AI apps ranked by what people keep using, not what they download
If you are shopping for AI apps, the ranking that matters is not the App Store chart. It is the much shorter list of apps still open on your phone a month after you installed them. As of September 2026 that list is narrow and it has a shape: almost everyone who keeps anything keeps one general assistant, and a growing minority keeps one specialist next to it. Adoption has stopped climbing. Depth has not.
We could not install forty apps and watch them for a fortnight, so this piece does the next best thing. It pulls the retention and usage evidence that firms with panel data have published, adds a live App Store snapshot we took ourselves, and says plainly where the numbers stop. Every figure below comes from a page we opened, and the date of each measurement is named, because these numbers move fast.
Adoption stalled at 64% while spending tripled
The single most useful number this year is the one that barely moved. Menlo Ventures surveyed 5,067 US adults with Morning Consult in July 2026 and found 64% had used AI, against 61% a year earlier. Globally the firm estimates the user base grew 11%, from 1.8 billion to 2.0 billion people. So that is a category which has run out of easy new customers.
What did move is what those people do once they are inside. Menlo estimates global consumer AI spend reached $40 billion in 2026, more than triple the $12 billion it measured a year earlier, and daily use among all US adults rose from 19% to 25%. And 55% of AI users now pay for at least one AI product, which is the figure that moved most. The spending is lopsided: the 14% paying $100 or more a month account for 60% of the total.
That matters because it changes what a good AI app has to do. Growth is no longer about getting installed. It is about being the thing someone opens on a Tuesday for the fourth week running, which is also the moment a free tier stops being free enough. We worked through what each free plan actually costs in our breakdown of AI free tiers, and the pattern holds here.
One assistant sticks, and three more sit in the rotation
Pew Research Center surveyed 5,119 US adults from 17 to 23 February 2026 and found 49% ever use AI chatbots, up from 33% in 2024. About a quarter use one daily, made up of 8% about once a day, 12% several times a day and 4% almost constantly. Broken out by brand, ChatGPT reaches 44% of US adults, and nothing else is close to that.
Menlo asked a different question of a different group, and the gap between the two is the story. Among people who already use AI, Menlo found ChatGPT at 60%, up from 45%, but Gemini at 58%, up from 36%, which cuts a nine-point lead to two. Claude nearly tripled, from 7% to 20%, though it started from a much smaller base. The average AI user now reaches for 3.0 general assistants, up from 2.2 a year ago, so these shares no longer add to a hundred.
| Assistant | Share of all US adults who ever use it (Pew, Feb 2026) | Share of AI users who use it (Menlo, Jul 2026) |
|---|---|---|
| ChatGPT | 44% | 60% |
| Gemini | 24% | 58% |
| Copilot | 17% | not reported |
| Meta AI | 14% | not reported |
| Grok | 8% | not reported |
| Claude | 6% | 20% |
| Character.AI | 3% | not reported |
Reading those two columns as a contradiction is the common mistake. Pew measures every US adult, Menlo measures only people already using AI, and the surveys are five months apart. Put together they say something neither says alone: Gemini’s reach is real, but it is concentrated among people AI already had.
Depth of use still separates them, though. Andreessen Horowitz, ranking by SimilarWeb visits and Sensor Tower monthly actives as of January 2026, puts ChatGPT at 2.7 times Gemini’s size on web and 2.5 times on mobile. Its data also shows sessions per user per month running 1.3 times higher on ChatGPT than Gemini on web, and 2.2 times higher on mobile. Roughly 20% of weekly ChatGPT web users also used Gemini in a given week.
