We analyzed LLM market share by combining usage-based data and web visit estimates to show how demand for large language models is distributed across AI labs and AI applications:
- The United States dominates global LLM usage in web visits and brand adoption, driven by ChatGPT and Gemini, while China operates largely behind the scenes. China generates high developer usage through API and programmatic workloads, rather than through visible consumer activity.
- OpenAI’s ChatGPT remains the leading application, but is steadily losing share to Google’s Gemini, which demonstrates the most balanced and sustained growth across both consumer and API usage.
- Chat-based applications account for nearly the entire AI market, while other categories remain niche and fragmented.
LLM market share comparison by country
Read the methodology to see how we measured and calculated these results.
The United States dominated web visits across all four months, consistently accounting for 85.5–90.5%. This reflects both the concentration of consumer AI products in English-speaking markets and the US base of most major platforms.
China shows a different usage profile:
- In November 2025, China accounted for 50.9% of developer usage without BYOK but only 7.5% of web visits. This implies heavy API or programmatic usage rather than consumer activity.
- By February 2026, the gap narrowed. Developer usage without BYOK share dropped to 31.9%, while web visits remained around 8.0%, suggesting more balanced usage or a shift in how AI is consumed.
LLM market share comparison based on applications
ChatGPT remains the market leader, but competitive pressure is increasing. Its market share dropped from 72.5% in October 2025 to 60.5% in February 2026, a 12-point decline over four months.
Gemini captured most of that shift by growing from 13.9% to 23.9%. The increase is most likely to be driven by distribution and steady model improvements.
LLM market share comparison based on AI labs
OpenAI saw the largest shift in the dataset. Its web visits share increased from 19.5% in November 2025 to 57.3% in February 2026, while Anthropic’s web visits dropped from 48.4% to 13.9%.
One of the reasons for the increasing interest in OpenAI can be the introduction of Apps in ChatGPT in November 2025, which included partners like Spotify, Canva, and Booking.com.1
In terms of developer usage without BYOK, Qwen’s strong presence in November 2025, at 39.3%, dropped dramatically to 3.6% in February 2026. In the meantime, Google showed the most balanced growth, increasing from 16.7% to 25.7%.
App categories market share comparison
Chat dominates the entire market, consistently accounting for 88–92% of the market share. Consumer AI usage is still centered on general-purpose conversational interfaces. Within the Chat category, ChatGPT and Gemini accounted for ~84% of the market share in February 2026.
LLM market share methodology
Step 1: Identifying AI labs and AI applications
- AI labs: Organizations that develop and maintain LLMs. For example, OpenAI, Google, Anthropic, DeepSeek, Qwen, X, Mistral AI, and Meta Llama are among the most prominent AI labs developing AI apps.
- AI applications: End-user tools, platforms, or agents that rely on one or more LLMs, such as ChatGPT, Claude, Grok, and Gemini.
Each AI application is then mapped to:
- One or more underlying AI labs.
- One functional category.
Step 2: Calculating app and lab-level market share
We used below metrics to calculate the market share for apps and AI labs:
Developer usage without BYOK
Tracks how many tokens a lab’s models process through OpenRouter’s paid credit system. The criteria exclude requests routed via BYOK (Bring Your Own Key), where developers supply their own provider API keys and are billed directly by the provider rather than through OpenRouter.
A high share here signals that a lab’s models are widely used in developer-built workflows: API-integrated applications, agent pipelines, RAG systems, and code generation tools. It is a supply-side signal tied to developer activity on the OpenRouter platform.
Limitations:
OpenRouter data reflects a self-selecting population: developers who pay for API access to route requests across multiple models. Consumer usage of ChatGPT, Gemini, or Claude via their native apps is not captured; the data primarily reflects requests routed through OpenRouter’s platform.
This matters for interpreting regional figures. OpenRouter’s State of AI report notes that its user base skews heavily toward API-integrated workflows, and that over 50% of usage originates outside the United States. Chinese developers, particularly those building on DeepSeek and Qwen, use OpenRouter actively for programmatic access. This inflates China’s token share relative to its actual consumer presence.2
BYOK also creates a data gap. When developers use their own API keys, OpenRouter routes the request, but the underlying provider bills them directly. As a result, those tokens do not appear in OpenRouter’s platform usage statistics.
This means OpenRouter’s token data misses high-volume enterprise users with BYOK agreements. At the same time, it gives more weight to teams that pay through OpenRouter credits.
Web visits
Tracks visits to AI products’ websites across desktop and mobile web using Similarweb estimates.3 The metric shows how many people actively visit these products online, indicating web-based demand and adoption. It does not measure in-app usage activity.
Limitations: Similarweb applies sampling and inference to produce metrics such as total visits, making estimates directional rather than exact. The metric can also underrepresent private, enterprise-only, or regionally restricted deployments.
Step 3: Categorizing AI applications
We grouped AI applications into categories based on their primary use case. While many tools span multiple functions, we assigned each application a single primary category to ensure consistency in data analysis.
- General-purpose chat: Applications focused on conversational interaction, reasoning, and broad task support. These tools account for a large share of consumer-facing LLM usage and play a central role in customer interactions.
- Programming and coding assistants: Code generation, debugging, and software development workflows. This category is closely linked to enterprise usage and developer productivity.
- Developer platforms and tooling: Platforms that enable developers and enterprises to build, deploy, or manage AI applications. These tools are central to AI integration and are often used by cloud providers and large enterprises.
- Search and answer engines: Applications optimized for information retrieval and synthesis rather than open-ended chat. These tools often combine language models with internet data to achieve higher accuracy.
- Vision and multimodal generation: Applications focused on image and video generation or understanding, often used in content creation and media-related industry verticals.
- Audio and speech: Tools centered on voice generation, speech interaction, or audio processing.
- Gaming and interactive AI: Applications designed primarily for entertainment, role-play, or interactive experiences.
FAQ
ChatGPT holds the largest share of consumer web visits among AI applications, accounting for approximately 60.5% of visits across tracked platforms as of February 2026, down from 72.5% in October 2025.
At the lab level, OpenAI leads on the power index with 51.8%, driven by its models being used across both its own products and third-party applications.
Yes. ChatGPT’s web visits dropped 12% between October 2025 and February 2026. Google’s Gemini captured most of that shift, growing from 13.9% to 23.9% of web visits over the same period. The decline reflects increasing competition rather than a collapse in usage, as total AI web visits continued to grow.
The United States accounts for 85.5–90.5% of web visits to AI applications. However, that figure measures consumer-facing usage. At the token level, which captures API and programmatic usage, China’s share is significantly higher than its share of web visits, reflecting heavy developer and infrastructure use of models like DeepSeek and Qwen.
There is no single standard. We used three signals: web visits from Similarweb, which count visits to AI applications; developer usage without BYOK from OpenRouter, which counts tokens processed through developer API routing.
Each metric captures a different layer of the market. Web visits reflect consumer adoption, and developer usage without BYOK reflects developer and infrastructure usage. No single metric covers the full picture, and all three have limitations.
Reference Links
Be the first to comment
Your email address will not be published. All fields are required.