Which AI products have real recurring usage now?

Has AI usage actually become a habit now?

Yes. AI usage has clearly become recurring for a large group of consumers and workers, although the habit is concentrated in a much smaller set of products than the size of the AI market suggests.

Sensor Tower's 2026 consumer data gives us one of the cleanest checks. Time spent in generative-AI apps was projected to rise from 17.2 billion hours in the first half of 2025 to 36 billion hours in the equivalent 2026 period. Generative-AI websites also generated more than 67 billion visits during the first quarter, while total time spent reached roughly 23 billion hours. Time spent grew 41% year over year, faster than visits at 28%.

The longer-term mobile data points in the same direction. Generative-AI apps produced more than one trillion sessions during 2025, and session growth was faster than download growth. People were increasingly returning to apps they already had rather than simply installing new ones.

Work usage has deepened too. OpenAI's enterprise study found weekly ChatGPT Enterprise messages increasing about eightfold over a year, while the average existing worker sent roughly 30% more messages. Ramp's latest spending data shows that well over half of businesses using either Anthropic or OpenAI use both, which tells us that AI is becoming a recurring software category rather than a one-product experiment.

The habit is real. The more useful question is which products people keep reopening once the novelty has disappeared.

Is ChatGPT still the clearest example of recurring AI usage?

Yes. ChatGPT still has the strongest overall case for recurring AI usage because it combines more than one billion weekly users with evidence that existing users are doing more inside the product.

OpenAI says ChatGPT now reaches more than one billion weekly active users. Sensor Tower independently found that the mobile app reached one billion monthly active users earlier this year, faster than any previous mobile app.

The weekly figure is especially useful here. Monthly active users can include someone who opens an app once every few weeks. A billion weekly users means the recurring audience itself is enormous.

The workplace data makes the case stronger. OpenAI found aggregate weekly Enterprise messages rising about eightfold, while messages sent by the average worker increased roughly 30%. Weekly users of Projects and Custom GPTs rose around nineteenfold, and roughly one fifth of Enterprise messages were already being processed through those reusable environments.

People are increasingly returning to the same Projects, instructions, files and custom assistants instead of starting every conversation from zero.

ChatGPT may lose individual categories such as coding to more specialized products. Across the whole AI market, though, no other standalone product currently shows the same combination of scale, weekly recurrence and increasing usage per existing user.

Recurring-use measure	What we see in ChatGPT
Weekly active users	More than 1 billion
Mobile monthly users	Roughly 1 billion reached
Enterprise message volume	About 8× growth over one year
Messages per existing worker	About 30% higher
Projects and Custom GPT weekly users	About 19× growth

Is Gemini actually becoming a daily habit?

Yes. Gemini has moved well beyond Google simply putting an AI product in front of a huge audience, with daily usage growing much faster than its already enormous monthly reach.

Google says the Gemini app has passed one billion monthly active users. Shortly before crossing that threshold, Alphabet reported that Gemini daily active users had tripled in a year while monthly active users had a little more than doubled.

Distribution through Android obviously helps create a huge monthly audience. It is harder to explain daily users tripling in a year through distribution alone.

Other behavior looks surprisingly deep. Google says 63% of Gemini users use voice, one in five Gemini Live interactions involves a camera or screen sharing, and more than 100 million Gemini users are active on iOS. Mac power users prompt roughly twice as frequently as users on other surfaces.

Gemini also generates more than 150 million images per day. Earlier Google disclosures showed that daily requests had increased more than sevenfold over roughly a year, considerably faster than the user base.

We should still be more cautious with Gemini than ChatGPT when comparing pure product stickiness because Google controls Android, Chrome, Search and Workspace. Yet Gemini's current numbers are too large and too frequent to explain away through distribution. It has become a real recurring consumer product.

Are Claude users actually sticking around?

Yes. Claude has some of the strongest recurring usage in professional AI, and the freshest business-spending data suggests that its position is still getting stronger.

Ramp's latest AI Index found 43.8% of U.S. businesses in its sample paying Anthropic for subscriptions or tokens, compared with 39.8% paying OpenAI. Earlier in the year, Anthropic was at 24.4%. That is an unusually fast shift for a business-software vendor.