The App Store measures applause, not attendance
Star ratings are the metric most people reach for, so it is worth showing what they actually contain. We queried Apple’s public lookup endpoint for the US store on 16 September 2026 and took the figures below straight from the response. The last column is ours: total ratings divided by years since first release, which is a crude pace measure and nothing more.
| App | Publisher | Rating | Ratings | First released | Ratings per year (our calculation) |
|---|---|---|---|---|---|
| ChatGPT | OpenAI OpCo | 4.83 | 10,332,633 | 18 May 2023 | 3.1m |
| Google Gemini | 4.72 | 2,236,007 | 13 Nov 2024 | 1.2m | |
| Grok AI | X Corp. | 4.88 | 1,417,362 | 8 Jan 2025 | 840,000 |
| Microsoft Copilot | Microsoft | 4.70 | 1,321,576 | 14 Jun 2013 | 100,000 |
| Character.AI | Character Technologies | 4.28 | 557,266 | 23 May 2023 | 168,000 |
| Perplexity | Perplexity AI | 4.82 | 505,673 | 27 Mar 2023 | 146,000 |
| Claude | Anthropic | 4.72 | 263,229 | 1 May 2024 | 111,000 |
| Meta AI | Meta Platforms | 4.68 | 249,341 | 9 Sep 2021 | 50,000 |
| Replika | Luka | 4.44 | 227,709 | 13 Mar 2017 | 24,000 |
Two things in that table should stop you. Grok carries the highest rating of the nine at 4.88, above ChatGPT, which no usage dataset supports as a ranking of stickiness. And Microsoft Copilot’s listing first shipped on 14 June 2013, nine years before ChatGPT existed, so its 1.3 million ratings have been accumulating on a listing that predates the whole category.
The reason is structural. A rating records one moment, usually early, usually when the app asked. But it never records the uninstall, because nobody rates an app on the way out. That is why every serious retention number in this piece comes from panel data rather than a store listing, and it is the same reason we priced tools against finished work rather than feature lists in our look at AI tools for business.
Uninstalls move fast, and March 2026 showed how fast
Sensor Tower’s State of AI 2026 report, published in June, contains the clearest churn event anyone has put on the record. In the United States, it reports, ChatGPT uninstalls peaked at roughly 200% above the app’s average during the week of 9 to 15 March 2026, after OpenAI’s agreement with the Department of War. Many of those users appeared to switch to Claude, the report says. Earlier that month, from 1 to 5 March, Claude recorded more daily downloads than ChatGPT.
ChatGPT led again every day after that, though, and the same report calls it the fastest mobile app ever to reach a billion monthly active users, in May 2026. Even so, its True Audience share across mobile and web fell below 50% for the first time in March 2026. Claude’s True Audience was up 452% year on year in May, and its US share rose from 4.4% to nearly 14% over the same stretch.
The catch is that a spike is not a trend. Sensor Tower also reports that ChatGPT, Gemini and DeepSeek together held nearly 90% of all time spent in AI assistant apps in the first quarter of 2026. Concentration that heavy means a 200% uninstall week barely dents the ranking, which is exactly why download charts are a poor guide to what survives.
Mixpanel’s State of Digital Analytics 2026, drawn from 3.7 trillion tracked events across more than 12,000 companies, adds the unglamorous version of the same point. In its AI section, stickiness in North America fell from 23% in 2024 to 21% in 2025, while Latin America climbed from 31% to 37%. Engagement in the category is not rising everywhere, and where it is rising it is not rising in the market that built the products.
The specialists people keep sit alongside the assistant, not against it
The assumption for two years was that general assistants would absorb the single-purpose tools. Menlo’s data says the opposite happened. Use of specialized AI products rose from 69% to 75% of AI users, and the share using both a general assistant and a specialist jumped from 60% to 72%. When switching is free, so is adding.
Consumers adopt what’s useful, and they keep what they trust.
Menlo Ventures, 2026: The State of Consumer AI
What kills a specialist is not obscurity. Among AI users who stick to general assistants only, just 11% say they are unaware of specialized products that would suit them. The larger blocks are privacy worries about niche tools at 27%, and not knowing which tool to trust at 21%. A specialist can be measurably better at the job and still lose to the app already open.
