Software Growth

Moat

A moat is what stops a competitor, or your own customer with an AI coding agent, from replacing your product. In 2026, SaaS moats come from trust, workflow, permissions, data and operations, not from code that is hard to write.

A moat is whatever keeps competitors from taking your customers and your margins. The term comes from investor Warren Buffett's "economic moat", a picture of a castle protected by water. In software, the question every founder gets asked, "what stops someone from building this?", is really a question about your moat.

In 2026 that question has a new edge. AI coding agents mean a competitor, or one of your own customers, can rebuild much of a typical SaaS product quickly. Jon Yongfook, founder of Bannerbear, puts it bluntly: competing on features is dead, because features are now cheap to build. This guide covers the moats that still hold up, drawn from three recent sources: Rob Walling and Einar Vollset's episode on the eight SaaS moats that still work, Pieter Levels' list of the moats left, and Yongfook's essay on how SaaS companies can compete or pivot.

How AI attacks a SaaS business

In episode 855 of Startups for the Rest of Us (October 20, 2026), Walling and Vollset list ten ways AI threatens a SaaS company. Some add competition: customers vibe code the app, competitors catch up on features, open source gets good enough, platforms ship your feature for free, custom software gets sold at SaaS-like prices, and the models themselves absorb your product. Others shrink demand: your customers stop existing, per-seat revenue disappears, the roles you serve vanish, or your step in the workflow becomes obsolete.

Moats that stopped working

Several things founders used to rely on now protect much less:

  • Feature parity plus one. Yongfook notes that a competitor's "10x more features" claim can be invalidated by a motivated developer in days.
  • Code that took a long time to write. Walling and Vollset call this out directly (from 24:09): difficulty of building is no longer a barrier on its own.
  • UX familiarity. "Our staff already know this tool" matters less when agents do more of the clicking.

Once there’s a manual, anyone can do the work, and anyone will.

Seth Godin / Be careful what you wish for (October 28, 2026)

Godin was writing about professional services, but the point applies to software: anything that can be fully written down can now be copied.

The moats that still work

None of these are on/off. Walling stresses that each sits on a spectrum, and most businesses have a little of several. For each, there is a question you can ask about your own product.

Pieter Levels' list is a good summary before the longer version:

Proprietary data or content [...] Network effects / community [...] Audience / distribution [...] Physical component [...] Capital [...] Regulatory

Levels explains each one: data or content only you have, users who add value for one another, the ability to get attention whenever you need it, hardware and real-world operations, owning assets and investing in AI companies, and licenses or legal barriers.

Pieter Levels / Post on X (October 12, 2026)

1. Buyers know and trust you

Buyers know and trust you. Your product sits at the center of recommendations from partners and advisors, your audience, a champion inside the customer, and your brand. A new clone has none of these connections.
Adapted from Rob Walling and Einar Vollset, Startups for the Rest of Us, episode 855, and Pieter Levels (October 12, 2026). Schematic.

This is more than a brand. It is knowing your buyer and how they buy, partners whose business is recommending you, and a champion inside each account. Walling's examples are QuickBooks, which accountants recommend, and Shopify, whose agency network defaults to it. Levels lists the same idea as audience and distribution: "you can get attention on demand". Yongfook agrees that when two products have the same features, the one with the audience wins.

One caution: Lenny Rachitsky and Dan Hockenmaier argue that distribution and brand are fuel rather than compounding advantages. They get a company moving, but they need to feed one of the deeper moats below.

Ask: who has built their business or career around recommending you?

2. You're the system of record

You're the system of record. The canonical state of the business lives in your app and other systems such as payments, door access, email and accounting are wired to it. Leaving means re-entering cards, reprogramming key fobs, rewiring integrations and retraining staff.
Adapted from Startups for the Rest of Us, episode 855 (Gymdesk example). Schematic.

Your software holds the definitive record of how the business runs, and other systems integrate with it. Vollset's example is Gymdesk: who teaches the 6 AM class on Tuesday, who has paid this month and who has a key fob all live there and nowhere else. Exporting the data is the easy part. Re-entering every member's card, reprogramming the fobs and rewiring integrations is not. Hockenmaier calls this the "workflow tax": the time a team spent building a product into how it works.

Ask: is there canonical state that is recorded only in your system?

3. You're allowed to act

You're allowed to act. An advisor tool produces a recommendation that a human must review and act on. An actor tool produces the result directly, for example by moving money.
Adapted from Startups for the Rest of Us, episode 855 (Stripe Connect example). Schematic.

A tool that recommends a payment is easy to replace. A tool that is authorized to make the payment is plumbing. Stripe Connect moves money between parties. Once software is wired in and acting on a customer's behalf, removing it is a project nobody wants.

Can your software do something in the world, or is it just an advisor?

Einar Vollset / Startups for the Rest of Us, episode 855 (October 20, 2026), from 30:10

4. Customers don't check your work

Customers don't check your work. Payroll, taxes and filings flow straight to done without a manual check. A vibe-coded replacement would mean checking everything again.
Adapted from Startups for the Rest of Us, episode 855 (Gusto example). Schematic.

