Daniel sent us this one, and it's a good one. He's asking about the metrics that actually matter when evaluating startups, specifically in a moment where AI has blown the doors off the barrier to entry and the number of new founders has surged. MRR, ARR, net revenue retention, CAC, LTV, churn, engagement, burn multiple, Rule of 40. The full toolkit. And underneath all of it, the harder question: which of these numbers actually distinguish a genuinely promising startup from one that's just telling a good story?
The timing on this is not accidental. AI-powered startups are up something like forty percent compared to last year. The number of people who can credibly show up to a pitch meeting with a working product, or at least something that looks like one, has just exploded.
Which means the signal-to-noise problem for investors has gotten dramatically worse. When anyone can build, the metrics become the conversation. Or they should be.
By the way, today's episode is powered by Claude Sonnet four point six.
Our friendly AI down the road, doing the heavy lifting. So let’s get into it, because there’s a lot of ground to cover here—especially how the ecosystem shift is reshaping what investors prioritize. Some of it is counterintuitive.
That ecosystem shift matters because it changes what investors are actually filtering for. When you had fewer startups, you could afford to spend time on narrative. Now, with Q1 data showing AI startups captured eighty-one percent of global venture capital funding—something like two hundred and forty-two billion out of two hundred and ninety-seven billion total—the competition for attention is ferocious. Investors are running pre-due-diligence filters before they’ll even take a meeting.
The metrics become a first pass. A gate, not a conversation.
And the gate has specific criteria. Traction, market size, capital efficiency. If you can't show measurable signals on those three axes, you don't get to explain yourself. Which is why understanding the full toolkit matters, not just the headline number.
The toolkit has layers. There's the revenue layer, which is your monthly recurring revenue, your annualized version of that, your net revenue retention. Then you've got the customer economics layer, cost to acquire a customer, lifetime value, how long it takes to pay back what you spent acquiring them. Then churn, which sounds simple but splits into at least four meaningful variants. Then engagement, which is where a lot of the AI-specific weirdness lives. And then efficiency metrics, which are basically the investor's way of asking whether you're spending money intelligently.
None of these live in isolation. That's the thing people miss. A great monthly recurring revenue number next to a terrible churn number tells a completely different story than either one does alone. The metrics form a system. You have to read them together.
Which is probably why investors who just chase the top-line number get burned. The story is always in the combination — and that starts with the revenue layer.
MRR and ARR are the numbers that end up in headlines, and they're also the numbers that get abused the most. There was a piece in the Economic Times recently about how ARR has become Silicon Valley's hottest metric and simultaneously its least trusted. There's a case where a founder, the CEO of a company called Cluely, just admitted he fabricated a seven million dollar ARR claim.
That's a remarkable thing to admit publicly.
It is, and it points to something structural. Usage-based pricing models, which are increasingly common in AI products, can swing revenue by twenty percent week to week. So even when founders aren't lying, the number is volatile in a way that a traditional subscription ARR figure isn't. You're comparing apples to something that isn't even consistently a fruit.
Which means the headline ARR figure, taken alone, tells you almost nothing about trajectory. What you actually want to know is whether that number is sticky.
That's where net revenue retention comes in, because it's the stickiness metric. The calculation is: take your revenue from a cohort of customers at the start of a period, and then measure what that same cohort generates at the end, including any expansions, upsells, contractions, and churned accounts. If you end up with more than you started with, your NRR is above one hundred percent. The best SaaS companies run above one hundred and twenty percent. What that means practically is that even if you stopped acquiring new customers entirely, you'd still grow.
Which is a completely different business than one with ninety percent NRR that needs to keep running hard just to stay flat.
Right, and the gap between those two companies might look identical in a headline MRR figure at any given snapshot. That's why sophisticated investors treat NRR as the more honest revenue number. MRR tells you where you are. NRR tells you whether the ground is moving under you.
Let's stay on this for a second, because there's a tradeoff hidden in optimizing for NRR that I think is worth naming. If you're pushing hard on upsells and expansions to inflate that number, you can mask underlying problems with the core product.