The growth has continued even as overall new AI adoption has started slowing. Anthropic is gaining share in a market where simply adding another wave of first-time AI buyers is getting harder.

Consumer economics have improved at the same time. Sensor Tower found Claude's U.S. mobile revenue per user climbing from below $0.50 in September 2025 to $2.76 by May 2026. Claude's U.S. "true audience" share across mobile and web more than tripled as well.

There is an important limitation. Ramp measures purchases, so a company paying Anthropic does not tell us how often every employee uses Claude. Still, the persistence of the growth, rising consumer monetization and heavy use of Claude Code all point in the same direction.

Claude looks much stickier than its smaller consumer audience might imply.

Has Claude Code become one of the stickiest AI products?

Yes. Claude Code is one of the strongest examples we have of AI becoming part of someone's working day rather than something they occasionally open.

The product's economics reached unusual scale quickly. Anthropic said Claude Code passed a $1 billion annualized revenue run rate only around six months after general availability, with subsequent growth pushing it much higher.

Revenue by itself would leave room for doubt because coding agents can burn through a lot of tokens. The way people use Claude Code makes the evidence stronger. Anthropic has repeatedly increased usage allowances, and its research increasingly separates long-running agentic sessions from normal conversational use because programming sessions can involve extended sequences of file reading, edits, commands, testing and debugging.

A developer can interact with Claude Code dozens of times while working on one feature, return during debugging and use it again on the next task. Recurrence is built into the job.

Anthropic's wider enterprise position supports the same reading. Ramp now shows Anthropic leading OpenAI in paid U.S. business adoption, and Anthropic has built much of that growth around coding and technical work.

Claude Code belongs in the very small group of AI products where repeated use is part of production itself.

Are Cursor users really coming back every day?

Yes. Cursor has unusually strong evidence of daily professional usage because we can see the output flowing through the product every day.

Cursor currently says more than 50,000 enterprises use the product, including 64% of the Fortune 500. It also reports more than 100 million lines of enterprise code being written with Cursor per day.

That daily output is much more useful for our question than another ARR milestone. Someone has to keep opening Cursor, prompting it, accepting or changing its code and working around the output for those lines to continue appearing.

Earlier in Cursor's growth, the company had already disclosed more than one million daily users. Since then, enterprise adoption and revenue have increased substantially rather than fading after the original AI-coding rush.

Cursor also benefits from where it sits. Developers live inside their editor for hours. Once AI becomes part of searching a repository, generating code, refactoring, debugging and reviewing changes, there are many reasons to invoke it again before the day is over.

Competition around AI coding is brutal, so Cursor's future lead is far from guaranteed. Its present recurring usage is much easier to defend: people are clearly using the product as a working environment rather than treating it as a demo.

Is GitHub Copilot still sticky now that Cursor and Claude Code are everywhere?

Yes. GitHub Copilot remains one of the biggest recurring AI work products, even though Cursor and Claude Code have taken much of the excitement among early adopters.

Microsoft recently said GitHub Copilot had reached 50 million users. Business and Enterprise seats were still growing, and Copilot revenue accelerated more than 60% quarter over quarter after GitHub expanded usage-based billing.

The product is also moving deeper into actual repository work. Microsoft says one in three pull requests on GitHub now involves an agent. That is an unusually strong recurrence measure because pull requests are part of real software delivery rather than a separate AI activity users have to remember to perform.

GitHub itself tracks daily, weekly and monthly active Copilot users alongside accepted code completions, chat requests, code reviews, CLI interactions and agent activity. Enterprises can therefore measure whether Copilot is actually being used instead of relying on seat counts.

Copilot has lost the luxury of being the obvious AI coding choice. It still has a huge advantage from living inside GitHub, IDEs, repositories and enterprise developer accounts.

For now, that distribution is translating into real repeated usage.

AI coding product	Best recurring-use evidence	Current read
Claude Code	Long-running coding sessions and heavy paid consumption	Extremely strong
Cursor	100M+ lines of enterprise code per day	Extremely strong
GitHub Copilot	50M users and one in three GitHub PRs involving an agent	Extremely strong

Are companies actually using Microsoft 365 Copilot after buying the seats?

