Digest

Cited Is Not Visited: The Gap Just Got a Number

September 28, 2026 · Vikram Jayanand

The Big Picture

The first randomized field experiment on Google AI Mode landed this week, and its finding is blunt: when AI Mode became people's default, clicks to outside websites fell by 18.8 percentage points and satisfaction with search went down, not up. Around it, Google launched a two-week spam update backed by a newly disclosed AI forensics system, AI Overviews started carrying inline links (some of which lead to AI Mode rather than the web), and ComScore showed the assistant market splitting fast. For B2B, the through-line is that being cited and being visited are now two different outcomes, and the platforms are making it harder, not easier, to measure the gap.

Google AI Mode & AI Overviews

A controlled experiment puts a number on AI Mode's traffic cost

Researchers recruited 1,100 U.S. Chrome users in March 2026, observed three days of normal behaviour, then randomly assigned them for seven days to normal Google, a treatment that hid AI Overviews and blocked AI Mode, or a treatment that redirected searches into AI Mode. In the AI Mode group, the share of searches leading to an external website fell by 18.8 percentage points. Participants ran 0.92 fewer Google search sessions per day, and the share using Bing, DuckDuckGo or Yahoo rose by 11.2 points. Satisfaction, usefulness, sense of control and trust in Google all declined. The no-AI treatment raised external click-through by an estimated 8.8 points, though Google changed the AI Overview HTML mid-study and the blocker only hid about 51% of them, so that figure is statistically adjusted.

The caveats matter. Before the experiment AI Mode was just 0.6% of participants' searches, and the treatment pushed that to nearly 95%. The study shows what happens when AI Mode is the default, not what happens when people choose it. The sample skewed younger, highly educated and left-leaning, and it ran for a week.

AI Overviews now appear on nearly four in ten desktop searches

ComScore's Q2 2026 AI Intelligence Report puts AI Overviews on 39.4% of U.S. desktop searches in June 2026, up from 25.8% in July 2025. Copilot Search appeared on 17.3% of Bing desktop searches. Desktop search volume grew too: 77 billion U.S. desktop searches in Q2 2026, up 8% from Q2 2024. ComScore measured presence, not clicks.

GEO / AEO / AIO

The assistant market is splitting

ComScore's share of AI assistant prompts moved sharply between January and June 2026: ChatGPT fell from 70% to 50%, Gemini rose from 17% to 30%, and Claude rose from 2% to 11%. Copilot held at 5%. In AI answers, YouTube was the most cited source for retail, entertainment and financial services; in retail it led at 23%, ahead of Amazon (14%), Reddit (12%) and Wikipedia (11%).

The obvious vendor response is to track more engines. Rank Math did exactly that this week, adding Gemini to its AI Visibility feature, which runs saved prompts daily, weekly or monthly depending on plan. Some agency roundups told readers to add Gemini and Claude to their prompt tracking. That is the wrong conclusion to draw from a share shift.

Why more tracking does not fix AI visibility measurement

ViMi's position is that AI visibility tools are not effective, because LLM output is probabilistic. An AI visibility score is a draw from a distribution, not a measurement of a stable quantity. Running more prompts, more often, across more engines narrows the interval around an average, but that average still describes no individual buyer's experience, and it cannot support the before-and-after claims these dashboards are sold on. Last week's arXiv study found identical reruns flipped the cite or no-cite outcome about one time in seven. Multiplying engines multiplies distributions; it does not turn any of them into a measurement.

This week added a second, more mechanical problem. Google's /goto redirect, now routing search-result clicks through Google's own servers, makes it harder and more expensive for rank trackers, SERP APIs and AI-visibility platforms to attribute a result to the right URL. Vendors' own advice is to validate unexpected visibility shifts against Search Console before treating them as real. In other words, the dashboard's movement now has a known non-performance cause as well as a probabilistic one.

The contrary evidence, stated fairly

Two points from this week push the other way and deserve a straight answer. First, panel data can detect platform-wide product changes: the Peec link data and ComScore's AI Overview share describe what Google shipped across many queries, and a step change on a known date is useful information. That is a different claim from "your brand's score rose because of what you did," which is the claim the dashboards sell. Second, Search Engine Land's contributor guide to auditing AI visibility without a paid tool argues that ongoing automated tracking is the real value paid platforms provide. Its own checklist makes the better case, though: its strongest checks have determinate answers. Can AI crawlers reach the site at all (robots.txt, cookie walls, location pickers)? Do your facts match across your site, Google Business Profile and directories? When you ask an assistant a specific factual question about your business, is the answer true?

Key stat: Between January and June 2026, ChatGPT's share of AI assistant prompts fell from 70% to 50% while Claude's rose from 2% to 11% (ComScore).

