The Top 100 Most Used AI Apps in 2025
Summary
- Consumer AI is stabilizing without becoming static: only 11 of 50 web properties were new, versus 17 six months earlier. Olivia Moore’s fifth semiannual ranking measures global monthly web visits and mobile active users—“usage, not revenue”—revealing durable attention beneath the launch-cycle noise.
- Vibe coding has become a credible consumer-prosumer category, with Lovable and Replit reaching the main web list while Bolt sits just below it. Lovable announced $100 million in ARR, and many leading platforms showed “100% or above” revenue retention during their first three months before potentially flattening below 100%, suggesting users may upgrade as projects become useful rather than merely sampling trials.
- Google placed four distinct properties on the web list, showing that distribution and product segmentation can challenge ChatGPT without displacing it outright. Gemini ranked No. 2 with roughly 10% of ChatGPT’s web traffic but about half its mobile traffic; AI Studio reached the top 10, NotebookLM ranked No. 13, and Google Labs—likely propelled by Veo 3—ranked No. 39.
- China’s AI ecosystem now supplies domestic champions, export-oriented products, and globally used agents simultaneously. Qwen, Doubao, and Kimi each ranked in the web top 20, while Chinese image and video models often reach U.S. users through intermediaries; Manus, meanwhile, announced a $90 million annualized run rate with Brazil as its largest traffic source and the U.S. second.
- Companionship remains one of consumer AI’s hardest categories to displace on mobile. JuicyChat, Joyland, and DreamGF joined repeat entrants including Character.AI, Janitor AI, SpicyChat, PolyBuzz, CrushOn.AI, A-Dot, and Candy AI—a dense installed base that Justine said would be difficult to displace.
- Model ownership is not the only route to durability: more than half of the long-running “all-stars” host or use other companies’ models, or act as model aggregators. Durability increasingly comes from the interface, community, reusable assets, and bottoms-up enterprise adoption—“the UI and the product experience matter just as much as the model.”
- The next usage wave may come from accuracy-sensitive products rather than another creative novelty. Better math, logic, reasoning, and reliability from Grok 4, the new version of Claude, and GPT-5 might tip financial modeling, presentations, health, education, and personal finance from “cool but inaccurate” into habitual use—but the hosts’ overriding forecast remains that consumer breakouts contain “so much randomness.”
Deep dive
1. The leaderboard is maturing, but attention still turns over
Olivia’s methodology spans every global website and mobile app, ranked through Similarweb monthly visits and Sensor Tower monthly active users, before selecting the top 50 AI-native properties from each. The deliberate lens is “usage, not revenue,” including free users who reveal what is “capturing consumer imagination.”
The web list is the cleaner signal for tracking change over time; the mobile list is less stable because app stores continually change their policies. Early mobile editions were dominated by direct ChatGPT copycats, but iOS policy changes have since altered that mix.
The first editions were “total chaos”: roughly half the list could change, traffic swung fast enough that data delays altered rankings, and many products spiked for one month before disappearing. This edition brought 11 new web names versus 17 six months ago, evidence that the market is beginning to stabilize.
Fourteen properties have appeared on every web edition: ChatGPT, Perplexity, Poe, Character.AI, Midjourney, Photoroom, Leonardo, Cutout Pro, VEED, ElevenLabs, QuillBot, Gamma, Hugging Face, and Civitai. Their endurance spans assistants, companionship, creative workflows, productivity, and model hosting.
2. Google’s four-property debut reframes the assistant race
Domain structure previously prevented independent measurement, but four unique Google properties now qualify. Gemini debuted at No. 2, with about 10% of ChatGPT’s web traffic yet roughly half its mobile traffic, mostly from Android users.
Google AI Studio, the developer-facing model sandbox, entered the top 10. NotebookLM ranked No. 13 after maintaining flat or rising monthly traffic since its September 2024 viral moment, apart from a slight summer decline that Olivia thought was caused by reduced academic use.
Google Labs ranked No. 39, and traffic rose 15% during Veo 3’s release month. Justine’s hunch was that Veo 3 drove much of the gain, though Labs also houses products including Doppl, Portrait, Whisk, and Project Mariner.
The broader assistant market is still verticalizing: users may choose Gemini, Claude, ChatGPT, or Perplexity for different jobs. Grok arrived near the end of the measurement period yet debuted at No. 4 on the web list; Olivia guessed that its companion and image-and-video features drove much of its app growth. Meta AI also began making a small dent on the web, but not on mobile.
3. Chinese AI reaches the ranking through three distinct routes
China’s domestic ecosystem now has its own scaled assistants because products such as ChatGPT and Claude are banned there and would have to comply with local regulatory, policy, and data-capture requirements. Alibaba’s Qwen, ByteDance’s Doubao, and Moonshot AI’s Kimi all placed in the web top 20.
Export-oriented Chinese products include DeepSeek and image or video offerings such as Hailuo and Kling, while Krea is an example of a U.S. property for images. Olivia highlighted particular strength in image and video, especially animation from a single frame or image, and said ByteDance’s Seedance “actually outscores Veo 3 on some of the arenas.”
Distribution obscures the underlying model share: U.S. properties such as Krea and Hedra bundle Chinese models, while Fal.ai and Replicate offer developer access to closed and open models including Kling, MiniMax, Wan, and Qwen. That usage benefits Chinese suppliers without necessarily appearing under their own consumer domains.
Manus represents the third route—built in China and used both domestically and abroad. It announced a $90 million annualized run rate; Brazil supplied its most traffic and the U.S. ranked second. Justine added that Brazil has more free AI traffic than paid traffic.
4. Vibe coding has revenue quality, but its output remains hard to see
Six months earlier, Bolt alone represented vibe coding; now Lovable and Replit made the main web ranking while Bolt moved to the brink list. The category’s emergence breaks beyond the long-standing trio of general chat, companionship, and creative tools.
Consumer Edge cohorts showed many leading platforms retaining “100% or above” of revenue through the first three months, before maybe flattening somewhat below 100%. Justine suggested that users may begin cheaply, publish something useful, then buy credits or upgrade—behavior consistent with both enterprise prototyping and serious solo builders.
Traffic to Lovable and Replit’s creation tools exceeds traffic to projects hosted on their domains. Either successful builders move to custom domains, or users are making “personal software” whose value to themselves, friends, or family is high despite modest public reach.
No mobile vibe-coding product ranked yet; none reached a high enough position to make this list. The mobile list remains crowded with companionship, including three newcomers alongside numerous repeat entrants.
5. Workflow, community, and accuracy are becoming the moats
The hosts revisited the early assumption that only companies spending tens of millions on proprietary models could retain attention. More than half the all-stars actually host or use other people’s models, or are model aggregators, demonstrating that “the UI and the product experience matter just as much as the model.”
Hugging Face and Civitai add non-data network effects through hosted models, LoRAs, datasets, rankings, comments, and mini-apps; ElevenLabs similarly benefits from “thousands and thousands of voices” uploaded to its library. These assets make the product richer as participation grows.
Gamma, ElevenLabs, and Photoroom are also “graduating to enterprise usage” through team plans, templates, and design libraries. One enthusiast can become “their own self-serve enterprise sale,” eventually revealing 10, 15, or 20 individual users who justify a company contract.
The hosts expect reliability gains to expand prosumer productivity, where “hallucinations are a real problem,” unlike creative products where “hallucinations are the feature.” Manus already qualified; Perplexity Comet and Genspark might follow, with health, education, personal finance, social, and structured coaching, therapy, and wellness among the open categories. Yet Justine’s larger prediction is that one or two entirely unexpected categories will emerge, because consumer breakouts contain “so much randomness.”