AI Overviews Are Now on Almost Every Branded Search - October 2026

Dominic Bryce Dominic Bryce 5 October 2026

There has been so much news that it's a double this week!

The biggest news is the massive increase in Google AI Overviews showing for branded searches (something I've seen begin on September too) - until now AIOs mostly displayed forinformational searches and longer, more complex prompt-like searches.

Why are Google showing AIOs on brand searches? I believe it is a way to keep users in the Google ecosystem. With increasing pressure from competitors like ChatGPT, Google now shows AIOs on brand searches to discourage users from visiting LLMs to do their initial research, only to return to Google just to find or navigate to a brand's website.

The other news items and stats actually show why Google would be worried about their AI competitors taking users away from them, with the PYMNTS data showing that more online shoppers are using AI during their retail research, and Shopify now accepting purchases from browser based agents also posing a potential future threat to Google.

Of course, the new Google spam update started at the end of in September 2026 to keep your eye on, and there was a repeat of analysis already out there done by Profound, finding that many AI citations churn quite regularly, showing the importance of fresh content, but that it's also possible to retain longer term citations too.

On with the news:

On Fri, 2 Oct 2026, 17:21 Dom Website AI, wrote:

>

Key Takeaways

  • AI Overview presence on branded keywords rose from around 26-27% in early September 2026 to 90.48% by 27 September 2026, per DemandSphere's tracking across all markets and devices it checks.
  • A separate scrape by Nectiv (Chris Long), run 30 September 2026, found AI Overviews present for 93 of 100 enterprise brands checked.
  • Shopify announced on 28 September 2026 that browser-based AI agents can now complete checkout, not just browse, through three new tools: get_checkout, update_checkout and complete_checkout, including Shop Pay.
  • PYMNTS Intelligence research (reported via Forkast, survey of 2,191 US consumers, August 2026) found around 61 million US consumers now start retail research with AI; average AI-assisted spend rose from $889 in June 2026 to $1,039 in August 2026, a 17% increase in two months.
  • In a controlled PYMNTS test, 40% of consumers still chose to buy on a marketplace even when told AI had found identical products at an equal price, with equal shipping and returns, from three retailers.
  • A companion PYMNTS/Visa Global Digital Shopping Index (Merchant Edition, 1,185 merchants) found only 11% of small merchants and 19% of large merchants consider themselves ready for AI shopping agents; only 15% have product data structured well enough for an agent to use.
  • Profound's analysis of 883,000 pages and 1.19 million page-and-engine citation lifecycles (8 September 2025 to 8 September 2026, across ChatGPT, Perplexity, Claude, Google AI Overviews, AI Mode, Gemini and Microsoft Copilot) found a median AI citation half-life of 11 days.
  • Pages cited by five or more AI engines are 5.5 times more likely to sustain citations long-term than pages cited by a single engine, per Profound's analysis.
  • Corroborating citation-decay studies: Trakkr (22 April 2026, 857,138 brand reports across 10,991 brands) found a median half-life of 31 days and found 73.5% of citation URLs appear once and never return; Stacker with Scrunch (26 March 2026, 3 million-plus citation events) found a half-life of around 4.5 weeks for non-network domains; LumenGEO (20 May 2026, 240 million citations) found a median half-life of around 4.5 weeks.
  • MERJ's Agent Interaction Monitoring research (published 29 September 2026, testing Claude's Computer Use, ChatGPT Agent/Atlas and Perplexity's Comet across 100 controlled tests, research period November 2025 to March 2026) found average task success rose from around 30% to around 60% over that period, but virtual scrolling (around 30%), canvas-based applications (around 34%) and icon-only buttons without labels (around 45%) remained weak as of March 2026.
  • Bloomberry's analysis of OpenAI's ChatGPT Ads tracking pixel across 2 million companies found around 15,000 adopters over six months; software and SaaS companies are 5.2 times over-represented relative to their share of the wider company pool, and US-based companies make up around 49% of adopters with a known country against a 25% baseline.
  • Similarweb's ChatGPT Ads Intelligence tool, tracking impressions since 30 March 2026, recorded 5,171 unique advertisers all-time as of 15 June 2026; Monday.com led with 5.05% of global impressions, and the UK passed 1% of global impressions after its 6 June 2026 launch.
  • Google's September 2026 spam update was released 24 September 2026 at 9:15am PDT with an announced rollout of up to two weeks; a first volatility wave hit 25-27 September and a second, separate wave hit 30 September to 1 October, per Search Engine Roundtable.
  • Google's Merchant Center policy updates, published 28 September 2026, shift Shopping ads enforcement from individual product disapprovals to account-level penalties.
  • Microsoft Advertising removed Max CPC bidding for new non-portfolio campaigns effective 1 October 2026, without stating what happens to existing campaigns already using it.

