Own data is the moat: why AI wrappers lost 30–65% while data products hold

Revenue charts of small AI products tell a consistent story. Tools that put a new screen on someone else's model are shrinking, and the ones that sit on data nobody else has are not. Here is what the numbers show, why it happens, and where a data asset usually starts.

Thin AI wrappers, products that put an interface on someone else's model, lost 30% to 65% of monthly revenue from their late-2025 or early-2026 peak, according to verified TrustMRR data as of 16 November 2026. Small products built on their own data charge far more per customer and kept growing.

All figures below were collected on 16 November 2026. Revenue marked "TrustMRR" is checked against the product's payment provider. Everything else is what founders or sellers say about themselves, and we label it that way.

What counts as an AI wrapper?

A wrapper is a product whose core is a call to a model it does not own: a prompt, a nicer screen, maybe a template library. It writes blog posts, turns a script into a short video, drafts an essay or a mobile app mockup. The model does the work. The product handles the input and the output.

That is not automatically a bad business. Wrappers are fast to build and some earned well into five figures a month. The problem is what protects them once the model provider, a search engine or a dozen clones offer the same thing.

How far did wrappers fall from their peak?

TrustMRR publishes a monthly revenue chart for each listed product. Here is the drop from peak to October 2026 for several of them.

ProductWhat it doesPeak monthly revenueOctober 2026Change
SEOBOTBlog articles for SEO$98.9k (Dec 2025)$34.7k−65%
fastwrite.ioAcademic writing in Word$24.8k (Mar 2026)$12.0k−52%
StoryShortFaceless short videos$44.5k (Jan 2026)$22.2k−50%
SleekApp design from a chat$32.4k (Apr 2026)$16.7k−48%
ChatEDUStudy help$13.2k (Dec 2025)$7.0k−47%
Czat.aiPolish-language chatbot$14.8k (Mar 2026)$9.7k−35%

Sources: product pages on TrustMRR (SEOBOT, fastwrite, StoryShort, Sleek, ChatEDU, Czat.ai). Two more products in the same sample, tools for students and job seekers, fell 30% and 39% over the same period.

An audience does not stop the slide. SEOBOT's founder has about 116,000 followers on X, and monthly revenue still went from roughly $120k in May 2025 to $31.7k in November 2026 (Latka, TrustMRR). When the output becomes something anyone can get for free, followers do not change the price.

Language does not protect a wrapper either. Czat.ai is a Polish chat assistant and fastwrite.io serves German students. Large models now write Polish and German well, and both products are down by a third or more.

Why are AI wrappers losing revenue?

The causes differ by product, but four show up again and again.

Google answers the question itself. Google shows AI Overviews on about 20% of informational queries, and deals for content sites on Flippa fell 33.5%, according to Flippa's 2025 review. A tool that sells generic articles competes with an answer that sits above every article. Textero, a student writing assistant, had around 10,000 search visits a day before a Google update cost it rankings; it later merged with a competitor (Starter Story).

The platform controls access to its data. GummySearch, an audience research tool built on Reddit, shut down after it could not agree a commercial licence for Reddit's data (GummySearch). If your product reads someone else's platform, that platform decides whether you exist.

The model provider ships your feature. The founder of Superpower ChatGPT, a browser extension for ChatGPT, has said he has gone numb to the idea that OpenAI could copy his features (Rosetta, Starter Story recap). Every product that adds folders, search or export to a chat interface lives with this.

API bills eat the margin. CodeGuide reported an OpenAI bill of about $2.8k out of roughly $3.5k in monthly costs (Starter Story). When most of your cost is someone else's price list, they also set your margin. In dictation, free open-source apps and lifetime licences at €9.99 have already pulled prices down (TechCrunch).

How do buyers value revenue that is shrinking?

People who buy small online businesses price this risk directly. A multiple is how many years of profit or revenue a buyer pays for a business. On Flippa in 2025 the average multiple for software was 2.7 times annual profit. The median for confirmed deals on Acquire.com was 3.9 times (Flippa, Acquire).

Two listings show what happens to a wrapper:

  • An AI content generator for small-business social media fell from $43.5k to $22.7k in monthly recurring revenue and is listed at about 0.8 times annual revenue (Flippa).
  • Twixify, a text rewriter, dropped from $43k to around $10k and asks about 0.2 times annual revenue, with a six-month buyback guarantee (Flippa).

