# What's defensible for small builders now that AI labs ship apps

> Anthropic and OpenAI now ship the apps their customers build. What a small builder can still defend: niche data, workflows, trust and distribution.

Author: [Flavio Copes](https://flaviocopes.com/about/) | Published: 2026-11-12 | Topics: [Business](https://flaviocopes.com/tags/business/), [AI](https://flaviocopes.com/tags/ai/) | Canonical: https://flaviocopes.com/defensible-small-builders/

On November 10, 2026, Anthropic [started rolling out](https://x.com/claudeai/status/2120206365759094857) Claude Whiteboards, Claude Sheets and Claude Data. The same day OpenAI launched [Businesses in ChatGPT](https://flaviocopes.com/businesses-in-chatgpt/), a dashboard where a company manages how ChatGPT presents it.

Klaas ([@forgebitz](https://x.com/forgebitz/status/2120223958066774417)) summed it up that evening:

> openai is going after google, anthropic is going after microsoft

If you build on these models, the company that sells you the model now also sells a product next to yours.

I build small apps and write a blog, and both depend on these companies. So I went through what I have that a lab can't ship next week, and where I'm exposed.

## The fear

The replies under that post say it well. Nadeem ([@gearsandcode_](https://x.com/gearsandcode_/status/2120365441851744594)) wrote that "openai wants to be your search bar, anthropic wants to be your org chart."

Enrique S ([@nrique_s](https://x.com/nrique_s/status/2120270180227953057)) added that the end game is the same for both: they climb one layer at a time and replace the incumbent at each layer. In an earlier post he listed the order he expects: freelancers first, then niche SaaS, then ERPs, then operating system companies.

A whiteboard, a spreadsheet and a dashboard are three products that many small companies sell today. Now they come with a Claude paid plan.

```mermaid
flowchart TD
  accTitle: Two paths from a model to a customer
  accDescr: A lab sells its model through an API to a small builder's product, which reaches the customer. The lab also ships its own app, which reaches the same customer directly.
  A[AI lab's model] --> B[API]
  B --> C[Your product]
  C --> D[Customer]
  A --> E[The lab's own app]
  E --> D
```

## What isn't defensible anymore

Some things stopped protecting a product:

- **Generic features.** A chat with your documents, a meeting summary, a chart from a CSV file. A lab can add any of these to an app people already have open. OpenAI says 1.2 billion people use ChatGPT each week.
- **A thin wrapper around a model's API.** If the product is a prompt and a text box, the lab's own app does the same job with the same model.
- **Being able to build it.** Coding agents write most of the code in my apps now. Anyone can do the same, so "it was hard to make" is no longer a reason to buy.

I wrote about that last point in [how to compete with AI-generated SaaS clones](https://flaviocopes.com/after-the-clone/). My [apps](https://flaviocopes.com/apps/) cost $20 each and every purchase includes the full source code, so I already can't charge for the difficulty of writing them.

Four things are left: niche data, workflows, trust and distribution.

## Niche data

**Niche data** is data the labs don't have and can't scrape. You collect it by using something, by curating it, or because you have access to a domain that they don't.

My example is the [hosting picker](https://flaviocopes.com/hosting-picker/). It holds the plans and prices of 21 hosting providers. Every price comes from the provider's own pricing page, and each host shows the date I last checked it.

![The hosting picker on flaviocopes.com with a shortlist of three hosts and the line that says prices were checked on October 1, 2026](https://flaviocopes.com/images/defensible-small-builders/hosting-picker.webp)

A model can answer "which host should I use?" too. But its prices come from whenever it was trained, or from whatever page a search returned. Mine have a date next to them, and the ranking rule is written on the page.

The [acquisitions](https://flaviocopes.com/acquisitions/) page and the [AI news](https://flaviocopes.com/ai-news/) page work the same way. They're lists I keep up to date, entry by entry, and the value is in the upkeep.

Be careful with what you call niche data. A list copied from public pages once is a snapshot that anyone can copy again. The data is worth something when you keep checking it. I went through the cost of that in [which business advantages can a solo developer build](https://flaviocopes.com/indie-business-moats/).

## Workflows

A general-purpose tool does the step that looks good in a demo. A **workflow** product does the whole job, including the boring steps.

[Video Producer](https://flaviocopes.com/apps/video-producer/) is how I publish videos. A coding agent edits the recording, checks every word of the transcript, exports in 4K, writes the title and description, and renders the cover. I do two checks and the upload. The app tracks every one of those phases.