AI Overviews have taken over branded search almost overnight

The single most striking data point this week is how fast AI Overviews have spread across branded search. DemandSphere's tracking of branded keywords shows AI Overview presence rising from around 26-27% in early September 2026 to 90.48% by 27 September, across all markets and devices it checks, with once-daily checks and no country or device split published yet. Search Engine Roundtable corroborated the direction independently, reporting AI Overviews appearing for almost every large brand name it checked except Google's own and news publishers, mostly positioned mid-page. A separate one-off scrape by Nectiv's Chris Long, run on a hundred enterprise brands, found AI Overviews present for 93 of them.

It is worth being precise about what is and is not established here. Nobody has yet measured a branded click-through-rate drop; the prediction that clicks will fall is Long's, not a measured outcome. Google has made no statement about the change, and there is no verified breakdown for the UK specifically, so I would not treat a UK figure as confirmed either way. What is well supported, across two independently run checks using different methods, is that an AI Overview answering a branded query has gone from occasional to close to the default in under a month. The practical move is simple: check what AI Overviews currently say about your own brand name and about your direct competitors' brand names, because that answer box is now functioning as a results page of its own.

Agentic commerce: Shopify opens checkout to AI agents

Shopify announced on 28 September that browser-based AI agents can now complete a purchase on merchant storefronts, not only search products and add them to a cart. Three new tools, get_checkout, update_checkout and complete_checkout, let an agent inspect a checkout's current state, change details like delivery address or shipping option, and place the order, including through Shop Pay, once the buyer has authorised it. Shopify's documentation is explicit that the agent runs inside the buyer's own browser, the buyer sees the same checkout state the agent does, and the buyer must handle Shop Pay login or payment challenges and confirm before complete_checkout fires. This is rolling out to all eligible merchants. It builds on the Universal Commerce Protocol, the same standard behind Google's native UCP checkout rollout inside AI Mode and Gemini that I covered previously, and both Meta's Muse and the Instinct shopping agent already have direct Shopify integrations.

The wider pattern is a split forming across retail. Amazon and Adidas are both reported to be blocking AI agents from completing purchases on their own sites, while Shopify, Google's UCP rollout and the platforms building on it are moving the other way. Separate PYMNTS Intelligence research, reported via Forkast from a survey of 2,191 US consumers in August 2026, found around 61 million US consumers now start retail research with AI, and that average AI-assisted spend rose from $889 in June to $1,039 in August, a 17% increase in two months. Tellingly, in a controlled test where consumers were told AI had found identical products at an equal price, with equal shipping and returns, from three different retailers, 40% still chose to buy on the marketplace anyway. The reasons given were easier or more reliable returns (37%), selection (35%), habit (33%) and customer service (32%), which points to trust and post-purchase accountability as the deciding factor, not checkout friction.

That trust gap is compounded by a readiness gap on the merchant side. A companion PYMNTS/Visa Global Digital Shopping Index survey of 1,185 merchants found only 11% of small merchants and 19% of large merchants consider themselves ready for AI shopping agents, and only 15% have product data structured well enough for an agent to process it reliably. Sixty-eight percent expect agents to drive at least 5% of digital sales within two years regardless. Put together, these three stories describe the same shift from different angles: agents can now transact, consumers are starting to let them, and most merchants outside the largest platforms are not yet set up to be chosen when they do.

AI citations decay fast - track them continuously, not as a snapshot

Profound's research team analysed 883,000 pages and 1.19 million page-and-engine citation lifecycles across seven AI engines (ChatGPT, Perplexity, Claude, Google AI Overviews, AI Mode, Gemini and Microsoft Copilot), covering 8 September 2025 to 8 September 2026. Their definition of a citation's "half-life" is the number of days after a page's peak citation share for it to fall to half that peak and stay there, measured on a trailing 14-day average. The headline figure is a median half-life of 11 days: 78% of cited pages reach half their peak within 14 days or less. The 22% of pages that do sustain citations manage a 27-29 day median half-life depending on the engine, Claude citations are 2.8 times more likely than ChatGPT's to survive eight weeks, and pages cited by five or more engines are 5.5 times more likely to persist than pages cited by only one.