Flippa says the premium goes to "AI apps with retention", not to the label itself. A review of 2026 exits found no premium for the word "AI" on its own (musthave.ai).

A drop is not the end, though. Twixify's seller says it held about $10k a month for a year and a half with no development and no ads. Wrappers can be decent cash businesses. They are just valued as if the cash is temporary.

What do products built on their own data look like?

The products in the same research that hold up best share one thing: the valuable part is a data set they collected, cleaned and keep updating. The model is just a tool on top of it.

LeadX combines company data, UCC lien filings (public records of loans secured against business assets) and web signals for business lenders and B2B sales teams. On TrustMRR it shows $31.6k in monthly recurring revenue from just 18 subscriptions, an average of about $1,756 per subscription. That is the highest figure among 30 AI products in the sample (TrustMRR). Its chart went from $1.6k in December 2025 to $40.7k in October 2026. The raw filings are public. The value is in pulling them together, matching them to companies and keeping them current.

Podscan transcribes podcasts at scale and sells search and monitoring on top of that archive. Founder Arvid Kahl wrote that it became profitable in April 2025 (Bootstrapped Founder). Getting there meant cutting monthly costs from about $30k to under $10k (Huxiu interview). The archive is hard to copy because it took months of compute to build.

Virlo collects data on short videos across TikTok, Reels and Shorts and sells it to agencies, while other startups use it through an API. Its founders say it makes more than $31k a month and credit their own data infrastructure as the advantage (Indie Hackers). That figure is self-reported.

What does a data moat cost?

It is not free, and the trade-off is worth stating plainly.

  • Sales take work. None of these three spreads by itself. LeadX shows no visible self-serve channel and likely sells directly. Virlo names search, paid social and direct outreach to agencies. Expensive B2B data is sold through demos and conversations, not app store rankings.
  • Infrastructure is heavy. Podscan's cost story is the clearest example: a data product pays to collect and store data every month, whether customers arrive or not.
  • Growth is slower at the start. A wrapper can be built in a weekend. A data set that is worth paying for takes months of collecting before anyone pays.

The pattern across the research is blunt: products that spread on their own are easy to copy, and products that are hard to copy have to be sold by hand.

Can niche expertise replace your own data?

Partly. Narrow tools for professionals grew fastest in 2026, and their protection is knowledge of the craft rather than data.

Stencil AI turns photos into tattoo stencils for working artists. According to the aggregator SaaSHub, one of its two founders is a tattoo artist. Its verified revenue grew from $16.1k in January to $62.5k in July 2026 (TrustMRR, SaaSHub).

But craft knowledge leaks into the product, and the product can be copied. Within months, the App Store already had at least three similar apps: StencilX, Stencify and Stencil Maker IQ. Expertise buys time. Data you keep adding to buys more.

Where does your own data set usually start?

Very few data products begin with data nobody has ever seen. LeadX starts from public filings. Podscan starts from public podcasts. Virlo starts from public video statistics. Their work is everything that happens after collecting the data. A practical sequence looks like this:

  1. Pick a narrow question that a business pays to answer. "Which companies just took on equipment loans" is a product. "Business data" is not.
  2. Start from open data. Government portals, research repositories and model hubs already publish a lot of raw material. You can search open datasets on DataHunter across Hugging Face, Zenodo and public data portals. Every result shows the license, format, size and last update, and you can filter to datasets that allow commercial use.
  3. Check the license before you build. A data set you cannot use commercially is a prototype, not a foundation. Write down where each source came from and under which terms.
  4. Add your own layer on top. This is the part competitors cannot download: labels you assigned, records you matched across sources, categories you defined, corrections from customers. It turns a public file into your data.
  5. Keep it fresh on a schedule. A one-time scrape ages fast. A data set that updates weekly, with a history going back months, gets harder to replace every week.
  6. Put a model on top last. At this point the model is replaceable, and that is the point. If a cheaper or better model appears, you switch and keep everything else.

How reliable are these numbers?

Only revenue shown on TrustMRR is checked against a payment provider. Descriptions, team size and channels on those pages are written by founders. Figures from Indie Hackers, Starter Story and similar interviews come from founders themselves and are best read as an upper bound. Flippa listings are seller-provided, with some checks by the marketplace. And every story here is a survivor: on TrustMRR, 67.4% of all listed projects make under $1k a month (TrustMRR stats), and failures rarely get written up. Treat the pattern as a signal, not a promise.

What to do next

Other guides

All blog posts