Author: andrew chen, a16z

Compiled by: Tim, PANews

I have been observing retention curve data for over 15 years.

I have seen thousands of retention curves, and this is one of the first metrics I ask to see when evaluating startups. I have sifted through thousands of databases and analyzed retention curves broken down by various segmentation dimensions. As a product builder, I have also looked at this metric from another perspective. I have run hundreds of A/B tests, drafted countless versions of user onboarding guides and notification emails, trying to change the shape of the retention curve.

A/B testing (also known as split testing or bucket testing) is a random experimental method used to compare two versions of a product (version A and version B). Its core purpose is to determine which version performs better in achieving predetermined goals by collecting data and analyzing user behavior.

From the results, there are some patterns.

Like physical laws, it is strange that over time, certain deterministic patterns keep emerging. Here are a few examples I want to share:

• You cannot improve a poor user retention rate. Yes, adding more notification features will not improve your retention curve. You cannot achieve good user retention through A/B testing.
• Retention rates only decline, they do not rise. And strangely, the rate of decay does follow a predictable half-life pattern. Early retention rates can predict later retention performance.
• Revenue retention expands, while usage retention shrinks. The good news is: although users will gradually churn, the users who remain sometimes spend more!
• Retention rates are closely related to your product category. There are both inherent reasons and cultivated factors. Unfortunately, you are destined not to make a hotel booking app a daily-use product.
• When user expansion and growth occur, retention rates tend to decrease. The highest quality users come from early and organic growth, while users acquired later tend to perform the worst.
• User churn is asymmetric; losing a user is far easier than winning them back.
• Calculating retention rates is very difficult. Seasonal factors do exist, newly launched test versions can interfere with data, and system bugs can occur. While D365 is a real metric, it should not be the only result considered.
• Viral growth with poor retention will ultimately fail. We have repeatedly verified this conclusion across multiple platforms and categories.
• Excellent user retention is a miracle. When you truly witness such a miracle, it is incredibly shocking.

We will analyze these points one by one.

You cannot save a poor user retention rate. You have seen this situation firsthand: you spend months developing a new product, and then it launches. The initial user retention data is dismal. At this point, product development has been underway for months, and it is difficult to turn things around; how can you improve retention? Suddenly, you have a bright idea: why not add push notification features to remind users to come back? Or add a bunch of new features? Or A/B test the landing page to improve conversion rates?

I think we all know how this will end. Unfortunately, when a product's retention rate performs poorly, it is often extremely difficult to turn it around; it can almost be said to be hopeless. Of course, marginal improvements might be achievable. Suppose your day-two retention rate is 40%, and the goal is to increase it to 50%; this is entirely feasible and worth the effort. But if the day-two retention rate is only 10%, it likely means that the product you built does not meet market demand at all, and all the localized optimization efforts around A/B testing and push notifications will not be enough to turn the fundamental situation around. When months of development time and sunk costs are already a reality, it is hard for people not to struggle desperately. But I believe that in most cases, it is best to make a decisive choice to pivot.

This pivot aimed at improving user retention requires a complete redesign of the app's homepage. If it originally presented as an information flow model, it may need to shift to a structured step-by-step process; if the core of the product is sharing functionality, perhaps the focus should shift to content creation and curation. You may need to describe the product positioning in a completely different way, or even align it with competitors. This must be a large-scale transformation across multiple dimensions; the more thorough, the better. Only in this way can the low user retention situation be turned around.

Retention rates will decline but will not rise. Retention curves typically present as very regular geometric curve patterns. For example, many curves I have observed exhibit the following pattern: regardless of the first-day retention rate, the seventh day will see a 50% decline; regardless of the seventh-day retention rate, the thirtieth day will see another 50% decline. Over time, the final retention rate may approach zero, and if lucky, it might maintain around 10%. This decay pattern is predictable.

You have never seen a curve that rises after initially declining; that is impossible. In other words, if early retention rates are not outstanding, then late retention rates are likely to be poor as well. You must start strong to finish well.