This finding is well corroborated by independent work using different methods and datasets. Trakkr's research, published 22 April 2026, covered 857,138 brand reports across 10,991 brands plus 108,650 URLs for 200 brands across 8 AI models from June 2025 to March 2026, and found 73.5% of citation URLs appear once and never return, a median half-life from peak of 31 days, and a week-over-week citation swing averaging 51.8%. Stacker, working with Scrunch and published 26 March 2026, analysed over 3 million citation events across more than 120,000 domains in 8 industries and 6 platforms over 26 weeks, finding a half-life of around 4.5 weeks for domains outside its publisher network, against nearly 10 weeks for domains inside it, though the authors are careful to say they cannot claim causation for that network effect. LumenGEO, published 20 May 2026 from 240 million citations, found a median cited-source half-life of around 4.5 weeks, with 40-60% of cited domains rotating month to month and 70-90% rotating over six months.

The exact figures differ because the studies measure slightly different things, pages versus domains, decay from peak versus overall lifespan, but the conclusion across all of them is consistent: AI citations turn over quickly, and distributing a topic across several credible, independent sources meaningfully extends how long any one of them stays cited. The practical implication for anyone tracking AI visibility is to treat citation monitoring as a continuous job rather than a one-off audit, and to keep key pages genuinely current with specific, regularly refreshed detail rather than publishing once and leaving them.

What AI agents still can't do on a typical website

Two pieces of research published this week make closely related points about the gap between what AI agents can read and what they can actually do. Slobodan Manic, writing on Search Engine Journal, argues that the text-only or markdown mirrors many sites are now building specifically for AI agents preserve readable prose but strip out exactly the layer an agent needs to complete a task: the forms, buttons and working links that let it act rather than just read. His framework splits a page's content into what is "describable" (readable prose, fine for a markdown mirror) and what is "doable" (an interactive element that needs to keep working, with or without the visual layer on top), and he argues that most sites need to fix basic semantic HTML, what he calls "the floor", before a declared tool surface like WebMCP, "the ceiling", will help at all. He cites WebAIM's 2026 accessibility evaluation, which found 95.9% of the top million homepages fail basic WCAG 2 checks, with unlabelled form inputs (51%), empty links (46.3%) and empty buttons (30.6%) as the most common failures, and notes that without a programmatic success confirmation, an agent completing an action has no way to tell whether it actually worked, which risks duplicate orders or submissions.

MERJ's own primary research, published 29 September 2026, puts numbers against that same gap. Testing Claude's Computer Use, ChatGPT Agent (Atlas), Perplexity's Comet and several frontier models across 100 purpose-built controlled tests covering six failure stages, plus live-site testing in 10 languages, over a research period from November 2025 to March 2026, MERJ found average task success rose from around 30% to around 60% across that period as the agents themselves improved. But several specific interaction types remained weak as of March 2026: virtual scrolling around 30% success (up from around 22%), canvas-based applications around 34% (up from around 24%), icon-only buttons without accessible labels around 45% (up from around 34%), and typing into fields that were disabled around 45% (up from around 34%). In one case, an agent clicked a "Cancel" button in 25 out of 25 runs because a misleading ARIA label told it that was the primary action, despite a clear visual hierarchy suggesting otherwise. Across ten languages, success rates dropped by 8.8 to 18.4 percentage points compared with English depending on the model.

MERJ's own caveat is worth repeating: their results "aged quickly" and are version-specific evidence about particular agent releases rather than permanent properties of AI agents in general, and the test data, while published in September, covers a period that ended in March. Taken together with Manic's argument, the practical starting point both pieces converge on is the same: real button and anchor elements instead of divs standing in for them, accessible names that actually describe what a control does, an ARIA audit for truthfulness rather than just presence, and persistent, announced error states rather than a toast that disappears after four seconds. That is foundational accessibility work, and it happens to be exactly what makes a page usable by an AI agent too.