Yes, although Microsoft 365 Copilot usage looks much less uniform than usage of Cursor or Claude Code.

Microsoft now reports more than 30 million paid Microsoft 365 Copilot seats, twice the 15 million figure it disclosed earlier this year. Around 90% of the Fortune 500 use Microsoft 365 Copilot.

Thirty million paid seats obviously matters, but enterprise software has plenty of examples where companies buy thousands of licenses that employees barely touch. The better evidence comes from what people consume after deployment.

Microsoft says usage-based AI-credit consumption in customer service grew fourfold quarter over quarter. Thousands of customers were already paying for and actively using its newer Cowork product shortly after usage billing started. Across Microsoft's first-party Copilot family, monthly active users have passed 150 million.

A meaningful part of the installed base is clearly doing real work with Microsoft's AI. The intensity probably varies enormously between a finance employee who occasionally summarizes a spreadsheet and someone using agents throughout the day.

We can confidently call Microsoft 365 Copilot recurring software now. We should be less confident that the average paid seat is heavily used.

Does Canva count as an AI product with real recurring usage?

Yes. Canva shows why some of the most durable AI usage may end up happening inside products people were already opening every week.

Canva has more than 265 million monthly active users and around 31 million paid users. Its Magic Studio AI tools have accumulated billions of uses, while AI generation now sits directly inside design, presentations, video, marketing content and image editing.

Canva does not need to teach someone a completely new routine. A marketer opening Canva to make this week's social posts encounters AI inside the workflow they already came to complete. The same applies to a student building a presentation or a small business creating an advertisement.

Standalone image generators have to persuade users to visit a dedicated destination each time. Canva can surface AI repeatedly while users move through work they already do.

The distinction becomes more important as general assistants improve at image and video generation. A creative AI tool with no larger workflow around it faces much more pressure than an AI feature sitting inside a design environment with hundreds of millions of active users.

Canva therefore has one of the strongest structural advantages in recurring creative AI usage, even though most people would still describe Canva as a design product first.

Is Notion AI becoming something teams use repeatedly?

Probably yes. Notion AI increasingly looks like recurring workplace software, although the public usage data is still thinner than what we have for ChatGPT or coding tools.

The strongest clue is how deeply AI has moved into the paid product. Earlier analysis from a16z found Notion's paid AI attach rate moving from roughly 20% to more than 50% within about a year, with AI-related products accounting for a large share of the company's new economics.

Notion has since pushed further into agents that can run on schedules and triggers. A user no longer has to remember to ask Notion AI the same question every week if an agent can perform the job automatically when something happens.

The company now charges credits for some of those agent actions, giving both Notion and its customers a direct way to see whether recurring automations are valuable enough to keep running.

We should still keep the confidence level below Claude Code or Cursor because Notion does not publish comparable daily or weekly AI-user metrics. The product design, attach rate and scheduled-agent model nevertheless make repeated use increasingly likely.

Are Lovable and Replit recurring products, or mostly one-off app generators?

Lovable and Replit both have real returning users now, but app building naturally creates a weaker daily habit than coding inside an IDE.

Lovable is the more interesting test because its early traffic looked vulnerable to an AI-hype cycle. Web activity fell sharply from its initial peak, which would have supported the idea that many people had tried vibe coding and moved on.

The business kept growing anyway. Lovable later reported net dollar retention above 100%, meaning retained customers collectively increased rather than reduced their spending. The company has also said users create roughly one million projects per week and have generated tens of millions of projects in total.

Replit shows a similar transition from experimentation toward running actual software. The company says it has more than 50 million users and is used by employees at 85% of the Fortune 500. Its Agent-driven revenue grew from only a few million dollars annualized to well into nine figures, while Replit increasingly emphasizes production applications, internal tools and deployed businesses.

Neither company gives us a clean public DAU-to-MAU ratio. That keeps us from putting them beside Cursor, where daily code output is observable.

