Gemini falls to 1.9% share as ChatGPT and Claude pull ahead

Three data sources rarely agree. When they do, the finding is hard to dismiss.

The Decoder reports that Pangram Labs tracked Gemini's share of AI model usage collapsing from 12% to 1.9%, while OpenAI held above 50% and Anthropic's Claude climbed from 4.3% to 14.9%. Similarweb traffic data and OpenRouter's API routing logs point in the same direction. This is not one analyst's read on a noisy market. It is a convergent signal.

The instinct in most marketing departments will be to treat this as a technology story: Google loses a product race, engineers update a roadmap. That instinct is wrong. For any brand whose authority strategy depends on appearing inside AI-generated answers, the distribution of model usage is a strategic variable, not a spectator sport.

Which model your audience uses determines which sources it trusts

Citation behaviour differs materially across models. ChatGPT, Claude, and Gemini do not retrieve and cite the same sources in the same proportions, partly because their training data, retrieval architectures, and reinforcement fine-tuning differ, and partly because the publishers and datasets each company licensed vary. A brand that built its LLM-visibility strategy primarily around optimising for Gemini's citation patterns over the past 18 months may have spent that budget on a shrinking audience.

The practical consequence is asymmetric. OpenAI's 50%-plus share concentration means that appearing credibly and consistently in ChatGPT's outputs is now the dominant objective. Claude's rise to nearly 15% is significant because Anthropic's user base skews toward developers, researchers, and knowledge workers: precisely the professionals who staff multilateral institutions, financial services strategy teams, and policy bodies. For organisations like CGAP or UNDRR, Claude is not a secondary model. It may already be the primary interface through which their staff research and draft.

Gemini's collapse from 12% to 1.9% is steep enough to warrant a concrete reallocation question: if roughly one in 12 AI interactions ran through Gemini a year ago and fewer than one in 50 do today, content or technical investments made specifically to improve Gemini retrieval deserve fresh scrutiny. That is not a reason to ignore Gemini entirely. Google's distribution advantages, Search integration, Workspace embedding, and Android defaults mean the model could recover share quickly if product problems are fixed. But the data as it stands argues against treating Gemini parity as a near-term planning assumption.

The Claude variable most brands have not priced in

The more underappreciated finding is Anthropic's growth. A rise from 4.3% to 14.9% over a comparable period represents a more than threefold increase in share. Claude's users tend to run longer, more analytical queries: the kind where a model cites external sources to support a multi-paragraph synthesis rather than returning a quick factual answer. That citation surface is where brand authority either accretes or disappears.

Organisations in financial services, industrial groups, and the UN system produce exactly the kind of dense, technical, policy-adjacent content that Claude users are asking about. If that content is structured to be legible to AI retrieval systems and is already visible in the sources Claude trusts, the model's growth is an opportunity. If it is buried in PDF archives or behind authentication walls, the opportunity flows to whoever wrote the accessible version.

The broader principle the Pangram data enforces: model-agnostic visibility is increasingly the right frame. A brand that appears consistently across ChatGPT and Claude, while maintaining basic hygiene for Gemini, holds more durable AI-search positioning than one optimised for a single model's quirks. Distribution shifts of the magnitude Pangram documents can happen fast. A brand's cited-source footprint needs to be diversified enough to survive them.