You absolutely can. High NRR driven by expansion into a small, loyal customer base is a different story than high NRR across a broad, diverse customer base. The first one is fragile. One or two large accounts churn and the metric collapses. So you want to look at NRR alongside logo count and how concentrated your revenue is. The top ten customers shouldn't be fifty percent of your ARR unless you're explicitly an enterprise play and you know what you're signing up for.
Okay, so from revenue into customer economics. CAC, LTV, payback period. These feel like they should be straightforward but they're not, are they.
They really aren't. Customer acquisition cost sounds simple: how much did you spend to acquire a customer. But the denominator is always in dispute. Do you include only sales and marketing spend? Do you include the cost of the sales team's time? Product-led growth complicates it further because some customers acquire themselves through a free tier, and the marginal cost looks artificially low until you account for the infrastructure carrying all those free users.
A startup can show you a very attractive CAC number that's basically a definitional choice.
The lifetime value calculation has the same problem on the other side. LTV is typically modeled as average revenue per account divided by your churn rate. Which means a small change in your churn assumption produces a massive swing in LTV. If you're modeling five percent annual churn versus eight percent, your LTV estimate can differ by sixty or seventy percent. Founders know this. Some of them use it.
The CAC payback period is maybe the cleanest of the three then. Just: how many months of gross margin does it take to recover what you spent acquiring that customer.
It's more grounded, yes. The average for AI startups has come down from around eighteen months last year to something closer to twelve months now, which reflects both better go-to-market efficiency and the fact that some AI products are landing in organizations faster because there's genuine urgency to adopt. But twelve months is still twelve months of capital tied up before you're in the black on that customer. For a capital-constrained early-stage company, that math matters enormously.
Churn is the variable that can quietly destroy all of it. Because a twelve-month payback period only works if the customer is still there at month thirteen.
Which brings us to gross versus net churn, and logo versus revenue churn, and this is where I think the misconceptions pile up fastest. Gross churn is the revenue you lost from cancellations and downgrades. Net churn subtracts expansion revenue from that. A company can show zero net churn, or even negative net churn, while still losing a meaningful percentage of its customer base, because the customers who stayed expanded enough to cover the ones who left.
Logo churn is the headcount version. How many accounts actually cancelled.
Logo churn is the one that predicts future problems before they show up in revenue. If you're losing thirty percent of your logos annually but your revenue churn looks fine because your big accounts are expanding, you have a retention problem in the long tail that will eventually catch up with you. There's a pattern in B2B SaaS where a company scales fast, the large accounts love the product, the small accounts never quite get there, and the logo churn in the small segment is a leading indicator of a ceiling on total addressable market penetration. You can't grow into that segment if you can't retain it.
There's a version of this that ends companies. Where the metrics look fine until they very suddenly don't.
There's a classic case in the enterprise collaboration space, not going to name the company, but the pattern was: explosive MRR growth, strong NRR, and logo churn in the small and medium business tier that everyone internally knew about but that wasn't surfaced in the investor deck. The large accounts held the revenue together for two years. Then two of the largest accounts consolidated vendors after acquisitions, and the company lost thirty-five percent of ARR in a quarter. The logo churn had been the warning sign the whole time. It just wasn't the metric anyone was watching.
Because it wasn't the metric anyone was reporting.
Which is the deeper problem with the whole category. The metrics that get reported are the ones that look good. The metrics that predict failure are the ones that require you to look.
Let's talk about engagement, because this is where AI products specifically get strange. DAU/MAU ratios, daily active users over monthly active users, are a fairly blunt instrument in a consumer context. But in B2B SaaS they've historically been treated as almost irrelevant. The argument was always: if the customer is paying and renewing, who cares how often they log in.
That argument has collapsed in the AI era, and here's why. A lot of AI products are priced on consumption. Tokens, API calls, seats with usage minimums. If your DAU/MAU ratio is low, it means users aren't building the product into their workflows. And if it's not in the workflow, the contract renewal conversation becomes a negotiation from weakness. The customer has optionality because they haven't become dependent.
Whereas a product with a DAU/MAU ratio above, say, sixty or seventy percent is one that people open every morning. That's a different renewal conversation entirely.
The benchmark for sticky B2B tools tends to be somewhere between fifty and seventy percent DAU/MAU. Consumer social apps aim higher, obviously, but for business software hitting sixty percent means the product is embedded. Below thirty percent and you're in the territory where the customer might not notice if you disappeared for a week.
