Kyle Poyar published a study in Growth Unhinged last week that does something most competitive research doesn't. It states its own limits.
He combed through the websites of 50 top SaaS and AI-native companies to see how they handle the question every software buyer is now asking: can't we just build this with Claude? He looked at compare pages, blogs, docs, pricing, trust centers. He turned what he found into five lessons and a decision tree.
And he said plainly that he wasn't going to pretend to know who has the biggest moat against Claude, or how durable those moats would turn out to be. Positioning was what he set out to investigate, so positioning is what he investigated.
That's a scope limit, not a gap. But it's the more expensive question, and I said I'd take a run at it.
What follows extends his data rather than disputing it. His findings hold. The question is what they tell you about durability, and the answer turns out to be: almost nothing, by construction.
What the data actually shows
Three numbers carry the study. All figures below are Kyle's, gathered in July 2026 and published in Growth Unhinged on 12 August.
The gap between 34 and 7 is the study's real finding. Most companies have chosen to be available inside Claude and have declined to explain why they still exist. Kyle names the risk himself: ship enough surface into the model and Claude becomes the interface to your own product.
That is a strategic position, not just a marketing gap. And nobody in the 34 appears to have priced it.
The tree, and what kind of object it is
Kyle's decision tree gives you five branches. Does your product make Claude better. Is Anthropic your customer or partner. Are you in a vertical Claude can't serve. Are you selling to enterprise. Are you selling to SMB or midmarket.
Each one is a good answer to a real question. My claim is narrower than it sounds.
Every branch on the tree is a content decision.
Not one of them describes a structure. They describe things you can say, and the fact that they are true when you say them doesn't make them barriers.
Here's the tell, and it comes from his own exemplars rather than from any framework of mine. Glean appears in the head-to-head section and again at branch four. Lovable appears in the complementary-positioning section and again at branch one. Harvey sits at branch three and carries the fourth lesson. These companies are running multiple branches simultaneously.
Which is entirely fine. Qualifying for five branches costs you nothing, because qualification describes what you already are. You either make Claude better or you don't. Anthropic is either a customer or isn't.
Barriers don't work that way. Each one costs something, and each one forecloses another. Building regulatory depth costs you speed. Building a network effect costs you early-stage focus. Nobody qualifies for a moat.
Moats do stack. That's the whole premise I write about. But they stack the way floors stack, one paid for at a time, not the way descriptions stack.
Five branches, five questions
Kyle's graphic pairs each branch with two companies actually running it. That roster is the useful part, because it lets you ask a second question of each branch without leaving his data.
What follows is one question per branch, using his exemplars. These are questions, not verdicts. I've run full diagnostics on a few of these companies and not on others, and the point isn't to score them here. It's that each branch has a different structural question hiding behind it, and the tree doesn't ask any of them.
The sharpest pair on the board, because they sit on the same branch with completely different things underneath.
Zapier's argument is that Agent Builder handles a handful of native connections while their MCP reaches roughly nine thousand apps with triggers and scheduling. Every one of those connections is . That took years and it isn't purchasable.
Lovable's page leads with running on Claude. Persuasive to a buyer today, and also a description of a dependency.
Kyle flags the expiration date on this branch himself. The branch doesn't distinguish between a company whose complement position rests on accumulated counterparty agreements and one whose complement position rests on the platform's continued goodwill.
Kyle calls this the best proof point available, and for conversion he's right.
Clay is the instructive case, because Clay is genuinely hard to displace. Their position rests on and assembled over years. None of it comes from the logo.
So the same asset can be excellent proof and almost no barrier. Those are different properties and the tree can't tell them apart. Worth noting that Fin is being acquired by Salesforce for $3.6B, and that outcome didn't come from an Anthropic case study either.
For Clay, most of it. That's the answer you want, and it's an answer the branch never asks for.
Both compete on domain expertise. Underneath, they're doing different things. One is knowledge in a professional services vertical. The other is operational embedding in the trades, where the software runs dispatch, scheduling and invoicing for the business.
Domain knowledge is copyable given enough hiring. is not, because the cost of removal lands on the customer rather than the competitor.
This is the branch most likely to hold, and Kyle's caution about the roadmap is well placed. But the reason it holds is time, and time is the one input a funded competitor can't compress.
Glean's page is the most technically serious artifact in the study. Workflow evaluations across 175 queries, plus specifics on sensitive data handling, audit logs, compliance APIs and single-tenant deployment.
All real. All per-tenant.
and make leaving expensive. Neither makes staying more valuable each month, and single-tenant architecture guarantees it, by design and for good reason.
Time-to-value and cost efficiency are the first two things every model release erodes. A cost claim in this category has a shelf life measured in months, and everyone running this branch knows it.
This is the only branch where the honest answer is that the position buys you time rather than protection. That's not nothing. It's just a countdown, and the branch doesn't tell you to start one.
Notice what the five questions have in common. Not one of them is answerable from a website, which is why Kyle's method couldn't reach them and why he said so.
Notice also the ordering. The branch with the most structure underneath sits at position three, behind two that describe nothing durable at all. A decision tree resolves on first match, so a company that could truthfully claim branches one, two and three presents as branch one.
The structure routes you toward your loudest available claim rather than your most durable one.
I've called this the inverted visibility trap elsewhere. The claims that are easiest to make and most immediately persuasive tend to describe the least structure. The claims that describe real barriers are slow to explain, boring in a sales conversation, and frequently invisible to the buyer.
The best play on the list has no branch
Kyle's fourth lesson is the strongest thing in the study, and it never made the tree.
