

Two findings this week cut the other way and deserve a fair hearing.

Ahrefs found the meaning of AI Overviews stayed stable (0.95 semantic similarity) even as wording and sources rotated. That is real stability. But what it supports is checking what the engines say about you (a determinate question: accurate or not), not counting how often you appear.

Writesonic's analysis of 9 million AI answers across nine platforms for 400+ enterprise brands found platform volatility in a tight 0.08 to 0.32 range while brand volatility spanned 0.04 to 1.00, and argues the brand is the variable, not the platform. If that holds, some brands' instability is diagnosable. The caveat: the dataset is Writesonic's own tracked clients and competitors, and a brand's volatility can reflect how contested its category is as much as anything it publishes. The fixes it points to (thin topical coverage, ambiguous entity information) are things you check directly on your pages, not things a score tells you.

The same Writesonic study adds a useful warning about citation counts: up to 75% of AI answers that cite one of a brand's pages never name the brand. A rising citation number can mean your research is being used while buyers never see your name. Its other findings are worth noting for B2B: 82% of citations on bottom-of-funnel commercial prompts went to third parties and only 3% to owned pages; 64% of citations came from ordinary pages, not listicles or AI-friendly formats; YouTube showed the strongest third-party association at 2.8x, ahead of LinkedIn (2.7x), Reddit (2.4x), G2 (1.9x) and Gartner (1.6x). These are correlations from a vendor dataset, not causal effects.

Key stat: Up to 75% of AI answers that cite a brand's page never mention the brand by name (Writesonic, 9 million answers).

• Replace "visibility rate" in reporting with three checks that have answers: can the engines retrieve your key pages (crawl logs, robots.txt, rendering)? Are the claims they make about you accurate? Did any real buyer conversation change (sales call notes, "how did you hear about us" fields, AI referral sessions that converted)?
• For commercial-intent topics, the evidence points to third-party coverage (reviews, analyst pages, comparison sites, YouTube). Audit what those sources say about you before writing more owned content.

Sources: Your AI search visibility keeps changing (Search Engine Land), What 9 million AI answers reveal about search visibility (Search Engine Land), Ahrefs: AI Overview change study, University of St. Gallen study (arXiv)

Traditional SEO & Organic Search

September spam update: second wave on 30 September

Google announced the September 2026 spam update on 24 September and said it would roll out over about two weeks, not the usual two days. The first wave of impact landed 25 to 27 September; a second hit on 30 September and 1 October, with site owners reporting sharp drops on those dates and Discover traffic swinging heavily. Glenn Gabe noted Reddit's AI-translated pages surging again in several non-English markets after being hit by the May core update and June spam update. Expect more waves; if it runs the full two weeks, it should finish around 8 October.

Google defines "main content" and how raters judge it

Google updated its helpful content documentation (dated 1 October) to define main content as any part of a page that directly helps it achieve its purpose, including tools, reviews, tabbed sections, titles and headings. It also states that quality raters assess main content on four attributes: effort, originality, talent or skill, and accuracy. Relevant Audience also reported a new warning about fake authors on the same page. For B2B, product and solution pages are main content too: thin feature lists and stock copy are what this language describes.

Desktop CTR fell in Q2, but check the measurement break

Advanced Web Ranking's Q2 2026 data shows desktop click-through at positions one and two down a combined 5.27 points from Q1, with mobile up a combined 6.59 points. SEJ flagged that the comparison spans the 27 April fix of a Search Console logging error that misreported impressions from May 2025 to April 2026. Because CTR is clicks divided by impressions, part of this shift may be the measurement, not user behaviour.

Penske and Chegg suits dismissed

Judge Amit Mehta dismissed the Penske Media and Chegg antitrust complaints over AI Overviews on 30 September, finding they had not plausibly alleged an agreement trading traffic for content. Nothing changes in how Google uses publisher content. Separately, The Information reported Google's AI contribution pilot pays about 100 publishers roughly 0.1% of their ad revenue, with participants unable to explain how payouts are calculated.

Key stat: Desktop CTR at positions one and two fell a combined 5.27 points in Q2 (AWR), though the comparison crosses a Search Console data fix.

• If traffic dropped on 25 to 27 September or 30 September, check for thin or scaled pages before assuming a technical issue. Do not make major changes until the rollout completes.
• Review your highest-value solution and product pages against the four rater attributes. Effort and accuracy are the easiest to fail on templated B2B pages.
• Annotate 27 April 2026 in any Search Console CTR trend you present internally.

Sources: Phase Two Of The Google September 2026 Spam Update (Search Engine Roundtable), Google Helpful Content Doc With New Main Content Sections (Search Engine Roundtable), Google says main content is critical (Search Engine Land), SEO Pulse: AI content fact-check, Gemini UTM tags, Q2 CTR (SEJ), Google Paying ~100 Publishers 0.1% Of Ad Revenue (Search Engine Roundtable)

New Protocols & Standards

Manual fact-checking now covers structured data and metadata

On 1 October Google added three sentences to its guidance on generative AI content. It now says models "don't retrieve facts" but predict likely word sequences, that it is "critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing," and that this applies to titles, meta descriptions, structured data and image alt text. Google announced no new penalty or ranking signal, and the scaled content abuse policy is unchanged. Still, the timing (one week into a spam update) and the word "critical" are notable. Many B2B sites auto-generate schema and meta descriptions at scale; those fields are now explicitly in scope.

