Gemini Traffic Doubles, ChatGPT Ads Favour Clarity & GEO Tactics That Damage SEO

SE Ranking tracked referral traffic from seven major AI platforms across 101,574 websites with Google Analytics over 12 months. The headline number is Gemini's growth: between November 2025 and January 2026, Gemini referral traffic increased 115%. In the same period, ChatGPT declined roughly 8% per month. The gap between them compressed from 22x in October 2025 to 8x by January 2026. By January, Gemini also surpassed Perplexity, sending 29% more traffic globally and 41% more in the US. ChatGPT still dominates at around 80% of all AI-sourced traffic, but the direction of travel is clear.

The Gemini 3 rollout in November and December 2025 appears to have been the catalyst. The study notes this directly: traffic growth coincided precisely with the product update. That pattern matters because it suggests Gemini traffic growth is tied to Google's product decisions, not just broader AI adoption. If Google continues improving Gemini and integrating it more deeply into Search, the growth trajectory could accelerate further. The conservative modelling in the study projects a potential overtake of ChatGPT as early as October 2026; the aggressive scenario puts it at June 2026 if current growth rates hold.

Adthena analysed over 40,000 daily ad placements on ChatGPT and found a consistent structural pattern in top-performing creatives. Headlines average 30 characters and around 5 words, almost always leading with the brand name. Body copy averages 116 characters and around 19 words, structured as two tight sentences: a proof point followed by an offer or conversion nudge. Concrete numbers, dollar signs, and specific value claims consistently outperform vague or emotional language. Free trial and demo offers are the dominant conversion mechanism. The tone is calm and measured, with exclamation marks essentially absent from high-performing ads.

The implication is that ChatGPT's ad environment rewards utility over storytelling. Users are in a task-completion mindset, not a browsing one, and the ads that perform reflect that. This is essentially the inverse of what works in traditional display or social advertising, where creative differentiation and emotional hooks tend to drive performance. The report frames it as "precision over persuasion" and that framing is useful because it connects to the same logic that drives AI citation behaviour more broadly: specific, clear, credible content wins over creative or stylised alternatives.

Lily Ray documents five GEO tactics that are causing organic traffic collapses: scaled AI content production, artificial date refreshing, self-promotional listicles (ranking your own product at number one in category comparisons), prompt-injection "summarize with AI" buttons, and excessive comparison pages. In each case, she has traffic data showing the pattern: rapid early growth followed by a hard drop during Google algorithm updates, with some sites losing visibility permanently. One company had to remove the case study articles that had celebrated its growth after subsequent traffic crashes made them embarrassing.

The structural argument is the most important part of the piece. Most major AI search products, including ChatGPT, use retrieval-augmented generation (RAG) that pulls from external indexes before generating responses. Britney Muller's cited research puts it plainly: "Every single URL you see in LLM output comes from search engine API." Ray also presents data showing a correlation between organic ranking drops and ChatGPT citation declines among affected sites. If you lose organic visibility, you are likely losing AI citation visibility at the same time. GEO is not an alternative to SEO, it depends on it.

Forrester's argument is that llms.txt solves a narrow problem but creates a maintenance burden and cannot express the relationship structures that AI systems actually need. Every product update, pricing change, or new case study requires updating both the live site and the file, and the format provides no way to signal hierarchy, deprecation, or authority relationships between entities. The piece proposes a four-layer architecture as the next step: enhanced JSON-LD structured data on commercial pages, entity relationship mapping (JSON-LD extensions or headless CMS endpoints that express how products connect to categories and solutions), versioned content API endpoints for FAQs and documentation, and provenance metadata attaching timestamps, authorship, and version information to every exposed fact.

The data point that anchors the practical case: pages with valid structured data are 2.3x more likely to appear in Google AI Overviews. Forrester also flags Anthropic's Model Context Protocol as the emerging standard for content API endpoints, suggesting that the infrastructure you build for AI-readable content now will need to be compatible with MCP as it becomes more widely adopted. The minimum viable approach he recommends for this quarter is a JSON-LD audit of commercial pages, a single structured content endpoint for frequently compared information, and provenance metadata on public-facing facts.

Natividad's central argument: search engines capture existing demand, they do not create it. The implication is that the upstream work of building distributed public evidence, across forums, reviews, community discussions, and third-party publications, is what determines whether a brand is visible in AI answers before anyone ever opens a search bar. The Seer Interactive case study makes this concrete. One negative review theme appeared 67 times in AI outputs before Seer published counter-evidence. After publishing retention data publicly, LLMs immediately stopped citing the negative claim. The evidence that existed in the public record shaped the AI answer; new evidence changed it.

The search behaviour data reinforces the point. In Q4 2025, Google accounted for 73.7% of all searches across 41 major US websites, but Reddit simultaneously outranked SaaS vendors on 50 to 66% of shared keywords across three of four verticals analysed. For queries of six words or more, Reddit's advantage reached 73 to 100%. User-generated content on third-party platforms is winning organic rankings and, by extension, the AI citations that follow from those rankings. Brands that only publish on their own domains are competing at a structural disadvantage on both fronts.