Fin publishes twenty questions telling prospects how to evaluate enterprise AI customer service agents: how to evaluate, what to ask, which metrics, what good looks like. Harvey has run BigLaw Bench since August 2024, evaluating itself against the foundation model companies and declining to reference other legal AI vendors at all. Legora and Listen are running versions of the same play.
Kyle's framing is that the best companies write the RFP, and he's right that it's the highest-leverage move in the study. If you define the evaluation criteria, you shape which product wins before the comparison starts.
It didn't fit the tree because it isn't a positioning answer. It's a play. The tree asks what you should say, and this is a thing you do.
And it still isn't a moat.
The test is copyability. Harvey has held BigLaw Bench for two years, which sounds like durability until you notice that Legora is running the identical play now and nobody had to ask permission. Self-published and self-scored means any funded competitor ships their own benchmark next quarter, with their own criteria, favoring their own product. Fin's twenty questions are genuinely useful and several of them are visibly shaped to favor Fin. That's good marketing. It's also entirely reproducible.
There's a version of this that converts into a barrier: when a third party adopts your methodology as the standard. A procurement template that cites your framework. An industry body that adopts your metric. At that point the criteria stop being yours and start being the category's, and the switching cost lands on everyone else.
Nobody in the study has reached that yet.
Own the RFP and you win the deal in front of you. That is worth a great deal. It is not the same as being hard to replace.
Positioning isn't a barrier
Set against the content plays, the things that hold have a common shape. Hamilton Helmer's formulation is the cleanest: a real advantage needs both a benefit and a barrier. Most of what gets called a moat in this category is a benefit with nothing behind it.
Those labels aren't decoration. They're four of the eight moats I score companies against, and they showed up because they're the four Kyle's exemplars are actually holding.
Which makes the absences worth naming. Nothing on the tree routes toward , where each new customer makes the product better for every other customer. Nothing routes toward that someone else controls and you have. Nothing toward economics or .
That's not an oversight on his part. Those four are hard to see from a website, because they don't produce marketing collateral. A company with a genuine network effect rarely writes a compare page about it. Which is exactly why a positioning scan can't find them, and why half the defensibility map is missing from the tree by construction.
Here's what the full map looks like on a company outside Kyle's fifty, so the picture isn't a verdict on anyone he named.
Five of eight at 2 or higher. Ecosystem is the lone 3, and it's the moat HubSpot markets least. Network Effects scores 1 because a new customer at one company doesn't make the product better for another. Regulatory and Physical are 0 because the category doesn't offer them.
See HubSpot's full diagnostic →Four of those eight bars would be invisible to a website scan. Ecosystem shows up because partners talk. Network Effects, Scale, Regulatory and Physical either don't produce marketing or, in HubSpot's case, register as absent because the category rules them out. You can't tell those two situations apart from the outside, which is the whole problem with reading defensibility off a compare page.
Two questions separate these from everything on the tree.
If both answers are no, you have a barrier. If either is yes, you don't have one yet. You may still have something worth holding: distribution, reputation, implementation capability, a real head start. Those matter. They just aren't what decides whether you're still in this category in eighteen months.
The test falsifies moats. It doesn't classify everything else, and it shouldn't try.
Both are worth having. A company with barriers and no message loses deals it should win. A company with messages and no barriers wins deals now and loses the category later. The failure mode isn't choosing wrong. It's not knowing which one you're holding.
The caveat inside the strongest branch
I want to close on Glean, because their page is the most technically serious artifact in the study and because the thing that makes it good is also the thing worth thinking hardest about.
Glean doesn't claim better AI. They claim a better context layer around the AI, and they bring workflow evaluations across 175 queries to support it. The comparison goes into handling of sensitive data for unauthorized users, audit logs, compliance APIs, single-tenant deployment. Technical depth used to be too much for marketing collateral. Here it positions them as the party qualified to tell you what to look for.
Single-tenant deployment is central to that case. Enterprise buyers want it and are right to want it.
It also means that nothing any Glean customer does makes any other Glean customer's deployment better. Each customer retains real value, and it accrues inside their tenant rather than across the base. One stock builds. The other doesn't.
That is a wall. Worth building, and this is a good one. It just doesn't grow.
It makes leaving expensive. It doesn't make staying more valuable every month.
Which gives you three objects, not two.
Most of what gets marketed as a moat in B2B software right now is the first thing wearing the language of the third.
Kyle's tree answers the first question: what should we say. The two tests answer the second: what would a competitor have to overcome. What they say is a real input, and I don't want to wave it away. Neither question is the cheap one.
The third is the one I keep coming back to. When the business runs a cycle, does anything accumulate on either side? Something the company keeps, and something the customer keeps. If only one stock is building, or neither, the barrier is just expensive to leave.
Business loops come in five types. User, content, trust, skill and capital. They differ in what accumulates on each side, and in what makes them stall. A loop that stops adding to either stock is a funnel wearing better clothes.
That's a different instrument and a longer argument, and it's what I'll be writing about next.
The 8 Moats Diagnostic
Four moats showed up in this piece. There are eight. The diagnostic scores all of them and produces the part you can't assemble on your own: the list marked structurally unavailable given how you're built today. You can guess at your score. You can't see which doors are already closed. Free, about ten minutes.
Start the diagnosticFor the third question, the one about the two stocks and which of the five loop types you're running, there's the Business Loop Diagnostic. Start with the moats.
Kyle Poyar's original study is at Growth Unhinged. The 50-company dataset is public and worth reading in full.