AI Clicks & Citation Gaps

A randomised field experiment involving 1,065 US desktop Chrome users found that Google's AI Overviews reduce organic clicks by 38% on queries where they appear. Zero-click searches rose from 54% to 72% when AI Overviews were present. Researchers recruited participants via Prolific and deployed a Chrome extension that randomly assigned them to one of three groups: standard Google Search, AI Overviews hidden, or all searches redirected to AI Mode. The experiment ran for two weeks per participant between January and February 2026 and was pre-registered with the AEA RCT Registry, making it the first randomised field experiment on AI Overview impact run under real browsing conditions.

The finding that closes off the user-benefit argument: self-reported satisfaction, perceived quality, and ease of finding information were nearly identical whether AI Overviews were present or removed. The study authors concluded that AI Overviews "divert traffic away from publishers without delivering measurable improvements in user experience." The effect was strongest when AI Overviews appeared at the top of the page, which happened in 85% of cases. For publishers and SEOs, this is the clearest evidence yet that the click reduction from AI Overviews is structural, not a temporary side effect of a new feature finding its footing.

Semrush analysed web visit data across more than 50,000 websites and 17 industries throughout 2025. The AI traffic headline is deceptively impressive: AI-driven visits grew 66% over the year, from 462 million to 767 million monthly visits. But the context matters. AI still accounts for just 0.14% of total web traffic. Google AI Mode grew from 1,600 to 38.2 million monthly visits over the period but represents only 0.01% of all traffic. The absolute numbers are growing fast; the relative share is still tiny relative to organic search, direct, and social.

The more significant finding is on the organic side. Organic search declined in 13 of the 17 industries analysed, with growth limited to visually driven sectors, beauty, apparel, food, and retail. Across all industries, organic search maintained a 16% share of total web traffic but that share is under sustained pressure. The headline AI growth rate and the organic decline are happening at the same time, and the AI gains are not yet large enough to compensate for the organic losses in most verticals. Retail saw the largest AI traffic increase at 343%, but even there the absolute volumes remain a fraction of organic.

Foundation Marketing and AirOps tracked 5.1 million AI responses and 57.2 million individual citations across ChatGPT, Gemini, Perplexity, Google AI Mode, and Google AI Overviews over 60 days from December 2025 to February 2026. The sample covered 50 brands across seven B2B verticals using 100 prompts per brand, split 65% unbranded discovery and 35% branded validation. The headline finding: brands own just 10% of citations in AI responses about them. The other 90% point to third-party sources: Reddit, YouTube, review sites, comparison platforms, and industry forums.

The gap becomes more pronounced at the discovery stage. When looking only at category-level prompts, the kind buyers use when building a shortlist before they know which brand they want, brand-owned domains account for just 2.2% of citations. Third-party sources took 97.8%. Reddit alone made up 20.8% of all external citations overall and 30.9% in unbranded discovery queries. The vertical breakdown adds texture: fintech brands were most cited via comparison sites, DevOps brands via developer platforms, and healthcare brands via medical authorities. Each sector has a different citation ecosystem, and a GEO strategy that ignores the specific platforms AI relies on for a given vertical is working with the wrong map.

Ben Tannenbaum, CEO of Aiso, tested 90 prompts across the beauty, legaltech, and IT sectors to understand what type of content ChatGPT retrieves when answering commercial questions. The headline finding: commercial prompts triggered ChatGPT to generate additional downstream queries, known as fan-out, in 78.3% of cases. Informational prompts did so just 3.1% of the time. Of the 20 fan-out triggers recorded across the study, 18 were commercial in nature and 2 were informational. The fan-out behaviour matters because it determines what additional content ChatGPT searches for to complete an answer, and that content becomes part of the citation set.

The practical implication is that content designed to rank on informational keywords is poorly positioned for ChatGPT citations on commercial questions. What performs better, based on the study's directional findings, is content that answers the questions a buyer asks at the point of comparison: best-of lists, vendor comparison guides, feature-driven category explainers, evaluation FAQs, and alternative product pages. These formats sit at the intersection of information and decision, which is where ChatGPT fan-out most often lands. The study acknowledges its sample is small and directional rather than definitive, but the commercial-prompts-trigger-commercial-fan-out pattern is consistent across all three sectors tested.

SE Ranking's Yevheniia Khromova sets out a five-step framework for deciding which prompts to track when monitoring AI search visibility. The core challenge the article addresses is that AI answer sets are unstable: only 35% of domains repeat consistently in AI responses, and two-thirds of domains that appear in one run are gone from the next. That makes prompt selection the most important decision in an AI tracking setup, because tracking the wrong prompts produces noise rather than signal. The five steps are: identify your prompt categories (informational, comparative, instructional, brand-specific, transactional), map them to buyer journey stages, source candidate prompts using seven methods, filter by competitive relevance and business intent, then set tracking volume and frequency.

The sourcing methods are the most practical section of the guide. Khromova recommends converting existing SEO keywords into conversational form, mining Google People Also Ask, asking LLMs directly what users in your category search for, analysing Reddit and forum threads for natural language questions, reviewing paid search data, reverse-engineering competitor websites, and using AI tracking tool suggestions as a starting point. The recommended starting range is 10 to 20 awareness prompts, 20 to 30 consideration prompts, and 5 to 10 brand evaluation prompts, tracked across two to three AI models for at least 30 days. The article's central framing is worth carrying forward: prompt tracking is a brand intelligence tool, not a performance marketing metric. The volume data is directional; the patterns over time are what matter.