Measuring what AI-driven traffic is actually doing

Kevin Indig's Growth Memo newsletter, featuring a conversation with Ramp's VP Growth George Bonaci, tackles a problem that is becoming harder to ignore: click and cookie-based attribution was never a particularly good way to measure brand and content channels, and AI-assisted buying journeys make that weakness more visible rather than creating a new one. The typical journey Indig describes is prompt, synthesise, then a direct or branded-search return visit, which means standard analytics credits only the return visit and the AI exposure that actually drove the decision disappears from the data entirely. Bonaci's framing is blunt: attribution "has become a crutch replacing critical thinking. Every attribution model is wrong." Ramp still runs multi-touch attribution and org-level goals, he says, but debates its limits daily, and attribution taken alone "would have led us to cancel all the brand marketing we do, all our stunts, all our direct mail, all our events."

The data behind the argument: Graphite's research suggests AI-driven traffic can be under-attributed by as much as 10 times; a Boston Consulting Group survey of 3,000 measurement professionals found 46% now combine marketing mix modelling, incrementality testing and multi-touch attribution together, with that group seeing up to 70% stronger revenue growth than those relying on a single method; and an IAB survey found only 39% of US buy-side decision-makers use all three methods. Indig's own research found a medium-strong correlation between AI share of voice and conversions. His two suggested alternatives are triangulation, combining an exposure metric like share of voice with a behavioural signal such as self-reported discovery or CRM tagging and an actual business outcome, so that divergence between the three tells you where to investigate, and incrementality testing through holdouts, geo experiments, phased rollouts and marketing mix models calibrated against real experiments.

Separately, and without any announcement from Google, Search Engine Journal's Roger Montti reports that Gemini has quietly started adding UTM parameters to some outgoing links, noticed around 24 hours before publication. The behaviour is undocumented: it is not clear under what circumstances the parameters are added, there is no Google statement or documentation to confirm it, and the article gives no breakdown by surface (web, app, Chrome, API). If a UTM survives the in-app webview, as the article argues, it would show up in raw server logs and be less likely to fall into an undifferentiated Direct bucket, which the article illustrates with a reported 9% referrer pass-through rate from the Gemini mobile app against 0% from Android Assistant, though the source of those specific figures is not given. By contrast, the article says ChatGPT only adds UTMs to links that are grounded in web sources. My advice here is simply to check your own server logs and GA4 source and medium fields for Gemini traffic before reporting anything based on this, since the implementation appears inconsistent and none of this has been confirmed by Google.

Seven assumptions about AI search worth re-checking

A useful myth-busting piece from Search Engine Land, published 28 September and written by Rob Tindula, tests seven commonly repeated claims against available data. A few are worth calling out directly. The claim that purely AI-generated content cannot rank is not supported: the article points to AI-generated articles from March to September 2025 that still rank and perform well, with the quality of the underlying process mattering more than who, or what, wrote it. The claim that llms.txt improves AI visibility also does not hold up: Google has said directly that it does not use llms.txt for AI search discovery, and a 90-day study cited in the piece found no measurable impact from implementing it. And the claim that AI is not driving meaningful traffic undersells what is actually happening: users who arrive at a site after an AI mention visit at 1.5 to 2.5 times their forecasted baseline rate, even though only around 2.5% of those downstream visits carry a trackable AI-referral parameter, which suggests AI-driven traffic is being undercounted in most analytics setups rather than genuinely absent.

Paid media: who is actually buying ChatGPT ads, and two platform changes to action

Henley Wing Chiu at Bloomberry ran an unusual piece of research: scanning the websites of 2 million companies, ranked by LinkedIn employee count, for OpenAI's bzrcdn.openai.com tracking pixel, which signals that a company has set up ChatGPT Ads attribution. That turned up around 15,000 adopters over six months. The adopter pool is heavily skewed: software and SaaS companies make up 17.0% of adopters against 3.3% of the overall company universe, a 5.2 times over-representation; information and internet companies are 3.1 times over-represented and advertising services 2.4 times; companies with 51-200 employees are 2.4 times over-represented, and Series B and Series C-funded companies are 2.5 and 3.1 times over-represented respectively. Geographically, around 49% of adopters with a known country are US-based against a 25% baseline, with Japan at 3.7 times and Brazil at 2.5 times. Reddit advertisers are 7.3 times more likely to also be ChatGPT Ads adopters than Facebook advertisers are. Chiu is upfront that pixel detection only shows tracking set-up, not actual ad spend or impressions served.

