--- title: "Why Your Marketing Analytics Miss Brand Exposure in AI Answers" description: "Analytics log clicks and paid impressions. Most AI answers are read without a click, so brand mentions in them leave little or no trace in your data." canonical: "https://underneath.agency/resources/why-analytics-miss-ai-visibility" published: 2026-10-07 updated: 2026-10-08 publisher: "Underneath (https://underneath.agency/agent)" entity: "https://underneath.agency/.well-known/entity.json" --- Guide · AI search # Why can’t our marketing analytics see our visibility in AI answers? Your analytics record what reaches your website or your ad platform, and most exposure in AI answers never reaches either. A brand can be named in a third of the AI answers in its category and still barely register in referral reports. The research now treats AI visibility as a separate measurement that has to be built, not pulled from existing dashboards. ## The short version 1. AI answers are mostly read without a click: in a panel of 900 US adults, only about 1% of visits to Google results pages with an AI Overview led to a click on a cited source. 2. In a field test with 1,100 US users, forcing Google’s AI Mode cut clicks to outside websites by 18.8 percentage points. 3. One software brand appeared in 33.8% of answers from one OpenAI model to 56 product questions, exposure no web analytics tool would have logged. 4. On one website, ChatGPT referrals grew 5.7 times in four months, but pages nobody changed grew 3.5 times, so most of the rise was the platform’s own growth. 5. A survey of 45 studies rates the evidence that AI citation scores predict clicks, conversions or revenue as “very low”. ## What do our current marketing reports actually record? They record ad impressions, clicks and visits; none of these captures an unpaid brand mention in an AI answer. [Kato and colleagues](https://arxiv.org/abs/2609.11915), in a 2026 paper on extending marketing mix models to AI, put the gap plainly: “conventional impression logs do not record nonsponsored occurrences of a firm’s name in generated answers.” Paid placements have the reverse problem. The platform records that a sponsored placement was shown, but not whether anyone noticed it. Marketing mix models, the statistical models many marketing teams use to split budget across channels, need an exposure count for each market and period. TV, search ads and social all supply one. AI answers do not. [Schulte and colleagues](https://arxiv.org/abs/2604.07585) add that AI providers offer no monitoring tool equivalent to Google Search Console, so even the questions people ask, and how often, are hidden from brands. ## Why don’t AI answers show up as traffic? Because most people read the answer and stop there, so your brand is seen but no visit is logged. The best evidence comes from Google’s AI features. In [a Pew Research Center browsing panel](https://arxiv.org/abs/2608.04831) of 900 US adults tracked in March 2025, about 18% of all Google searches showed an AI Overview, the AI summary at the top of Google’s results. Only about 1% of visits to those pages led to [a click on a cited source](https://underneath.agency/resources/do-people-check-ai-sources). People clicked any search result on 15% of pages without an AI Overview, but on only 8% of pages with one, and ended their browsing session more often (26% against 16%). The study is observational, so it shows an association, not a cause. A randomized test points the same way. In [a preregistered field experiment](https://arxiv.org/abs/2608.18352) with 1,100 US participants in March 2026, forcing Google’s AI Mode reduced clicks to outside sites by 18.8 percentage points. Hiding AI Overviews raised them by 8.8 points. We cover what that means for traffic in [our article on AI Mode](https://underneath.agency/resources/ai-mode-default-traffic-loss). Standalone assistants such as ChatGPT look similar. A [panel study by Scrunch AI](https://arxiv.org/abs/2607.04282), a vendor of AI visibility software, covered US and British users in early 2026. It found that 34.1% of sessions containing an AI assistant showed no visit to any outside website, against 19.5% for sessions built around search. ## Is AI referral traffic a good stand-in for AI visibility? No: a referral is a different event, and its growth mostly reflects the AI platforms’ own growth. Kato and colleagues make the first point directly: referral sessions measure something else, “because users may read an answer without following a link.” A brand can be named often and receive almost no visits. The second point comes from [Watanabe and Nakayashiki](https://arxiv.org/abs/2606.04362), who studied their own website’s server logs and Google Analytics data. From January to May 2026, total ChatGPT referrals grew 5.7 times. But pages they never touched grew 3.5 times over the same window. The pages they did change grew 6.1 times, and their best estimate of their own effect was a lift of about 1.82 times. Even that estimate did not pass their strictest check, so they call it suggestive. Our guide on [crediting ChatGPT referral growth to GEO](https://underneath.agency/resources/chatgpt-referral-growth-and-geo) covers this test in more depth. Their data also shows how fragile the numbers are. The share of ChatGPT visits counted as engaged rose from 0.486 to 0.896, and most of the jump came when bot filtering changed in mid-March. Monthly totals rebuilt from daily data differed from direct monthly queries by up to 3.3%. A dashboard can move for reasons that have nothing to do with your brand. ## What would it take to measure AI exposure properly? You would have to build the exposure count yourself, from repeated answers, question volumes, engine shares and attention. Kato and colleagues set out the recipe. How often a brand appears in sampled answers must be combined with how many