User Retention Rate Analysis

According to a16z- AI products and "AI Tourists" are changing everything we knew about retention and how we understood it. Till very recently M1 (30 day/ Month 1) retention was gospel. Founders swore by it, Investors wanted it. You acquired users on Day 0, you checked how many stuck around at Day 30 and you knew you had something. The retention graph would then start flattening and stabalize- signalling that these users have found something "they want". "AI tourists" are changing that. People will happily pay $20 to try a shiny new AI tool. They’ll poke around for a month or two… and then vanish. That doesn’t mean your product is broken. It means you’re seeing tourists churn before you find your residents. And the process is taking longer. In this new world, M1 tells you almost nothing- because there is so much Tourist noise. The real test starts at M3. - By Month 3, the hobbyists are gone. - What’s left are users who’ve discovered real, repeatable value. - This is your true base. Now here’s the kicker: the leading AI products aren’t just retaining- their curves sometimes start to “smile.” Which basically means instead of the retention curve only stabilising it starts moving up- -As capabilities improve, churned users come back. -Retained users expand usage. That’s why we have to start thinking in terms of: M3 is the new M1 (who survives the tourist churn) M12 is the new gold standard (how strong your core really is over a year) The most powerful metric in AI today isn’t M1 retention. It’s M12 ÷ M3. That ratio tells you how well your committed users behave once the noise is gone. It’s a forward indicator of long-term net dollar retention (NDR), LTV, and whether you can really scale. So if you’re building in AI: Stop obsessing over M1. Track your curve to M3, then to M12. That’s where the truth is hiding.

The hidden cost of 10 million orders For context: Femi shared that Chowdeck delivered 10M orders in 2025 to 2.1M registered users. Incredible milestone. But the math tells a different story: 10M orders ÷ 2.1M users = ~4.8 orders per user/year That means the average Chowdeck user orders once every 2.5 months. For more context: • Uber Eats targets 8-12 orders per user annually • DoorDash power users order weekly Chowdeck doesn’t have an acquisition problem. They’re already #1 by downloads and order volume. They have a frequency problem. Here’s why this matters: If each existing user places just one extra order in 2026, that’s 2.1M additional orders at near-zero acquisition cost. At ~₦3,500 AOV → ₦7.35bn in recovered revenue. So where does frequency actually break? From what I’ve seen across food & commerce platforms, it usually leaks in 3 places: 1. The forgetting curve Users order once, have a good experience, then… forget the app exists. Solution: Behavioral email triggers ⚫️ 7 days since last order: “Miss you! Here’s what’s new” ⚫️ 14 days: “Your favorite restaurant has a new menu” ⚫️ 30 days: Free delivery to bring them back 2. The discovery gap Users order from the same 2-3 restaurants and never explore. Solution: Personalized recommendations ⚫️ “Based on your love for [Restaurant X], try [Restaurant Y]” ⚫️ “New restaurants near you this week” 3. The silent churn Users stop ordering and no one notices until they’re gone. Solution: Reactivation campaigns ⚫️ 60 days inactive: “We saved your favorites” ⚫️ 90 days: “Here’s ₦500 off your next order” Why most platforms miss this: They optimize heavily for growth, not retention. But the economics are brutal: • New user CAC: ₦2,000–₦5,000 • Reactivation via email/push: ₦200–₦500 Retention is ~10x cheaper than acquisition. As Nigerian food delivery apps expand into quick commerce, bills, and super-app territory, the winners in 2026 won’t be the ones with the most users. They’ll be the ones who get existing users to come back more often. Because at scale: frequency > acquisition. … What are your thoughts on this? 👇

Netflix doesn’t wait until month 12 to learn you’re gone. The platform knows by episode 3. B2B SaaS churn works the same way: 71% of cancellation intent surfaces in the first 30 days. Essentially, day 1 - 30 is the verdict window. - Only 28% of users who fail to reach first value inside two weeks renew a year later. - Accounts that activate three core features in month one renew at a 92% clip versus 58% for single-feature tourists (per Gainsight Pulse). - CS teams that run a 30-day “decision audit” see renewal forecast accuracy tighten from around 18% to +/- 7%. Yet most companies schedule the first serious check-in 90 days before renewal, which is LONG after the jury has left the building. Try doing this: 1. Map a Time-to-Impact SLA: first value <14 days, second value <30. 2. Treat early warning signals like pipeline slips. No daily log-ins by day 5? Auto-trigger a guided tour. 3. Escalate risk the same way sales escalates exec involvement. If NPS is < 6 in week three, drop an exec note rather than a generic survey. 4. Push product usage data to CS in hourly feeds, not weekly roll-ups. Retention is the delta between first-month reality and twelfth-month pricing. Nail the former and the latter becomes paperwork. Forecast renewals on behavior you can still change, not anniversaries you can only regret.

