a16z in-depth report: In the AI paid market, businesses have emerged that don't need to make it onto the mainstream traffic charts.

The consumer AI market is forming a clear dual-track structure: traffic prosperity and payment concentration coexist, and the real commercial value is hidden in the latter. The latest a16z Top 100 Consumer AI Apps tracking report shows that only 7 companies appear simultaneously on the three lists of web traffic, mobile monthly active users, and consumer spending, while 29 of the top 50 vendors by spending do not appear on any traffic list at all. The payment side exhibits extreme power-law characteristics. The top 1% of paying users contribute 19.5% of observable spending, higher than the 16.6% combined share of the bottom 50% of users; these heavy users spend an average of $903 per month on AI, mainly purchasing coding, productivity, and creation tools. As of August this year, only 4.5% of U.S. consumers had at least one active personal paid subscription to ChatGPT, Gemini, or Claude. This structure is influencing the strategic direction of the entire industry. Subscriptions helped AI companies cover model costs in the early days, but they also lowered the ceiling for adoption. New business models such as personal agents, advertising, and transaction commissions have gradually emerged, pointing to the same question: how to turn AI from a paid piece of software into a low-threshold service in ordinary people's daily lives. The lists are becoming more stable, and payment data has become an independent dimension for the first time This report introduces for the first time AI spending data observed by YipitData based on U.S. consumer debit and credit card panels, placing it alongside the existing Similarweb web visit volume and Sensor Tower mobile monthly active users to form three independent observation dimensions. The report also excludes products mainly aimed at NSFW scenariosunder the previous methodology, such products would occupy more than 20% of the web traffic list. The liquidity of the lists themselves has clearly declined. This issue added only 11 first-time products, the fewest in seven tracking periods. a16z analyst Olivia Moore summarized the core contradiction behind this phenomenon: "Consumer AI use is broad, but for almost everyone it is not deep." Nearly half of U.S. consumers say they use AI, but only 25% reach daily usage frequency. The mismatch between payment data and traffic data is the most noteworthy finding of this issue. Only 7 companies appear simultaneously on the web, mobile, and spending lists: ChatGPT, Claude, Suno, Perplexity, Photoroom, Canva, and Notion. ChatGPT is the only product ranked first on all three lists. Claude bets on high-value subscriptions, fighting a different battle from ChatGPT ChatGPT's overall lead remains undisputed. In August, its web visits were about 2 times those of Gemini and 6 times those of Claude; mobile monthly active users were 2.5 times those of Gemini and 14 times those of Claude; in U.S. consumer paid subscriptions, ChatGPT also maintained about a 3x advantage over Claude or Gemini. Claude's growth logic, however, shows a clear gap with its traffic ranking. In U.S. consumer subscriptions, Claude had already surpassed Gemini earlier this year; even after Google migrated historical paying users to AI plans to boost subscription numbers, the two were only roughly tied. Claude had not yet entered the web list in the first issue in September 2023, and now its traffic has surpassed DeepSeek and Perplexity. The more critical difference is reflected in subscription structure. Claude's most expensive personal subscription, Max, starts at $100 per month, and 7.3% of its consumer paying users have chosen this tier; Google's corresponding tier accounts for 1.3%, and ChatGPT's 1.1%. Without relying on advertising, Anthropic is more actively mining subscription depth. However, Claude's growth is not linear. YipitData's global desktop panel shows that Claude's average daily sessions declined in July and August; in the U.S. electronic receipt panel, new additions slowed and churn rose. At the same time, ChatGPT reaccelerated after launching GPT-5.6 Sol, Terra, Luna, and ChatGPT Work in July. The three companies are in fact fighting three different battles. Over the past six months, Anthropic, Google, and OpenAI together released 127 new products: Anthropic bet almost entirely on prosumers, such as Claude Design, Code Review, and Claude Science; Google leaned toward creative models, such as Lyria 3 Pro and Gemini Omni; OpenAI covered both enterprise and consumers, rolling out everything from ChatGPT Work to image, health, personal finance, and job-seeking tools. Top 1% of users support the payment base, multi-subscription behavior is rare The payment base is expanding, but concentration has not diluted. In YipitData's U.S. electronic receipt panel, 4.5% of eligible consumers subscribed to at least one mainstream AI product in August this year, up from 2.1% a year earlier. Among users who already pay, multi-product subscription behavior is also rare. Of those who pay for one AI product, only 13% also pay for another AI tool. Most users' decision logic is not to build an AI tool matrix, but to find a product that solves a clear problem. The profile of the top 1% of paying users is close to that of prosumers: they frequently purchase automation and product-building tools, such as n8n, fal, Manus, and Nous Research's Hermes Agent, while also showing a clear preference for creation tools, such as Higgsfield, Figma, and HeyGen. These users spend an average of $903 per month on AI, and what they are buying is essentially productivity, not entertainment. This structure directly explains the mismatch between the traffic list and the revenue list. Among the top 50 vendors by spending, 29 do not appear on any traffic list at all, indicating that building deep usage among specific professional groups is already enough to support considerable commercial scalewithout relying on mass traffic. Personal agents compete for the transaction entry