Gemini links carry UTM tags (undocumented)

A Reddit user found UTM parameters on links from Gemini to websites, and John Mueller confirmed he sees them too. Google has not documented when Gemini adds them. If they persist, they give analytics a way to identify Gemini visits without relying on referrer data. This is a measurement with a determinate answer: a visit either arrived from Gemini or it did not.

A cautionary false alarm

A widely shared screenshot showed a ?st_source=ai_overview parameter on AI Overview and AI Mode links, prompting hopes of proper attribution. It turned out to be a Chrome extension rewriting URLs, presumably to sell the data. Verify any new signal before building reporting on it.

llms.txt: still no evidence for search visibility

No new standard emerged this week. For context, the most thorough recent evidence review (Agent Experience, 1 September) found no public evidence that llms.txt improves AI retrieval or citation. Google states Search ignores it; Ahrefs found 97% of llms.txt files got zero requests in May 2026; SE Ranking's 300,000-domain study found no relationship with citations. The one positive signal: coding agents (Claude Code) fetched it more than AI retrieval bots. Useful for developer documentation, not a visibility tactic.

Key stat: 97% of llms.txt files received zero requests in May 2026 (Ahrefs server logs, 137,210 domains).

• Add a human review step for AI-generated title tags, meta descriptions, alt text and schema.
• Create a GA4 segment for utm_source values from Gemini and compare against referrer-based Gemini sessions.

Sources: Google Tells Sites To Fact-Check AI Content Before Publishing (SEJ), Google Gemini Adds UTM Parameters For Referral Attribution (SEJ), Google Tests URL Tracking Parameters, update: browser extension (Search Engine Roundtable), Does llms.txt actually work? (Agent Experience)

Google Ads Updates

A/B testing arrives for Performance Max asset groups

Google Ads is rolling out controlled experiments inside PMax asset groups: test new headlines, descriptions, images and video against existing creative instead of swapping assets and comparing before and after. The catch is that you cannot edit the asset group while a test runs. For B2B advertisers with low conversion volume, plan longer test windows; a creative test needs enough conversions on each side to mean anything.

Business names can differ from destination domains (from 19 October)

From 19 October, eligible advertisers can use a recognised business name that does not exactly match the domain users land on, provided the name reflects the brand, there is a verified direct relationship with the domain owner, and the products are offered on that domain. Resellers, affiliates and intermediaries are excluded. Useful for B2B groups with product brands on parent-company domains, or acquired brands not yet migrated.

Automated promotions on by default (from 12 October)

Google will start pulling deals and coupons from advertisers' websites into promotion assets for Search and PMax campaigns from 12 October, on by default. It applies to campaigns with linked location assets and no promotion assets attached, so most B2B accounts will be unaffected, but check if you have location assets linked.

Elsewhere in paid search

Loyalty Customer Match expanded to AI Mode and more regions. Microsoft Advertising made optimization experiments generally available. OpenAI added bulk product creation and product review status to ChatGPT Ads. Google is testing a sponsored products grid that looks like the organic product grid.

Key stat: Automated promotions switch on by default on 12 October for eligible Search and PMax campaigns.

• Pick one PMax asset group with the most conversion volume for your first creative test, and freeze other changes for its duration.
• If you run location assets, decide before 12 October whether to opt out of automated promotions. A scraped "20% off" banner on a B2B ad can contradict your pricing conversation.

Sources: Google Ads brings A/B testing to Performance Max asset groups (Search Engine Land), Google Ads Business Name Policy Change (Search Engine Roundtable), Google Ads will automatically pull promotions (Search Engine Land), Google Loyalty Customer Match Expands To AI Mode (Search Engine Roundtable)

Suggested Actions This Week

• Read your branded AI Overviews. Search your brand, brand plus pricing, reviews, alternatives and "vs" queries. Log every factual error and trace it to the source page. Fix what you own; contact the owners of what you do not.
• Check retrieval, not rates. Confirm your top 20 commercial pages are crawlable by Googlebot and the main AI crawlers (server logs, robots.txt, rendered HTML). This has a yes-or-no answer.
• Push back on visibility-rate reporting. If an agency or tool reports an AI share of voice with a range, ask what decision it changed and how they separate your work from platform changes. Ask for buyer evidence instead: AI referral sessions that converted, and what prospects say on calls.
• Hold changes during the spam update. If you saw drops on 25 to 27 or 30 September, audit thin and scaled pages but wait for the rollout to finish before acting.
• Add human review to AI-generated metadata and schema. Google now names these fields explicitly.
• Set up Gemini UTM tracking in GA4 and compare it with referrer data.
• Decide on automated promotions before 12 October, and queue a first PMax asset-group experiment.