Most churn analysis in digital products focuses on a simple yes or no - did the user leave or not. But churn is not just about if, it is about when. The timing matters. That is where survival analysis, or time-to-event analysis, comes in. It is a set of statistical methods designed to answer questions like: How long does the average user stay? How does the risk of churn change over time? Which user groups leave sooner and which ones stick around longer? Survival analysis works especially well in digital product research because it can handle censored data - users who are still active when your observation period ends. Instead of ignoring them or making arbitrary assumptions, the method uses all available information. This means you can work with incomplete churn outcomes without throwing away valuable data. It also adapts naturally to real-world product behavior. Many products have usage in fixed cycles like weekly logins or monthly subscriptions. User behavior can change during their journey, such as upgrading to a premium plan or decreasing engagement after a poor experience. Some users churn and later return, sometimes multiple times. Survival analysis methods have extensions that can account for all of these realities. If you are only using classification models to predict churn, you are leaving insights on the table. Classification tells you who might leave. Survival analysis tells you when they are most at risk, how risk changes over their lifetime, and what factors influence that timing. That knowledge is critical for designing targeted interventions, personalizing retention strategies, and understanding long-term engagement patterns. Modern best practices blend classical survival models like Kaplan–Meier curves and Cox regression with adaptations for digital products, such as discrete-time survival for interval-based data, time-varying covariates to reflect evolving behavior, competing risks models to separate different churn types, and recurrent events models to track leave-return cycles. For small datasets, robust techniques like penalized estimation, bootstrapping, or Bayesian survival can stabilize results.

Every day you wait, another customer quietly walks away. Most SaaS companies react to churn. The smart ones predict it. We tested a system to get ahead of cancellations. Retention loop improved 17%. First, segment users by usage: low, medium, high. Then ID at-risk users. Think: low usage, haven't logged in recently. Send personalized “We miss you” emails with dynamic images (Send47). Follow up with an Awaz AI voice agent offering help. If they re-engage, tag them as active in your CRM. If no response, trigger your cancellation sequence. Proactive outreach, not reactive, made a difference in customer retention and reduced our CAC. More breakdowns in the link in bio.

Why Cohort Analysis Unlocks True Retention Insights 🚀 We recently ran a cohort analysis in AMC to measure not just how many customers buy, but how many come back. Instead of looking at broad repeat rates (which often blend new and long-time buyers), this approach isolates first-time purchasers in a single month and then tracks their behavior over time. Here’s what we found looking at January 2025 first-time buyers: 🔶 1,442,328 customers made their very first purchase with the brand in January. 🔶 440,385 of those customers returned within 2 months. 🔶 408,808 returned again within 4 months. 🔶 238,677 returned again within 6 months. This isn’t just a measure of volume — it’s a direct look at customer stickiness and the long-term impact of acquisition campaigns. But the real power of cohort analysis is how it guides audience strategy: 🔸 Exclusions: If you know a product typically lasts 4–6 months, you can exclude recent purchasers from campaigns in that window, avoiding wasted impressions. 🔸 Retargeting timing: Once you see when repurchase behavior spikes (e.g., months 4–6), you can retarget those exact customers with replenishment messaging right before their expected reorder. 🔸 Campaign efficiency: This ensures DSP and Sponsored Ads are working together — prospecting when buyers are new, suppressing them while they’re “in the product lifecycle,” and re-engaging them at the optimal moment to maximize LTV. By running the same query across multiple months, brands can: 🔸 Benchmark retention and spot seasonal dips. 🔸 Identify which products bring in high-LTV buyers vs. one-time shoppers. 🔸 Align DSP + Sponsored Ads investment with long-term growth, not just immediate ROAS. #DSP #AMC #Amazonads #advertising #amazondsp #BTR #BTRmedia