[SEO Bot](https://flaviocopes.com/apps/seo-bot/) runs 16 checks on a site. Each report becomes a task list grouped by source file, with the exact text to find, and it hands that list to a coding agent working in the site's folder. It also keeps reminders, so weeks later I go back and check whether the fix worked.

A spreadsheet inside a chat app doesn't know which source file a page comes from, and it doesn't come back a month later to ask if the new titles got more clicks.

The messy steps are different in every job, which is why a company building for a billion users skips them.

## Trust

People buy from a person they know.

I've been writing tutorials on this site for years, and the [newsletter](https://flaviocopes.com/newsletter/) has 130,000 subscribers. When I release an app, many of the people who buy it have been reading me for a long time.

Trust also comes from what you refuse to do. Nobody pays to be in the hosting picker's ranking, and none of its links are affiliate links. I say that on the page, and the ranking rule is there for anyone to check.

A lab has a brand that everybody knows. It doesn't have a person you already know behind a $20 app.

## Distribution

**Distribution** is owning the way people find you: a newsletter, a blog, search rankings, an audience on X.

This is the one the launches of this week touch directly. Elizabeta Kuzevska ([@ekuzevska](https://x.com/ekuzevska/status/2120295948693786977)) replied to Klaas:

> If ChatGPT becomes where discovery happens, the shortlist forms there. The vendor only finds out when the buyer shows up already decided.

I saw this when I tested Businesses in ChatGPT. I asked for wool running socks and ChatGPT recommended five brands. Two of the brands it named got no link to their own website. The links went to retailers.

A newsletter is the part of distribution that no lab sits in the middle of. I have the email addresses, and an email reaches the reader without anyone ranking it first.

Search is different, as you'll see in my numbers below.

## Supplier risk

Klaas also [warned](https://x.com/forgebitz/status/2119723031920599133) about the other side of this, the day before:

> at some point, the ai labs will restrict api/model access

His reasoning is that if the labs keep going after the app layer, they can cripple the competition.

You can't prevent that. You can make it cheap to switch.

SEO Bot doesn't call a model's API. It hands its reports to the agent you already use, Cursor or Codex. [Searchglass](https://flaviocopes.com/apps/searchglass/), my Search Console CLI, ships a skill that installs for Codex, Claude Code or Cursor:

```bash
searchglass skill install --target codex
```

If one lab closes a door, the tool keeps working with the others.

The one model call in this site's own functions runs on Llama 3.2, an [open-weight model](https://flaviocopes.com/open-weight-models/), on Cloudflare Workers AI. I can move it to another host without asking anyone.

My advice is to keep the model behind one function in your code, and to test your product with a second model before you need to.

## My own setup

My business runs on search and on these models. Here is which of the four I have.

| | What I have | Where I'm exposed |
| --- | --- | --- |
| Niche data | The hosting picker (21 providers), the acquisitions and AI news pages | Prices and lists are public, so the value lasts only while I keep them current |
| Workflows | Apps built around my own jobs, like Video Producer and SEO Bot | They need a coding agent to do the work, and the labs make those agents |
| Trust | Years of tutorials under my own name, and apps that ship with their source | It's tied to me, so it doesn't grow without my time |
| Distribution | 130,000 newsletter subscribers, and Google | Most new readers arrive through a search box I don't own |

Distribution is where I'm most exposed.

Search traffic to this site grew a lot over the last year, and that looks healthy. But most of the growth comes from posts about AI coding tools, like [pstack](https://flaviocopes.com/pstack/), [T3 Code](https://flaviocopes.com/t3-code/) and [the model comparison](https://flaviocopes.com/ai-coding-models/).

So I'm writing about the products of the same companies, and the readers arrive through Google, which is one of them. If people start asking ChatGPT or Claude those questions, the clicks can drop without anything changing on my site.

The newsletter is my hedge, because it reaches people without a search box in between.

On the model side I'm less worried. The tools I work in let me pick models from different labs, so if one lab restricted access I'd change the model and keep working.

## Everything around the code

AI made the building part easy, for me and for everyone else, including the labs.

What's left is the data I keep current, the jobs I know well enough to automate end to end, the readers who know me, and the list of email addresses. Those took years to build. If I were starting today I'd begin with the email list, because it's the only one of the four that nobody else sits in front of.