This builds on the previous moat. Vollset uses Gusto, which files payroll taxes for him with several states. He doesn't check those filings every month. Trust that deep is hard to transfer: a cheaper or home-built replacement means taking back the risk and going back to checking everything.

Ask: do your customers still review your output, or have they stopped?

5. Someone official already said yes

Someone official already said yes. A clone can be built quickly but then waits at the approval, certification or accreditation gate, which your product has already passed.
Adapted from Startups for the Rest of Us, episode 855 (SignWell example), and Pieter Levels (October 12, 2026). Schematic.

Regulation, licenses, accreditations and certifications take months to get, even when the code takes days. Vollset's example is SignWell, an e-signature tool: placing a signature on a PDF is easy, but nobody wants a million-dollar contract questioned because the signing tool was vibe coded. Levels lists "licenses and legal barriers" as a moat too, while admitting in a follow-up post that it's the one he won't pursue.

Walling makes a useful distinction: serving regulated customers is not the same as holding the approval yourself.

Jon Yongfook offers a quick test for how exposed a product is: does it depend on people taking part?

Create software that depends on human participation. [...] safe for the next few years at least.

Yongfook's example is document signing: he argues AI won't eat the category and serious companies won't vibe code it.

Jon Yongfook / Post on X (October 27, 2026)

In the same thread he adds that anything only computer to computer, such as an API, gets absorbed eventually. Treat this as buying time, not as permanent protection. One reply pointed out that an agent holding power of attorney could sign for you, which shrinks the human step to a delegation setting. Yongfook agreed that will happen eventually, just not for a few years, because lawyers will fight it.

6. Everyone's already in it

Everyone's already in it. The effort to switch grows from one person, to a team, to a company, to a whole community.
Adapted from Startups for the Rest of Us, episode 855 (Slack and school apps), and Pieter Levels (October 12, 2026). Schematic.

These are network effects. One person can switch tools in an afternoon. A team has to agree and move together. A company or a community is a migration. Slack keeps customers who complain about it, and school parent portals survive despite poor reviews, because moving thousands of parents is worse. Levels describes the same moat as community: "users make it more valuable for each other". He made the nomads.com community nearly free to grow it.

Ask: how many people would have to agree before a customer could leave?

7. Every customer makes you smarter

Every customer makes you smarter. Customers use the product, real-world outcomes are observed, the product learns, and it gets better for every customer. A fresh clone starts at zero.
Adapted from Startups for the Rest of Us, episode 855 (Stripe Radar example), Pieter Levels, and Lenny Rachitsky and Dan Hockenmaier (October 2026). Schematic.

No single merchant could ever build Stripe Radar.

Rob Walling / Startups for the Rest of Us, episode 855 (October 20, 2026), from 38:03

When usage from every customer improves the product for all of them, a competitor starting from scratch can't catch up by writing code. Walling sets three tests: it helps every customer, customers can take their data but not what you learned from it, and it keeps learning. Vollset adds that the learning should come from real-world outcomes, not something a rival can simulate with synthetic data. The episode calls this learning loop the strongest moat of all (from 54:46).

Levels' version is proprietary data: "data nobody else has, or unique content you can create". He has since started selling a paid nomads.com city data API. Hockenmaier counts data as one of only four compounding advantages, alongside switching costs, network effects and scale.

8. It's hard to keep running

It's hard to keep running. The code is a small layer an agent can write. Below it sits the larger operational work: deliverability and IP reputation, proxies and CAPTCHAs, abuse and spam, being on call at 3 AM, and compliance.
Adapted from Startups for the Rest of Us, episode 855, and Jon Yongfook, Bannerbear (September 2026). Schematic.

Is it hard to build and code, or is it hard to keep running?

Rob Walling / Startups for the Rest of Us, episode 855 (October 20, 2026), from 40:47

Some products are easy to code and painful to operate: web scraping through proxies and CAPTCHAs, email deliverability, edge networks like Cloudflare. The work only makes sense when it's shared across many customers, which is why they pay. Walling recalls Levels saying he vibe codes almost everything but still pays for ScrapingBee. Yongfook makes the same case for "the unsexy things nobody wants to build", such as hosting, email delivery and payments, where the hard part was never the code.

9. Atoms, not just bits

Atoms, not just bits. The software layer of a product can be cloned in a weekend; the physical layer of manufacturing, supply chain, certifications, inventory and locations cannot.
Adapted from Pieter Levels (October 12 and 16, 2026), Jon Yongfook, Bannerbear (September 2026), and Startups for the Rest of Us, episode 836. Schematic.

Hardware, physical locations and real-world operations can't be copied over a weekend. Yongfook writes that a hardware pivot "has more of a moat than almost anything else" on his list, while warning that it needs capital and has thinner margins. Levels is acting on it with in-person nomads.com meetups and a retro PC hardware project. In episode 836, Vollset lists hardware-software coupling first among the AI-era moats acquirers value.