That picture is corroborated by Similarweb's separate tracking of actual ad impressions, which has monitored ChatGPT Ads Intelligence across more than 30 countries since 30 March 2026 and recorded 5,171 unique advertisers all-time as of its 15 June 2026 report. Eight of its top ten advertisers by impression share are B2B SaaS or productivity brands, led by Monday.com at 5.05% of global impressions, with Vanta, Aikido Security, Wix, Home Depot, Shopify and Cursor also featuring. The US holds 71.6% of impressions, and the UK only passed 1% of global impressions after its launch there on 6 June, though that grew 18 times week on week immediately afterwards. For a UK-based consumer brand, the clear read is that ChatGPT Ads are still early and heavily tilted towards B2B software, so there is time to watch how competitors use the format before committing meaningful budget.

On the operational side, Google has rewritten several Merchant Center policies, covering dangerous products, counterfeit goods, misrepresentation, sexual content and editorial requirements, to shift enforcement focus from disapproving individual products towards account-level penalties, and has standardised the appeals process through the newer Merchant Center Next interface. The clearest practical warning in the update is that submitting a bulk appeal for a large set of offers risks immediate failure if any violation is found within that set, so reviewing flagged products individually before appealing is now the safer route. Separately, Microsoft Advertising removed Max CPC as a bidding option for new non-portfolio campaigns from 1 October 2026, taking away what many advertisers had relied on as a hard cost ceiling that automated bid strategies cannot exceed. Microsoft's announcement does not say what will happen to existing campaigns already using Max CPC, which is a notable gap compared with its 2025 bid consolidation that included an explicit migration path, so it is worth reviewing any live Microsoft campaigns on Max CPC now rather than waiting to find out what changes.

A practical framework for AI search optimisation

Independent SEO consultant Aleyda Solis has updated her AI search optimisation framework and checklist, last revised 26 September 2026, into a 12-step operating process built around commercially relevant AI search journeys. It runs across three layers: Presence, defining target prompts and measuring baseline visibility across AI platforms; Readiness, diagnosing and fixing gaps in technical infrastructure, content extractability, entity consistency, third-party citation, commercial data and localisation; and Impact, reporting outcomes and validating improvement through recurring testing. A few distinctive points stand out: it separates pages that earn citations from pages that actually drive clicks, measures visibility separately per AI platform rather than as one blended score, prioritises third-party corroboration of brand claims over self-published claims, and localises by market rather than simply by language. It also carries an explicit warning that prompt injection, fabricated reviews, manufactured community signals and citation gaming may produce short-term visibility gains but will eventually be addressed by the platforms, trading durable trust for a temporary bump. The framework comes with a downloadable worksheet and a 12-row action table linking out to nine supporting guides, and states its underlying Google Search, Bing and OpenAI documentation references were verified as of 26 September 2026.

The September spam update is still rolling - a second wave hit on 30 September

Google's September 2026 spam update was released on 24 September at 9:15am PDT, explicitly applying globally and to all languages, with Google stating it "may take up to two weeks" to roll out, an unusually long window compared with the typical two days for most updates. A first wave of volatility was felt from 25 to 27 September, and a second, separate wave hit on 30 September into 1 October, according to Search Engine Roundtable's Barry Schwartz and reports from site owners tracked by Glenn Gabe, with Discover and News traffic volatility reported again during the second wave. Gabe has also flagged a resurgence of Reddit AI-translated pages in several countries during this period, cause not yet established, and says some sites hit by earlier spam updates appear to be recovering, though that is his working hypothesis rather than a confirmed pattern. Google's own October 2026 webmaster report, published 1 October, confirms the update "started off strong and is still rolling out."

A few related items from the same report are worth noting. Gemini 3.8 Flash's earlier citation and link bugs inside AI Mode have now been fixed, alongside various interface adjustments to AI Overviews and AI Mode, and Google has begun piloting an AI Contribution programme to compensate publishers whose content appears in AI features, though payouts are reportedly minimal at around 0.1% of ad revenue across roughly 100 participating publishers so far. Separately, a US judge ruled that Google's ad tech division will not be forced to divest, instead ordering behavioural changes with ongoing oversight, and Google is appealing a separate EU requirement to share search history data with competitors. Google's own guidance on the spam update remains to review its spam policies and expect that improvements may take months to show. My advice is to wait for the rollout to fully complete, likely close to the full two-week window Google announced, before drawing firm conclusions from any ranking movement, and to compare Search Console data from 24 September onward with that second wave date in mind.