point, and platforms have begun taking sides Personal agents are seen as the direction closest to a new commercial entry point. Six months ago, such products were not mature enough for most consumers; now, startups such as Instinct, Tomo, Poke, Lindy, and Town claim to have hundreds of thousands of users. Meta launched Muse, OpenAI launched Dots, and xAI launched Grok Bot, as big companies have also followed suit. iMessage has become the current popular channel. Instinct started gaining traction there in August 2026, and DoorDash also launched its own iMessage ordering agent. What these products are jointly competing for is the entry point: being able to send texts, emails, and make calls like a real human assistant. The early user profile is still skewed toward hardcore users. After tracking 7 early user group chats with a total of 1,502 samples, AssistantBenchmark founder David Pawlan found that the most frequently discussed tasks were coding and technical automation. Agents are still at the stage of handling workflows for technical users and have not truly entered the daily lives of ordinary consumers. Instinct provided a set of aggressive data: founder Noah Shinn said 40% of users link a personal credit card within three weeks; once a purchase occurs, these users spend an average of $1,300 per month through Instinct; the company's annualized transaction volume is said to have exceeded $1 billion, half of which comes from travel. Platform attitudes have diverged. Amazon cut off Meta Muse's access in less than two weeks; while Shopify, Instacart, OpenTable, Expedia, Ticketmaster, Plaid, and others signed official integrations. Meta Muse launched on September 9, reportedly reaching 250,000 daily active users in its first week and more than 5 million downloads in less than a monthbut this scale is still far below the launch speed of Meta Threads at the time, which exceeded 15 million downloads in the first 22 days in the same two markets. Big tech expansion instead carves out survival space for startups Mature products have entered at scale. Canva and Notion entered the top ten of the web list, and Figma entered the top fifteen; Superhuman (formerly Grammarly) ranked fourth on the spending list for the first time. Google alone has five entry points on the web: Gemini, NotebookLM, AI Studio, Labs, and Antigravity. The full-scale expansion of big tech has instead made startups' differentiation paths clearer. A differentiated model is the first viable direction. Suno ranked 19th in web traffic and 7th in spending; ElevenLabs ranked 25th on the web and 10th in spending. Midjourney, HeyGen, Kling, and Topaz Labs also entered the spending list. In creative tasks, unique voices, visual styles, and training data can sometimes offer a greater competitive advantage than general large-model capabilities. A multi-model experience constitutes the second path. Cursor supports multiple models, and OpenRouter lets developers connect to different providers. Users' core need is not to be tied to a particular model, but to find the most suitable capability for a specific task. Lovable, Cursor, Base44, and Replit have all entered the web list; Cursor rose from 41st to 35th, and Base44 entered the list for the first time at 46th. A clear audience positioning is the third direction. OpenEvidence serves doctors and entered the web list at 47th; Venice focuses on private AI and entered the list for the first time at 43rd. Specific workflows, privacy protection, and content credibility may be as important as underlying model capabilities. Adjustments to content policies also leave market gapsNSFW products being excluded from mainstream lists does not mean demand has disappeared. Redoing the product experience is the fourth path, with the goal of freeing users from the blank prompt box. Plaud combines a hardware device with a subscription and entered the consumer spending list for the first time at 16th. Meta extended Muse to smart glasses and AI electronic pets, while OpenAI's acquisition of io Products hints at a larger hardware layout intention. Subscriptions support the early stage, while advertising and transaction fees point to a larger market Top consumer AI products almost all charge. Among the 44 AI-native products on the web list, all have achieved some form of commercialization: 84% offer paid subscriptions, 64% use usage-based charges or extra credits, only 14% have advertising, and 2% make money through transactions or platform fees. This is the exact opposite of the profit structure of the previous generation of consumer internet. Advertising contributed 97.6% of Meta's revenue and 73.2% of Alphabet's revenue; among the world's top 20 consumer subscription products, 65% belong to media businesses. The practical reason AI companies choose subscriptions is the high cost of models, making it impossible in the early stage to rely entirely on scale for revenue. But other models have already begun to appear. OpenAI said ChatGPT advertising reached a $1 billion annualized revenue run rate in August, based on 1.2 billion weekly active users. OpenEvidence estimates it covers 50% to 60% of U.S. doctors and has also begun to rely partly on advertising for monetization. Transaction commissions are seen as the natural charging path once the agent model takes hold: after users complete attributable purchases through AI, settlement is based on affiliate commissions or take rates. Both Instinct and Muse have described plans to charge from transactions in the future, but neither has implemented them yet; purchases initiated within ChatGPT are currently also not charged. The next key question for consumer AI has shifted from "whether anyone is willing to pay for AI" to "how to get people who are unwilling to subscribe to software to start using it frequently." Subscriptions will continue to serve professional user groups, while advertising, transaction fees, and agent entry points are more likely to bring AI into broader daily life. This article is reprinted from "Wall Street See" (Wallstreetcn), author: Zhao Ying; GMTEight editor: Liu Jiayin.