Which is a terrifying place to be when your contract is up for renewal.
The deeper version of this is cohort retention. Not just whether users are active today, but whether the cohort that onboarded six months ago is still as engaged as the cohort that onboarded last month. You can have a healthy aggregate DAU/MAU and still have a retention cliff at month three or month six where engagement drops sharply after the initial novelty wears off. That cliff is the real signal.
AI products are particularly vulnerable to that cliff, aren't they. There's a pattern of people trying something, being impressed, and then not integrating it into daily practice.
The numbers bear that out. There's been consistent reporting on enterprise AI tool adoption where initial activation rates look strong and then ninety-day retention tells a different story. The challenge is that AI tools often require workflow change, not just adoption. You don't just add the tool, you have to reorganize how you work around it. Companies that don't invest in that change management piece see the cliff.
Which means the cohort retention curve is actually measuring something about the company's customer success function as much as it's measuring product stickiness.
That's a really important reframe. A steep drop at day sixty isn't necessarily a product problem. It might be an onboarding problem. The question is whether the company knows the difference and is fixing the right thing.
Okay, efficiency metrics. This is the part of the conversation that I think has shifted most dramatically in the AI era. Burn multiple, magic number, Rule of 40. These were always important but they've taken on a different character.
The Rule of 40 is probably the cleanest place to start because it's the one that shows up most in investor conversations. The calculation is straightforward: take your revenue growth rate as a percentage, add your profit margin as a percentage, and if the sum is above forty, you're in healthy territory. A company growing at sixty percent annually with a negative twenty percent margin clears it. A company growing at twenty percent with a twenty percent margin also clears it. The rule is trying to capture the tradeoff between growth and profitability in a single number.
The implicit argument being that you can be unprofitable if you're growing fast enough, but at some point the market stops believing that.
That inflection point has moved. In the zero-interest-rate environment of a few years ago, the Rule of 40 was almost a formality. Investors were funding companies that were growing at triple digits with deeply negative margins and calling it fine. The recalibration since then has been significant. Now the Rule of 40 is treated as a floor, not a ceiling, particularly for late-stage growth rounds.
The burn multiple is the more granular version of the efficiency question, right. It's asking how many dollars of cash you're burning to generate each new dollar of net new ARR.
You take your net cash burned in a period and divide by net new ARR added in that same period. A burn multiple below one is exceptional. One to one and a half is good. Above two starts to raise questions. Above three is a problem. What the metric is really asking is: is your growth capital-efficient, or are you basically buying revenue at a loss and hoping scale eventually fixes the economics.
The AI founder wave has changed what's achievable here. Because the cost to build has come down so dramatically.
This is where the comparison to traditional SaaS becomes interesting. A traditional SaaS startup in, say, 2019 or 2020 needed substantial engineering headcount just to get to a functional product. Infrastructure costs were meaningful. Time to first revenue was typically measured in years. An AI-native startup today can reach a working product faster, with a smaller team, which structurally improves the burn multiple in the early stages.
The benchmarks have moved. What was a good burn multiple for a traditional SaaS company at series A isn't necessarily the bar for an AI startup anymore.
The benchmarks have moved, but there's a catch. The early-stage efficiency advantage is real. The challenge is that AI infrastructure costs can scale in non-linear ways. If your product is built on top of a large language model and your usage grows, your cost of goods sold grows with it in a way that doesn't happen with traditional software. So the burn multiple can look great at low revenue and then deteriorate as you scale, which is the opposite of the pattern you want.
The magic number is trying to measure something related but slightly different, isn't it. It's more specifically about sales efficiency.
The magic number takes your net new ARR for a quarter, multiplies by four to annualize it, and divides by the prior quarter's sales and marketing spend. Above one is the benchmark. It's essentially asking: for every dollar I spend on go-to-market, am I generating more than a dollar of annualized revenue. It's a tighter lens than the burn multiple because it isolates the go-to-market motion rather than the whole business.
AI startups can look deceptively good on this metric early on because a lot of initial growth is word of mouth or developer-led adoption, which doesn't show up in the sales and marketing denominator.