Do not respond to the share shift by expanding prompt-tracking coverage. Respond by checking that Gemini's and Claude's retrieval can reach your key pages.
Run factual accuracy checks: ask each major assistant specific, checkable questions about your pricing model, integrations, certifications and positioning, and log errors as defects to fix at the source.
Treat YouTube as part of your search program. It is the most cited source in several verticals, and B2B explainers and demos are an under-used route into AI answers.
Measure the outcome that matters directly: ask new pipeline how they found you and whether an AI assistant was part of the research.

Sources: Google AI Overviews now appear in 39% of U.S. desktop searches (Search Engine Land), Rank Math AI unifies SEO tools, adds Gemini visibility tracking (Search Engine Land), 9 checks to audit your AI visibility without a paid tool (Search Engine Land), Google's /goto change: The latest challenge for third-party SEO data (Search Engine Land), Google's /goto Redirect: What It Means for Rank Tracking and AI Citations (SUSO)

Traditional SEO & Organic Search

The September 2026 spam update is a slow one

Google released the September 2026 spam update on 24 September. It applies globally and in all languages, and Google says it may take up to two weeks. The three earlier 2026 spam updates each finished in under three days (March in about 19.5 hours, June in just over two days, August in two days and 16 hours). Google did not say which policies it targets, and said volatility seen on 23 and 24 September before the announcement was not this update.

SAFE: Google's new forensic system for AI-generated abuse

The day after the update, SEJ reported on a short Google research paper describing SAFE (Scaled Abuse Forensics Examiner), which Google says has been deployed. SAFE uses a root orchestrator agent coordinating three specialist agents: one for content (synthetic artifacts and spirit-of-policy violations), one for inorganic behaviour such as burst publishing and synchronized uploads, and one that maps networks through shared infrastructure. The paper leans toward video and withholds its test results, and it does not tie SAFE to Search rankings. SEJ's reading is that it may be part of the spam update. The practical point for B2B: detection is moving from "is this text AI-written" to "does this look like a coordinated production operation," which is exactly the pattern of programmatic content hubs spun up at scale.

Bots in your analytics

News UK told Press Gazette that about 20% of its website traffic over six months was non-human. Google Analytics added hostname allowlists to help filter spam traffic. If your engagement rates look worse than they should, clean the data before you rewrite the content.

Key stat: Earlier 2026 spam updates finished in under three days; this one may take up to two weeks.

Annotate 24 September in analytics and hold off on reactive content changes until roughly 8 October.
If you run programmatic or templated content at scale, audit it now against Google's scaled content abuse policy.
Set up GA hostname allowlists and bot filtering before your Q4 reporting cycle.

Sources: Google Rolls Out September 2026 Spam Update Globally (SEJ), Google releases September 2026 spam update (Search Engine Land), Google Has Deployed A New AI Spam Detector Called SAFE (SEJ), Google Analytics adds hostname allowlists (Search Engine Land), Fifth of News UK website traffic was non-human (Press Gazette)

New Protocols & Standards

Universal Commerce Protocol checkout switched on by default

Google began emailing Merchant Center users that eligible Shopify stores are now enabled for native checkout inside AI Mode and Gemini, running on the Universal Commerce Protocol. "Eligible products are automatically included." It is opt-out, via Shopify's Agentic sales channel. Most B2B sellers are not on Shopify checkout, but this is the template for agentic transactions: the order happens on Google's surface, and your analytics never see the session.

Markdown mirrors strip the wrong layer

A widely shared SEJ column argued that text-only or markdown versions of pages served to AI systems (the idea behind llms.txt-style mirrors) drop the wrong thing. Removing the visual layer is fine, because machines do not need it. Removing the structural layer means an agent can read about your demo request form but cannot use it. For B2B sites preparing for agents, the priority is clean semantic HTML with working forms and buttons, not a parallel prose copy.

Search Console adds a multimodal search type

Search Console's Performance report now has a multimodal filter covering Google Lens, Circle to Search, image uploads and Chrome's "Search this image." It shows page-level clicks and impressions but no queries. Google also added CrUX ad metrics (ad count, density, CPU and network weight) to its page experience documentation, with John Mueller noting "not everything is a direct search ranking factor."

UK regulator proposes AI assistants on search choice screens

The CMA proposed yearly search choice screens on Android and Chrome in the UK, with AI assistants such as ChatGPT and Perplexity eligible to appear if they meet technical and security requirements, and a requirement that search providers fairly attribute publisher content. Consultation closes 9 October; a final decision is expected by year end.

R4T-Diffusion: a cheaper query fan-out

Google Research described R4T-Diffusion, a query fan-out framework using a compact 53.9-million-parameter model trained to mimic a larger one, claiming 12 to 20 times faster fan-out. Google has not confirmed deployment. If fan-out gets cheap, expect AI answers to pull from more sub-queries per prompt, which favours pages that cover the comparison, integration and objection questions around a topic.

Key stat: R4T-Diffusion claims 12 to 20 times faster query fan-out than autoregressive approaches, using a 53.9M-parameter model.

