Every Retention Metric Lies a Little

Retention numbers carry a strange authority in SaaS. They get printed on board slides, quoted in fundraising decks, and used to justify headcount. Very few people in the room ask what the number is actually made of.

That's the problem worth writing about. Not that the metrics are wrong, but that each one is a compression of a messy reality, and compression always loses something. The operator's edge isn't knowing the formulas. Everyone knows the formulas. It's knowing what each number quietly leaves out.

NRR is a revenue metric wearing a customer metric's clothes

SaaS Capital's 2025 retention benchmark puts median NRR at 102% for companies with ACVs between $25,000 and $50,000, with a top quartile of 111%. Precise, useful, and easy to misread.

SaaStr's critique of NRR is the one I keep coming back to, because it names the specific distortions rather than gesturing at nuance. A strong NRR can be flattered by price increases, by tiering changes that push customers into higher-priced packages, by multi-year contract structure, and by a weakening new-logo engine that makes the existing base look proportionally healthier than it is.

Run the arithmetic on that. A company that raised prices, restructured its packaging, and lost a tenth of its customer count can still post a respectable NRR. Nothing about that number is dishonest. It just isn't answering the question most people think they're asking, which is whether customers are choosing to stay.

Worth keeping the base rate in view too: ChurnZero's 2025 Customer Revenue Leadership Study says 74% of participants get most of their revenue from existing customers, and High Alpha's 2025 benchmark says companies above $50M ARR get roughly 60% of new ARR from expansion. When most of the growth comes from the installed base, NRR stops being one metric among many and starts being the whole story. That's exactly when you want to know what's inside it.

GRR is cleaner, not clean

Gross revenue retention is the better read on whether customers stay when they have a real choice. High Alpha's 2025 work says GRR has stabilized across ARR cohorts, with retaining roughly nine out of ten customers now the norm. That's a reasonable bar to hold yourself to.

But GRR has its own distortion, and it's the same one SaaStr flags: contract length. A three-year agreement doesn't prove retention, it postpones the verdict. Your GRR looks stable because the decision hasn't been asked for yet. If a large share of your base is mid-term on long contracts, you're measuring lock-in and calling it loyalty.

I'd argue GRR deserves more board airtime in categories where AI is compressing switching costs. When replacing a tool gets meaningfully cheaper and faster, the gap between "customers stay" and "customers can't leave yet" narrows into something you'd rather find out early.

The non-revenue layer already moved

Gainsight's Customer Success Index 2025 shows the metric mix shifting in a direction that makes sense. Customer health score is now the most common primary non-revenue metric at 72%, product usage sits at 56%, and CS qualified leads reached 52%, up from 27% in 2023. Meanwhile NPS as a primary non-revenue metric fell to 53% in 2025 after peaking at 69% in 2024.

The read there is that leaders increasingly want metrics that can trigger an action, not sentiment snapshots that arrive after the fact. Fine. But health score carries the most personal lie of all the metrics, because you built it. It's a composite of whatever inputs you chose, weighted however you decided, which means it measures your assumptions about value as much as it measures the customer.

If your inputs are login frequency and ticket volume, you've built a metric that stays green for a heavy user who quietly hates the product and is already scoping a replacement. Health scores age badly when nobody revisits the inputs.

The CSQL jump is the more interesting number. Gainsight's own analysis points out that teams risk underreporting their value if they don't formally track CSQLs. That's a metric that lies by omission when it's absent, which is a different failure mode and an easier one to fix.

Time to value is the most useful and least standardized number in post-sale

Look at what the benchmarks say and the definitional problem becomes obvious. Userpilot's 2025 benchmark, drawn from 547 SaaS companies, puts median time to value at 1 day, 12 hours, and 23 minutes. Onboard argues that TTV should be shorter than launch completion and proposes under 30 days as a benchmark for most B2B SaaS, under 7 days for self-serve.

Those aren't in the same universe. That's not a contradiction to resolve, it's evidence that the metric is almost entirely definition-dependent. Self-serve activation and enterprise implementation are different things wearing the same label.

Amplitude's framing is the one that holds up: TTV is the time from sign-up to meaningful benefit, not the time to finish setup. Amplitude also cites a 69% relationship between strong seven-day activation and strong three-month retention, which is the argument for caring about the early window at all.

Most TTV lies are definitional. Teams measure days to launch, because launch has a clean timestamp and value doesn't, then present it as time to value. The fix isn't a better dashboard. It's writing down what "meaningful benefit" means for each segment before you measure anything.

Partner contribution is where the measurement is thinnest

PartnerStack's 2026 GTM partnerships research says only 42% of companies use multi-touch attribution across the funnel, against 31% using first-touch and 19% last-touch.

So roughly half the market is measuring partner contribution with a model that structurally assigns credit to a single moment. Those partner-sourced pipeline numbers aren't fabricated. They're answering a much narrower question than the slide implies, and the narrowness disappears the moment the number gets summarized upward.

This matters more now that partner motions touch onboarding, adoption, and rescue plays rather than just pipeline. If partners affect retention, then partner-attached retention needs an attribution rule that Sales, Finance, and Partnerships would all recite the same way. Most companies can't clear that bar yet.

Stack them in layers, and write down what each one hides

The answer isn't a better metric. There isn't one waiting to be discovered.

Revenue retention still belongs at the board layer, because that's the language of the room. Underneath it, you want earlier and harder-to-game evidence: time to first value against a defined benefit, adoption depth rather than raw usage, renewal forecast accuracy, risk-state transitions, expansion conversion, CSQLs, and partner-sourced or influenced pipeline under an attribution model nobody renegotiates per deal.

Then do the unglamorous part. For each number on your operating review, write one line describing what it can hide. NRR hides pricing and contract structure. GRR hides contract length. Health score hides your own input assumptions. TTV hides the definition of value. Partner-sourced hides the attribution model. That one-line discipline does more for metric credibility than another dashboard ever will.

One honest caveat, since I've quoted a lot of research here. Most of this work is vendor-produced or survey-based: Gainsight, ChurnZero, PartnerStack, Userpilot, High Alpha. It's current, specific, and closer to workflow reality than broad market reports, which is why it's useful. It's directional evidence, not industry law. Which, fittingly enough, is the same caution this whole post is about.

If this is the kind of thing you think about, there's more in the same vein on the blog: https://www.santiagomarin.net/blog