Which inflates the magic number artificially. Then the company tries to scale go-to-market, adds a sales team, starts running paid acquisition, and the magic number collapses. It wasn't that the go-to-market was efficient. It was that there wasn't really a go-to-market yet. The early adopters came to you. The question is whether you can go to them.
The AI founder wave has created a generation of startups with very attractive early efficiency metrics that may not survive contact with a real sales motion.
That's the structural tension. And it's why investors who understand this space are increasingly asking to see the magic number before and after the company started investing in sales. The delta between those two numbers is the real signal.
Right, and that delta is key. But for an investor or founder trying to apply this framework, where do you start? We’ve covered a lot of ground, and not all of it is equally tractable.
If I had to pick two metrics that I'd want to understand deeply before anything else, it would be net revenue retention and CAC payback period. Not because the others don't matter, but because those two together tell you the most about whether the business has durable economics. NRR above a hundred and twenty percent means your existing customers are growing their spend faster than you're losing accounts. CAC payback under twelve months means you're recovering your customer acquisition cost quickly enough that growth doesn't require permanent subsidization from new capital.
Those two are harder to manipulate than top-line ARR. The Cluely situation is a good reminder that ARR can be fabricated or inflated or defined creatively. NRR requires you to actually track individual customer cohorts over time. It's a more demanding calculation.
The practical implication for someone evaluating a startup, whether as an investor or a potential employee or a customer, is to ask for the cohort data behind the NRR number. Not just the headline figure. Ask what the NRR looks like for cohorts that are twelve months old versus twenty-four months old. If NRR is strong at twelve months and then deteriorates at twenty-four, that's a product that people try and gradually abandon. The headline number was masking a slow bleed.
On the CAC payback side, the question worth asking is what's included in the CAC calculation. Some companies define customer acquisition cost narrowly, just paid media and sales commissions. Others include the full loaded cost of the sales team, onboarding, implementation support. The companies that use the narrow definition can show a twelve-month payback that's actually eighteen or twenty when you account for the full cost of winning and activating the customer.
That definitional slippage is everywhere. Which is why the burn multiple is a useful cross-check. If someone claims a twelve-month CAC payback but the burn multiple is two point five, something doesn't reconcile. You're burning two dollars and fifty cents for every dollar of new ARR, which is inconsistent with the efficient acquisition story they're telling.
The metrics form a system, as you said at the top. Inconsistencies between them are the tell.
For founders, the actionable version of this is: build your internal dashboard around the metrics that are hardest to game, not the ones that look best. Track logo churn even when it's uncomfortable. Track your magic number before and after you start scaling sales. Track cohort retention at ninety days and one eighty days, not just the aggregate active user count. The discipline of measuring the hard things is what separates founders who understand their business from founders who are telling themselves a story about it.
That's exactly what separates the startups that survive the next funding round from those that don't.
Honestly, that survival rate question is what keeps me up at night. We're in a moment where the metrics to evaluate startups have never been more sophisticated, the information has never been more available, and yet the signal-to-noise problem is arguably worse than ever. AI lowers the barrier to building, but it also lowers the barrier to telling a convincing story about building.
The open question for me is whether the metrics themselves will keep pace. Usage-based pricing already breaks the traditional ARR model in ways the industry is still working through. Agentic AI products, where the software is actively doing work rather than just being a tool, introduce entirely new questions about how you measure value delivered. Is it tasks completed? Is it outcomes achieved? Is it time saved for the human on the other end?
The outcome-based measurement problem is going to be the next frontier. And it's hard. If an AI agent closes a support ticket autonomously, how do you price that, how do you measure retention around it, how do you calculate the LTV of a customer using a product where the unit of consumption isn't a seat or a feature but a result. The whole framework we've been discussing assumes a relatively stable relationship between usage and revenue. Agentic AI breaks that assumption.
Which means the investors who figure out the right leading indicators for that model first will have a significant edge. The ones still asking about DAU/MAU ratios for an agentic product are going to miss what's actually happening.
The landscape is going to keep demanding new mental models. That's probably the most honest thing we can say about where this is heading.
Big thanks to Hilbert Flumingtop for producing this one. And Modal keeps our pipeline running smoothly, as ever, one serverless GPU at a time. This has been My Weird Prompts. If you've found this useful, leave us a review on Spotify, it helps people find the show. We'll see you next time.