Still, the idea that vibe-coding products are purely one-session toys has become hard to defend. Lovable and Replit have built meaningful recurring cohorts; those cohorts simply return on a less predictable schedule than full-time developers do.

Are people choosing one AI assistant, or coming back to several?

Many serious AI users are coming back to several products, and the freshest business data makes that pattern clearer than before.

Ramp found that 52% of businesses paying either Anthropic or OpenAI now pay both. Earlier in the adoption cycle, the overlap was much smaller. Anthropic has therefore managed to pass OpenAI in Ramp's U.S. business sample without producing an equivalent collapse in OpenAI usage.

The overlap gets even stronger among advanced AI buyers. In Ramp's analysis of businesses using model-serving platforms, 85.8% also used OpenAI and 93.2% used Anthropic. Those businesses spent roughly 23 times more per employee on AI than the median AI-buying company.

Heavy users tend to accumulate AI products. Developers can use Cursor as their editor, Claude Code for agentic programming and ChatGPT for research. A company can buy Anthropic for coding while keeping OpenAI for other teams.

Retention in AI therefore behaves differently from retention in categories such as CRM. A customer can become a regular Claude user while remaining a regular ChatGPT user too.

Behavior	What it tells us
52% of Ramp businesses using Anthropic or OpenAI pay both	Multi-product usage is mainstream among AI-buying firms
85.8% of model-platform users also buy OpenAI	Advanced users keep OpenAI in the stack
93.2% also buy Anthropic	Anthropic is especially common among heavy AI adopters
Heavy adopters spend far more per employee	More AI intensity often produces more tools, rather than consolidation

Which AI products naturally get used most often?

AI coding products currently have the strongest natural recurrence, while general assistants win on the sheer number of different reasons people can come back.

Coding produces an unusually dense loop. Developers read code, write code, debug, run tests, search repositories, review pull requests and make fixes throughout the day. Cursor says more than 100 million lines of enterprise code pass through its product daily. Microsoft says one in three GitHub pull requests now involves an agent. Claude Code can remain active through long sequences of commands and edits.

General assistants work differently. ChatGPT and Gemini can be used for research, writing, school, planning, analysis, images, shopping and everyday questions. Any one task may happen less frequently than coding, but the number of possible return occasions is enormous.

Embedded AI has another route to recurrence. Canva AI appears whenever someone designs something. Microsoft Copilot can appear in Word, Excel, Outlook and Teams. Notion agents can keep running around an existing workspace.

The products with the best recurring usage tend to sit close to work people already repeat. Creating a new habit from scratch is much harder than attaching AI to one that already exists.

Does an AI subscription prove that people keep using the product?

No. An AI subscription tells us someone is willing to pay, but it can badly overstate how often the product is actually used.

This is especially important now because AI revenue metrics have become messy. Some companies annualize a recent month of consumption revenue, even though token usage can jump sharply from one month to the next. An impressive run rate can therefore reflect intense current consumption without telling us how durable that consumption will be.

Enterprise seats have a similar problem. Microsoft selling 30 million Copilot seats is strong commercial evidence. To understand recurrence, however, we learn more from the fourfold increase in customer-service AI-credit consumption or the growth of actual Copilot interactions.

Free usage creates the opposite problem. ChatGPT and Gemini have huge populations of users who pay nothing and still return constantly.

For this article, the best proof comes when several measures agree: users return weekly or daily, existing users do more over time, companies expand consumption, projects persist across sessions and the product sits inside work that has to be repeated.

That gives us much more confidence than ARR on its own.

Are AI agents getting real recurring usage yet?

Yes in software development, and increasingly in other kinds of work, although coding remains far ahead.

GitHub provides the clearest evidence. One in three pull requests on the platform now involves an agent. Claude Code has turned multi-step programming sessions into one of Anthropic's biggest businesses. Cursor has also moved deeper into background and agent-based development rather than limiting AI to autocomplete.

The pattern is starting to spread outside engineering. OpenAI has reported very fast growth in enterprise Codex usage across legal, sales, recruiting and marketing. Those functions started from much smaller bases, so the growth rates should be treated carefully, but people are clearly trying recurring delegation outside programming.