relevant questions are asked, the share each AI system handles, and the chance a reader notices the name. Their own sample shows why a single figure misleads. Across 2,240 answers to 56 product questions, the target brand appeared in 33.8% of answers from GPT-5.6 Luna and 27.8% from GPT-4o. Yet the order flipped by language: GPT-4o led by 6.79 points in English, while GPT-5.6 Luna led by 18.75 points in Japanese. Repeated sampling matters too. In [our consistency study](https://underneath.agency/research/ai-recommendation-consistency-study), a single ChatGPT answer showed 57.8% of the brands that five answers to the same question named between them. Most firms have none of the other inputs, such as question volumes by engine or measured notice rates. The authors flag that their method needs exactly the data most firms lack. ## Can we link AI visibility to revenue yet? Not reliably: published research has not shown that AI citation scores predict clicks, conversions or revenue. [Martinez’s critical survey](https://arxiv.org/abs/2607.14035) of 45 studies from 2023 to 2026 grades that claim “very low” confidence. It rests on one suggestive quasi-experiment, the single-site study above, plus a few industry claims. One industry study reported a 20% traffic lift against a control group, but without the group sizes or uncertainty needed to rely on it. Kato and colleagues test their method on simulated sales data, not a real company’s results. That makes it a design for future measurement, not evidence of a return. ## What should you do about it? Treat AI visibility as its own channel with its own data, instead of waiting for analytics to reveal it. 1. Stop reading flat AI referral traffic as proof of absence. The studies above show exposure without clicks is the normal case. 2. Measure mentions directly. Ask a fixed set of buyer questions repeatedly, on each engine your buyers use, and report how often your brand appears. 3. When you change content, leave a comparable set of pages untouched. Watanabe and Nakayashiki’s untouched pages are what revealed that most of their growth was the platform’s. 4. Brief your analytics or mix-modeling team on the missing inputs: question volumes, engine shares and notice rates. Do not let AI exposure enter a model as a guess. 5. Ask any vendor what its visibility number counts, how many runs it uses, and whether it has ever been checked against sales. If you want help setting up that kind of measurement, see [our generative engine optimization service](https://underneath.agency/services/generative-engine-optimization). ## What does the research not tell us yet? It does not yet show how many people notice a brand in an AI answer, or what that is worth. - No study in this set measures notice rates for brand names in AI answers; Kato and colleagues need them but assume them in simulation. - Their answer sample covers one brand, two OpenAI models and one day of collection in September 2026. - The referral evidence comes from one website, one engine and a bundle of changes made together. - The click evidence is strongest for Google’s AI features; for standalone assistants it comes from one vendor’s opt-in panel. - Whether AI visibility drives sales has not been tested in a controlled way. ## Frequently asked questions ### Does Google Analytics track visibility in ChatGPT? No, it only records visits that arrive through a link. In one site’s data, ChatGPT referrals grew 5.7 times in four months while pages that were never changed grew 3.5 times, so even those visits mix your efforts with platform growth. ### Why is our AI referral traffic low if AI assistants mention us? Because most readers do not click. In a 900-person US panel, only about 1% of visits to Google pages with an AI Overview led to a click on a cited source. ### Can marketing mix modeling measure AI search? Only once you build the missing input. Researchers propose combining how often your brand appears with question volumes, engine shares and notice rates, data most firms do not yet collect. ### How do we know if our AI visibility work is paying off? Compare changed pages or questions against a similar set you left alone. Without that control, a rise in traffic may simply reflect the AI platforms growing. ## Sources - Kato, Honma and Kato (2026), [Generative Marketing Mix Modeling: A Causal Inference Framework Linking GEO and GEM to Business Impact](https://arxiv.org/abs/2609.11915), arXiv:2609.11915. - Schulte, Bleeker and Kaufmann (2026), [Don’t Measure Once: Measuring Visibility in AI Search (GEO)](https://arxiv.org/abs/2604.07585), arXiv:2604.07585. - Chapekis, Lieb, Shah and Smith, Pew Research Center (2026), [Investigating Click Behaviors On Google Search Result Pages That Produce an AI Overview](https://arxiv.org/abs/2608.04831), arXiv:2608.04831. - Wang, Gleason, Bart, Wilson and Metaxa (2026), [AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence](https://arxiv.org/abs/2608.18352), arXiv:2608.18352. - Iannelli and Ai, Scrunch AI (2026), [The New Shape of Search: How Conversational AI Recomposes Information Seeking](https://arxiv.org/abs/2607.04282), arXiv:2607.04282. - Watanabe and Nakayashiki (2026), [Disentangling Answer Engine Optimization from Platform Growth: A Log-Based Natural Experiment on ChatGPT Referral Traffic](https://arxiv.org/abs/2606.04362), arXiv:2606.04362. - Martinez (2026), [Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026)](https://arxiv.org/abs/2607.14035), arXiv:2607.14035. - Underneath (2026), [Ask an AI the same question 5 times: do the brands change?](https://underneath.agency/research/ai-recommendation-consistency-study) --- This is the Markdown twin of https://underneath.agency/resources/why-analytics-miss-ai-visibility. The HTML page is canonical. Publisher: Underneath, https://underneath.agency/agent. Site index: https://underneath.agency/llms.txt.