How 1,824 people build apps with AI

Fabricate AI had 1,824 builders in the 90 days from June 23 to September 20, 2026. A builder is a signed-in person who ran at least one build in that window. We know the device for 1,458 of them, because their own browser sent analytics data. Among those, 53.8% built on a desktop computer, 44.1% on a phone and 2.1% on a tablet, so 46.2% built on a phone or tablet.

The share has grown. Counting builders by the month in which they built, it was 35.2% in July, 51.6% in August and 49.8% in September 1 to 20. Since August, about half of the builders we can classify have used a phone or tablet.

A cross-check that uses only the operating system seen on the build pages gives 45.9% on iOS or Android, iPads included, among 1,451 builders with a known system. That matches.

What it means in practice: for anyone who makes software for people building with AI, phones and tablets are close to half of this audience, not an edge case.

Builders came from 120 countries. Among the 1,458 whose location we know (June 23 to September 20, 2026), 22.5% were in the United States and the rest, 78% rounded to the nearest whole number, were elsewhere. Brazil (8.4%) and the United Kingdom (5.2%) came next. The other 104 countries each had fewer than 20 builders, and together they account for 465 builders (31.9%).

By continent, the split was Europe 27.8%, North America 26.4%, Asia 19.8%, Africa 11.5%, South America 11.2% and Oceania 3.3%.

Location comes from each builder's own browser. Our server-side events carry our hosting provider's network address rather than the builder's, which reads as the US for everyone, so they were not used here.

Among the 1,458 builders whose browser reported a language (June 23 to September 20, 2026), 58.0% had it set to English and 42.0% to another language. Portuguese (8.8%), French (6.0%), Spanish (5.2%) and Arabic (5.1%) came next. A cross-check that uses only the build pages gives 42.1% non-English among 1,451 builders.

Browser language is not necessarily the language people write their prompts in. We have too little data on the language of the prompts themselves to report it, so read this as a description of the audience, not of the prompts.

The limits from the country section apply here too: builders who sent no browser data are not counted, and people in the 48 restricted countries are largely absent.

For anyone who makes tools for this audience: 42.0% of these builders run their browser in another language, so English is not the only language worth testing in.

For each builder we looked at where their first recorded visit to Fabricate AI came from, and grouped builders by the month of their first build. ChatGPT went from 12.6% of builders with a recorded first touch in July to 33.6% in August and 26.2% in September 1 to 20. Google search accounted for 30.3%, 28.1% and 27.0% over the same months.

Over the whole 90 days, 363 of the 1,445 builders with a recorded first touch (25.1%) came from ChatGPT. The 379 builders with no recorded first touch are left out. Other AI assistants barely register: Gemini, Copilot, Perplexity, Claude, DeepSeek and Grok together sent fewer than 20 builders in 90 days, so they are not reported separately.

These are not our ads. Every visit tagged as coming from ChatGPT had no campaign or medium tag, while our own paid ads on OpenAI, which began on September 14, carry campaign tags.

In our case, ChatGPT has mattered about as much as Google search as a way for new builders to find us since August. Whether that holds for other products is a question this data cannot answer.

When ChatGPT cites a link, it adds utm_source=chatgpt.com to it. Many of those clicks arrive with no referrer, so a report that groups visitors by referring site files them as direct traffic. Of the 363 builders we attribute to ChatGPT over the 90 days, 231 arrived with no referrer and only the tag.

Counting only the chatgpt.com referrer finds 129 builders, 8.9% of the 1,445 with a recorded first touch. Counting the tag as well finds 363 (25.1%). Referrer-only attribution undercounted ChatGPT about 2.8-fold.

If you track where your users come from, count a visit as ChatGPT when either signal is present: a chatgpt.com referrer, or a utm_source that contains chatgpt. Check first that the tag does not come from your own campaigns. In our data, none of the tagged visits did.

From August 1 to September 20, 2026, 66.2% of the 314 builders whose first visit came from ChatGPT built on a phone or tablet. That compares with 50.0% of the 284 from Google search and 32.3% of the 279 who arrived directly.

The pattern fits people tapping a link that ChatGPT cited while using it on a phone. That is our interpretation, not a measurement: the data records the device and the source, not the reason.

Using the operating system (iOS or Android) instead of the device type gives the same shares for ChatGPT, Google search and direct visitors. Both come from the same browser identification string, so this checks the classification rather than measuring it independently.

What it means in practice: if AI assistants send you visitors, look at your first screen on a phone before you look at it on a desktop.

We measured the length of the prompt that created each app without reading the prompts: the length is recorded as a number. Across 566 apps created from a typed prompt by 476 people (September 5 to 20, 2026), the median first prompt was 1,330 characters. The median hides two groups: 29.5% were under 200 characters, a line or two, and 27.2% were 5,000 characters or more, long enough to be a pasted specification.