Microsoft is seeing usage-based agent consumption rise inside customer service, while Notion now allows agents to run from schedules and triggers without the user opening a chat first.

Coding is already past the proof-of-concept stage. In broader office work, we can see the habit forming, but we still lack enough long-term data to say that autonomous agents are as routine for a typical knowledge worker as coding agents are for a developer.

Which popular AI products still have weaker proof of recurring usage?

Standalone image and video generators generally give us weaker public proof of broad recurring usage than assistants, coding tools and AI built into existing work software.

Some creative professionals use products such as Midjourney extremely heavily. The difficulty is that public metrics in creative AI often focus on cumulative generations, Discord membership, registered users, traffic or revenue. Those numbers tell us the product is popular, but they reveal much less about how many people return every week for months.

Competition from larger platforms makes the question tougher. Gemini generates more than 150 million images per day. Canva has already processed billions of AI-assisted creative actions. ChatGPT includes image creation inside an assistant people visit for many other reasons.

Dedicated generators therefore have to earn the next visit on the strength of creative work alone.

AI companion products have a different problem. Some produce extraordinary session lengths among committed users, yet averages can hide a smaller group of very heavy users alongside many people who experiment briefly and leave. Cohort retention matters much more than total hours there.

Several of these products are clearly sticky within particular audiences. The evidence for broad, durable recurrence simply remains weaker.

What makes people keep coming back to the same AI product?

The stickiest AI products today remember the work, sit close to a repeated task and save enough effort that reopening them becomes automatic.

Cursor works inside the codebase. Claude Code can continue working across files and commands. ChatGPT Projects preserve context, files and instructions. Notion AI operates around an existing workspace. Canva AI appears inside designs people return to edit. Copilot already lives inside GitHub, Microsoft 365 and enterprise accounts.

Context gives those products an advantage every time someone comes back. Switching means rebuilding at least part of the setup, history, files, preferences or workflow.

Task frequency matters just as much. Software developers code every working day. Office workers write, search, summarize and communicate constantly. Marketing teams keep making content. Products tied to those activities get repeated opportunities to prove useful.

The strongest products also leave behind visible work: code committed, documents produced, tickets resolved, designs created or customer cases handled. That makes their value easier to notice than a chatbot that occasionally produces an impressive answer.

Model quality can change quickly. A product's place inside someone's routine tends to move much more slowly.

So which AI products have real recurring usage now?

ChatGPT has the strongest overall recurring usage today, while Claude Code, Cursor and GitHub Copilot show the clearest high-frequency professional habits.

ChatGPT stands out because more than one billion people use it weekly and existing workplace users are becoming more active. Gemini has also crossed into genuine mass recurrence: its monthly audience has passed one billion, daily users have tripled, and daily requests have grown even faster. Claude is smaller among consumers but exceptionally strong with professionals, with Anthropic now leading OpenAI in Ramp's latest U.S. business-adoption data.

Among work products, coding is the clearest category. Claude Code, Cursor and GitHub Copilot sit inside tasks developers perform all day, and we can see the recurring output directly through agent sessions, daily code generation and pull requests.

Canva and Microsoft 365 Copilot show another durable model: put AI inside software people already use. Notion looks increasingly convincing for the same reason, especially as recurring agents become part of the workspace.

Lovable and Replit have also passed the point where they can be dismissed as launch-driven curiosities. Their recurring usage is real, although building an app is naturally more episodic than writing software every day.

The clearest pattern across all of them is simple. Products become sticky when AI gets attached to something the user will need to do again tomorrow. Right now, those repeated workflows tell us much more about durable usage than downloads, registered users or headline ARR.

Group	Products with the clearest evidence	Our confidence
Strongest overall recurring product	ChatGPT	Very high
Highest-frequency professional usage	Claude Code, Cursor, GitHub Copilot	Very high
Mass recurring assistants	Gemini, Claude	High
AI embedded in existing habits	Canva, Microsoft 365 Copilot, Notion AI	High to medium-high
Real but more episodic recurrence	Lovable, Replit	Medium-high
Harder to prove broadly	Standalone image/video generators, some AI companions	Mixed