10. Go where the tech isn't

Go where the tech isn't. Tech-savvy markets are crowded with competitors. Trades, logistics, agriculture, local government and small manufacturing are harder to reach, which keeps competitors out too.
Adapted from Jon Yongfook, How to Pivot Your SaaS Company in the Age of AI (Bannerbear, September 2026). Schematic.

Trades, logistics, agriculture, local government and small manufacturing often still run on spreadsheets, phones or decades-old software. Yongfook points out that you can't launch on Product Hunt and expect plumbers to show up. These customers are hard to reach, and he argues that's the point: "But that's the moat." Compare vertical SaaS.

11. Become the all-in-one

Become the all-in-one. Eight separate tools, each with its own login and bill, are replaced by one product with one login, one bill, one set of permissions and one set of integrations.
Adapted from Jon Yongfook, Bannerbear (September 2026), and Lenny Rachitsky and Dan Hockenmaier (October 2026). Schematic.

Being the one thing that replaces a pile of things is a solid moat.

Jon Yongfook / How to Pivot Your SaaS Company in the Age of AI (September 2026)

Companies can now replace individual vendors with internal tools, but maintaining eight of them isn't practical. A single product that covers all eight offers one login, one bill and one set of permissions without the maintenance. Yongfook adds a condition: "as long as you actually do all of it well." Hockenmaier points to Rippling's "compound startup" approach, which spread across HR, payroll and finance on shared infrastructure.

Capital is a hedge, not a product moat

Levels also lists capital: owning assets and investing, especially in AI companies. In his October 16 post he describes adding AI stocks, bonds and land, and being open to selling two of his apps. That protects a founder's wealth. It doesn't stop anyone copying a product, so treat it as a personal strategy rather than a moat for your SaaS.

The same post goes through his whole list, moat by moat, with what he is changing in each of his products:

I am now scrambling to chase any remaining moats [...] chasing moats it is!

He then goes through each moat with specifics: a paid nomads.com data API, a paid hoodmaps.com API in progress, a nearly free community, an email and Telegram audience of 187,279, more meetups, hotelist.com, a retro PC hardware project, and more money in bonds, land and startups. On regulation, he says he has nothing.

Pieter Levels / Post on X (October 16, 2026)

Going on offense

The moats above are defensive. Walling and Vollset end episode 855 with three offensive moves:

  1. Keep improving the business with AI. They see this as table stakes.
  2. Serve agents. Agents will increasingly use your software. Vollset describes two stances: lock them out, or make it easy for them, for example with an MCP server you can gate and charge for. Yongfook describes doing this at Bannerbear: customers ask their coding agent to generate videos through the Bannerbear MCP server.
  3. Become the worker. Sell the outcome instead of the tool. Vollset argues this is the main opportunity, because a product that does the work and learns from the results builds the learning loop in moat 7.

Seth Godin reaches a similar list from another direction. In Building a second-wave AI business (October 6, 2026), he warns that selling cost savings is a race to the bottom and offers a "theory of profit" for AI projects that can become real businesses. In his framework, a project builds up its own specific, trusted data. It either has a community whose members see each other or gives people an obvious reason to spread the word. It solves a costly problem for buyers who have budgets, and it can start without data centers or a large team. He adds a check on that data that doubles as a warning about vendor lock-in:

It’s built to work for users, not to trap them.

Seth Godin / Building a second-wave AI business (October 6, 2026)

Score your own SaaS

Yongfook's advice is to pick two or three; you can't do everything. For each moat, rate yourself honestly from "none" to "strong":

  • Who recommends us for a living?
  • What state lives only in our system?
  • Do we act, or only advise?
  • Have customers stopped checking our output?
  • Do we hold an approval a clone would need?
  • How many people must agree before a customer can leave?
  • Does each customer make the product better for the others?
  • Is our product hard to code, or hard to keep running?
  • Is any part of what we deliver physical?
  • Are our customers hard for others to reach?
  • Could we replace several tools our customers pay for?

A simple stress test still applies: if a competitor copied your product tomorrow, what would still make customers stay?

The classic view

Older frameworks still hold. Jason Cohen's essay on moats separates a competitive advantage from a durable one, using a taxonomy by Jerry Neumann: scale, complementary assets, learning curves, brand and switching costs. He stresses that moats come from deliberate multi-year investment. Hamilton Helmer's 7 Powers calls the same idea "power", and NFX's Network Effects Manual covers the network moat in depth. The AI-era lists above mostly sharpen these: switching costs became "system of record" and "allowed to act", and learning curves became data and learning loops.

Common mistakes

  • Calling a head start or a feature list a moat.
  • Counting "we use AI" as a moat when the same model is available to everyone.
  • Building hostile vendor lock-in that angers customers into leaving the moment an alternative appears.
  • Assuming an audience is a moat on its own, without a product that keeps the customers it brings.

Sources

Back to the SaaS glossary