After the first build, messages get short. Prompts typed into the build chat afterwards had a median of 73 characters (2,692 messages from 876 people over the 90 days, from browsers that allow analytics).

A second measurement over the same dates, taken in the browser when a signed-in person submits a first prompt (487 submits from 318 people), gives a median of 1,500 characters, 27.3% under 200 and 29.8% at 5,000 or more. That matches.

If you are about to write a first prompt: there is no length you have to hit. A one-line idea and a full specification are both common ways to start. This report does not compare how apps turned out by prompt length.

For anyone who builds AI tools, the spread matters more than the median: the same product has to fill the gaps in a one-line idea and also work through a document thousands of characters long.

Fabricate AI's build agent loads a specialised guide when a request needs one. One guide covers user accounts, login and sessions; another covers keeping data on the server. The agent's own records of which guides it loaded cover 98.8% of first builds in this window, so we counted them.

Across 360 first builds by 312 people, for apps first built from September 12 to 21, 2026, the agent loaded the accounts and login guide for 100 apps (27.8%, 96 people) and the server-side data storage guide for 192 apps (53.3%, 176 people). It loaded a guide for dashboard-style software-as-a-service interfaces for 51 apps (14.2%, 50 people). Fewer than 20 people got the marketing landing-page guide, so that figure is suppressed.

These are the agent's judgements of what each request needed, not what people asked for. Fabricate AI's build rules add accounts and login to any app that keeps records belonging to individual users, even when the person never mentioned login. That is why we say these apps needed login, not that people asked for it.

The storage guide also covers deployment conventions, so 53.3% is an upper bound for apps that need a database, not an exact count. It was loaded for 96% of the apps that needed login and for 36% of those that did not, so it follows the need to keep data rather than being a side effect of login. Payments and file uploads cannot be measured this way, because the agent has no separate guide for either.

A cross-check on login: when the accounts guide loads, Fabricate AI adds sign-in files to the app. From September 12 to 20, 95 of the 96 first builds that loaded the guide went on to work on those files, and no build worked on them without it.

If you are planning an app that keeps records for different people, expect it to need accounts and login. 27.8% of new apps did here.

For apps first built in September 2026, the median time from the start of the first build to the first preview on the builder's screen was about 2.5 minutes. Two ways of measuring it agree within about 10 seconds.

By the event's own timer, the median was 141 seconds and the 90th percentile 246 seconds (25th percentile 114 seconds, 75th percentile 193 seconds). That timer only exists from September 5, so it covers 457 apps from 416 people first built September 5 to 20. By timestamps, from the first build starting to the first preview being sent, the median was 152 seconds and the 90th percentile 262 seconds, across 663 apps first built September 1 to 20 that showed a preview.

87.0% of first builds (663 of 762) put a preview on screen at all.

The clock starts when the first build starts. The app is created a moment before that, right after the prompt is accepted: a median of 1.5 seconds earlier (90th percentile 7.6 seconds). Time spent signing up, for visitors who were not yet signed in, is not included. The clock stops when the preview is delivered to the builder's screen.

If you are building: expect to see something within minutes, and judge the result by whether it works, not by whether a preview appeared.

A preview can appear before the app works, so we also measured the time to a first verified working version. That means a build that finished on its own, deployed, and passed Fabricate AI's automated check, which opens the deployed preview in a real browser and confirms that it renders. It is not a test of every feature.

We followed 485 apps first built from September 1 to 13, 2026 (408 people), each for at least seven days, until September 20. 64.7% reached a verified working version at some point: 60.8% within 10 minutes of their first build starting and 62.9% within an hour. For the apps that got there, the median time was 204 seconds (3.4 minutes) and the 90th percentile 382 seconds (6.4 minutes).

The other 35%, about a third, never reached one in the time we observed. Most apps that were not working after 10 minutes never got there: the gap between 60.8% at 10 minutes and 64.7% overall is small. The data shows where an app stopped, not why.

What to expect: when a verified working version arrives, it usually arrives within minutes.

Why people stop: event data shows where a build or a person stopped, not why.

What kind of app people ask for: no event classifies first prompts and we did not read them, so we cannot say how many were shops, booking tools or games. The guide counts above are the closest measure.

Payments and file uploads: the build agent has no separate guide for either, and no event records the need.

Builders with missing browser data: for the 366 builders (20.1%) who sent no browser data, device, country and language are unknown, and 379 builders have no recorded first touch.

Other AI assistants: together they sent fewer than 20 builders in 90 days.

The whole market: every number here describes one product's users. Other AI app builders have different audiences and will see different numbers.

Every number in this report comes from Fabricate AI's own product analytics. Build events, such as a build starting, finishing or showing a preview, are recorded by our servers. Device, country, language and first touch come from builders' own browsers, and only when they allow analytics.