It is an established fact that it is easier to retain existing customers than to acquire new ones, especially in the current macro environment where budgets are frozen and selling any new software involves a lengthy sales cycle. Keeping a close track of your retention metrics has thus become all the more important. It helps to segregate bad cohorts from good ones, divided by different dimensions such as industry and region. Additionally, it assists in identifying churn signals based on billing and usage. Some of the common metrics that companies can track are: 1. NRR/NDR cohort for Contracted ARR, Billed ARR, and revenue: A cohort of revenue from customers present in a specific time period, including expansions. 2. GRR/GDR cohort: A cohort of revenue from customers present in a specific time period, excluding expansions. 3. Logo retention %: A cohort of the number of customers remaining across months, starting at a specific time. 4. ACV cohort: ACV of customers in that particular cohort. 5. 12-month NDR: Net Dollar Retention for a 12-month period across cohorts. 6. 12-month GDR: Gross Dollar Retention (excluding expansions) for a 12-month period across cohorts. 7. Usage cohort: Usage of a particular set of customers over time. Types of insights companies can generate include: - A particular product and industry of customers are bringing down the entire NRR. - While NRR for contracted ARR is decreasing, ACV is increasing and stabilizing after a point, aligning with the right ICP. - Early cohorts of customers are not performing well. The company acquired a lot of SMBs, leading to higher churn. Benchmarks- NRR (This will vary with size and type of company): - $1-5 Mn: 107% - $5-20 Mn: 110% - $20-50 Mn: 106% - > $50 Mn: 109% At Mantys (YC W23) we work very closely with the customers to automate their cohort building process and segment them by different dimensions in seconds instead of updating 200 excel sheets. For consumption based companies, this becomes even more important to understand the usage trends over time. #cohortanalysis #retention #saas

We helped our customer cut churn by 39% in just 4 months. Another side effect: their MRR also doubled thanks to smart pricing experiments and fixing their churn problem. You can say that it’s impossible to increase revenue with the help of finance. But here’s how we did it: 1️⃣ Data Consolidation & Unit Economics Setup We consolidated all the product and financial data (from Stripe, Apple Store, Quickbooks and their own product) in Fuel. We also created user cohorts and a centralized dashboard to provide daily insights for each cohort, including: ▪ MRR ▪ Lifetime Value (LTV) ▪ Lifetime (LT) and when exactly customers churn ▪ Retention (& Churn Rate) ▪ Cost of Acquisition (CAC) 2️⃣ Churn Analysis Deep Dive  With all this data, customer’s product team figured out why customers were leaving: ▪Who was churning? ▪When were they churning? ▪Why were they churning? In Fuel Dashboard, we show the segments that were struggling and pinpointed the moments where they dropped off. 3️⃣ A/B Testing Pricing Strategies Once they understood the "who" and "when," the customer’s product team ran 4 A/B tests on different pricing models. Using Fuel, they compared key metrics across all test groups: ▪ Revenue metrics (e.g., MRR, ARPU). ▪ Conversion metrics (e.g., sign-ups, upsells). ▪ Retention metrics (e.g., CLTV, renewal rates). After 2 months of testing, they identified the pricing strategy that worked best. May The Profit Be With You 💚

Most teams track retention. Very few actually understan what drives it. They’ll report churn or DAU, throw the numbers on a slide, and call it a day. But those metrics don’t tell you what’s actually driving the change. That’s why I love the “Duolingo Model” for retention. Instead of just tracking churn at the end of the journey, it: ✅ Breaks users into clear states (new, current, dormant, resurrected…) ✅ Tracks how users move between them day by day ✅ Surfaces early signals when behavior is shifting ✅ Lets you run “what-if” scenarios without waiting months for churn data This is powerful because it forces product and data teams to ask sharper questions: Where exactly are users dropping off? Which small improvements compound into meaningful retention gains? At Nextory, I’ve been experimenting with this model to account for multiple profiles and behaviors per account (a very different challenge from Duolingo). The insights have been worth it. If you work in product or data, you need some version of this in your toolkit. 📖 I wrote a full breakdown of how the Duolingo Model works, its limitations, and how to adapt it to your own product (link in the comments 👇)