Test your key conversion paths (demo request, pricing, contact) with JavaScript disabled and in a text browser. If an agent cannot submit the form, fix that before building any markdown mirror.
If you sell into the UK, add the CMA consultation to your watch list; it could make AI assistants a default entry point for a share of Android users.
If you have a Shopify store for any product line, make a deliberate call on native checkout and give AI Mode orders a home in revenue reporting.

Sources: Google opens native AI checkout to eligible Shopify merchants (Search Engine Land), Google Merchant Center Notifications For Native Checkout (Search Engine Roundtable), The Text-Only Version Of Your Website Strips Out The Wrong Layer (SEJ), Google Search Console Adds Multimodal Filter (SEJ), CMA strengthens proposals on search choice (GOV.UK), Google Announces New Query Fan-Out Framework: R4T-Diffusion (SEJ)

Google Ads Updates

Smart Bidding learning is now framed around 50 conversions

Google updated its guidance to say bid strategies can take up to around 50 conversion events or three conversion cycles to calibrate after a change, although accounts with enough history may learn faster. Its conversion-goal guidance now explicitly covers Search, Shopping and Performance Max, and notes PMax can take longer when most traffic comes from outside Search and Shopping. For B2B lead gen, where a campaign may log a handful of qualified leads a week, this is the most important line of the week: a learning period can run for months, not days.

AI Max gets a unified report and AI Brief goes multilingual

Marking a year since launch and more than 500,000 advertisers, Google announced a single unified AI Max report showing the search term that triggered the ad, the assets shown and the landing page reached, in one view. The AI Brief closed beta, which lets advertisers give AI Max written instructions about their business, audience and messaging, expands to Dutch, French, German, Italian, Japanese, Portuguese and Spanish. Auto-upgrades of campaigns using automatically created assets and campaign-level broad match into AI Max are scheduled to finish by the end of September, while the Dynamic Search Ads sunset has been pushed to begin in February 2027.

Customer Match adds IP address and timestamp matching

Advertisers can now upload a user's IP address and interaction timestamp as Customer Match identifiers. Unlike email and phone, these must be uploaded unhashed, and the options are not available for users in the EEA, UK or Switzerland. For B2B, where much of the audience browses from corporate networks, IP matching could widen audiences built from CRM and webinar data, but it needs a proper consent and disclosure review first.

API v25.2, Microsoft, and ChatGPT ads

Google Ads API v25.2 (23 September) adds percentile-based competitive benchmarks, asset-group-level tracking templates and URL parameters for PMax, and a migration path from Smart campaigns to PMax drafts. Microsoft Ads' API added deeper search reporting and more PMax controls. OpenAI expanded ChatGPT Ads to seven more markets in Southeast Asia and Taiwan, now over 60 countries, on Free and Go plans only.

Key stat: Google now says Smart Bidding calibration can take up to around 50 conversions or three conversion cycles.

Calculate how long 50 conversions takes for each B2B campaign. If it is more than a quarter, consider optimising to a higher-volume qualified signal (for example, an MQL stage imported from CRM) rather than closed-won only.
Pull the unified AI Max report as soon as it appears and check which search terms and landing pages AI Max is choosing that you did not.
If you run AI Max, write an AI Brief that states who you do not sell to. Exclusions are where broad automation costs B2B budgets most.
Review consent language before testing IP-based Customer Match, and exclude EEA, UK and Swiss users.

Sources: Google Ads reframes Smart Bidding learning around 50 conversions (Search Engine Land), AI Brief in more languages and a new AI Max reporting feature (Google), Dynamic Search Ads are upgrading to AI Max (Google), Google Ads adds IP address and timestamp matching to Customer Match (Search Engine Land), Google Ads API v25.2 adds competitive benchmarks and PMax upgrades (Search Engine Land), ChatGPT Ads expands to Southeast Asia and Taiwan (OpenAI)

Suggested Actions This Week

Split reporting into citations and clicks. The AI Mode experiment and the AI-Overview-to-AI-Mode link test both show they are separating.
Annotate 24 September and avoid reactive changes until the spam update finishes (around 8 October). Audit any templated content at scale.
Answer the assistant share shift with retrieval checks, not more prompt tracking: confirm Gemini, Claude and ChatGPT crawlers can reach your key pages and that nothing sits behind a cookie wall or selector.
Run a factual accuracy pass on what the major assistants say about your pricing, integrations and certifications, and fix errors at the source.
Add "did an AI assistant play a part in your research?" to your inbound lead form or discovery call script. That is the buyer signal a dashboard cannot give you.
Work out how many weeks 50 conversions takes for each Smart Bidding campaign, and move low-volume campaigns to a qualified, higher-volume conversion signal.
Check your demo and contact forms work for a non-visual, script-light client before considering any markdown mirror.
Validate any sudden rank or AI-visibility movement against Search Console, given the /goto redirect.

← Back to Insights