The survivors follow a rule you can see in the a16z lists. Where the big labs shipped a good-enough native feature, standalone traffic compressed: Midjourney ranked in the top 10 on the first edition of that list and sits at number 46 now. But where they did not, specialists held. Suno kept its number 15 slot and ElevenLabs has appeared on every edition since September 2023. Coding tools are the clearest case of a specialist owning a workflow, which is why we tested them on one shared repository in our coding model comparison.
Sora is the cautionary entry. It spent 20 days at the top of the US App Store and hit a million downloads faster than ChatGPT did, then downloads fell and it missed the mobile list entirely. Sensor Tower still counts upwards of 3 million daily active Sora users, which is the honest shape of a viral launch: a large residue, and nothing like the chart position implied.
What keeping one of these actually costs
Retention and price are the same conversation once the free tier runs out. These are the list prices we read on 16 September 2026, and the currency differs by vendor because these pages geolocate. Google served us pounds, so that row is not comparable with the rest; Anthropic and Apple served us dollars.
| Plan | Price | Read on |
|---|---|---|
| ChatGPT Go | $8.00 | App Store in-app purchases, US |
| ChatGPT Plus | $19.99 | App Store in-app purchases, US |
| ChatGPT Pro 5x | $100.00 | App Store in-app purchases, US |
| Claude Pro | $17 per month on annual, $200 up front, or $20 monthly | anthropic.com/pricing |
| Claude Max | From $100 per month | anthropic.com/pricing |
| Google AI Plus | £7.99 per month | one.google.com, UK |
| Google AI Pro | £18.99 per month | one.google.com, UK |
Against those tiers, Menlo reports the typical payer spends $20 to $49 a month, which lands squarely on one mid-tier subscription. Half of payers are on flat monthly plans and 26% hold annual ones, while 15% mix a subscription with usage charges. Who pays changes the number too: users whose employer or school covers it spend a blended $95 a month, against $55 for those paying out of pocket.
The upshot for anyone picking an app is unromantic. One paid assistant plus one free specialist is the median configuration the data describes, and the people who exceed it are mostly running agents. Among Menlo’s respondents, 92% of agent users pay for AI against 55% of AI users overall, and 63% of them use AI daily. If you are not running agents, the second subscription is usually the one that gets cancelled.
What none of this data can tell you
No company in this piece publishes a day-30 retention curve for its own app, so nobody outside those companies can rank these products on retention directly. Everything above is a proxy: panel estimates from SimilarWeb and Sensor Tower, self-reported survey answers from Pew and Menlo, and store metadata from Apple. Each has a different failure mode, and they do not cancel out, because they lean on overlapping panels.
a16z says so itself, at the end of the same report that produced its ranking.
The implication for this list: our rankings increasingly undercount the AI products people use most.
Olivia Moore, The Top 100 Gen AI Consumer Apps, 6th Edition, a16z
That undercount is not hypothetical. Claude Code reached a $1 billion annualised revenue run rate in six months without being a website or a phone app, and OpenAI reported 2 million weekly Codex users as of early March 2026. Neither of those registers in a web-visit or mobile-MAU ranking, which is the whole problem. The same goes for desktop notetakers, which a16z puts at a combined 20 million visitors across the top five.
Menlo’s own footnotes are careful in a way the headlines are not. Its 2025 wave ran in late spring and the 2026 wave in early July, which it says dampens measured growth on school-linked and work-linked activities. It also allows that some of the 36% counted as non-users may be hitting AI inside Google or Gmail without recognising it. The NBER working paper on ChatGPT usage that Menlo cites for cohort depth runs only to July 2025 and has two OpenAI employees among its authors, which the paper discloses.
What would change the picture is a single number none of these sources has: a published cohort curve, per app, showing how many people who installed in January were still active in March. Until one exists, the most defensible reading is the boring one. One assistant survives per person, a specialist survives when it owns a workflow the assistant cannot, and everything else is a download that has not been deleted yet.
Get the daily rundown
One email each weekday with the AI news that matters, every claim linked to its primary source.
Free, one email each weekday, unsubscribe in one click. We never sell or share your address.