There are some noteworthy exceptions to this rule that need to be pointed out:

• Some products are very hardcore (e.g., online poker). The user retention rate for such products may be relatively low, but the users who remain are often extremely loyal and spend significantly, proving that this model can succeed.
• For products with network effects (possibly social networks, collaboration tools, or other products with network effects), new users may initially be active, but their activity may temporarily decline. However, if the product can leverage an increasing number of users to reactivate old users, a slight recovery in retention rates may occur. This situation is extremely rare, but when it happens, it is astonishing.
• Revenue retention expands while usage retention shrinks. One of the best and most important characteristics of the retention curve is that it can apply to both users and revenue. So far, we have been discussing user retention, but unfortunately, user retention always shows a downward trend, which is not ideal. On the other hand, revenue retention is interesting because the users who remain often spend more money on your platform over time.
• This is one of the biggest advantages of B2B SaaS products. Take Slack as an example; if you look at user group data, you will find that its retention curve also shows a downward trend, just like other products. Some people accept it, while others do not. But for companies that invest time in deploying Slack, the product begins to grow naturally, and the revenue you receive from these businesses increases accordingly. The revenue retention curve does not decline but instead rises, which is a very magical phenomenon, but unfortunately, it does not apply to most consumer products. It is this characteristic that makes B2B products have a smoother business model than consumer products.
• The model of consumer applications is closer to Amazon; you may initially only purchase books and music, but as the product's functionality expands, you gradually start using it to buy more and more items. For this reason, the lifetime value of users in the product essentially has no upper limit. We have observed similar phenomena in Uber: although the user group may decline over time, the spending on rides initially used for airport pickups gradually expands to restaurant outings or commuting scenarios. Therefore, the user retention curve shows a downward trend, but the revenue retention curve continues to rise.
• Retention rates are closely related to product categories. I have previously written about the inherent and cultivated factors regarding retention rates. The reality is that many products have natural usage scenarios, such as collaboration tools or programming software, which you may use daily, but the upper limit of usage days is five active days out of seven in a week. In contrast, a vulnerability alert system hopes users do not use it frequently. Consumer products are similar; people check news, messaging, and social applications daily, but typically do not frequently use medical reference guides. Some applications may have low usage frequency but high retention rates, such as weather or banking applications. In contrast, categories like games may be addictive and frequently used, but people usually churn within weeks after consuming the content.
• The importance of inherent and cultivated factors lies in the fact that they reveal the reality that many new products find it fundamentally difficult to break through. If you are developing a social travel application, but people do not travel frequently, then creating a product centered on friend interaction will be very challenging. A wiser approach is to accept its low-frequency usage attribute, enhance monetization capabilities by controlling the transaction process, or integrate high-frequency usage scenarios like restaurants and nightlife, as Yelp does, while retaining travel functionality. It is indeed difficult to go against the trend, and there is very little we can do.

For this reason, if you want to create an application with extremely high retention and usage frequency, you may need to choose areas that users already consider core daily products for development. This means that successful applications are likely to take up the usage time of other daily products, just as my frequent use of ChatGPT has significantly reduced my Google search frequency; when I started using Substack to read and write blogs, I gradually abandoned various social news applications.

When user scale expands, retention rates often decline instead of rising. Even if you are lucky enough to create a high-retention product, people often instinctively extrapolate the behavior patterns, monetization capabilities, and usage habits of existing users directly to a broader market, believing that multiplying several good small data points with core big data will naturally yield impressive macro results. But the reality is often: as the user base grows, problems begin to emerge. For example, when you start expanding to Android users and international markets, acquiring more customers through paid marketing and other channels, you will soon find declines across all key metrics.

The reason is that high-quality users tend to appear early. The user groups with the highest monetization potential, strongest willingness, highest digitalization level, and most active online behavior typically start using the product early through friend recommendations. As new users are acquired from other channels later, the product may not meet their needs as well. For instance, if you develop an iPhone app for university students in Western countries, when expanding to emerging market Android users, the metrics will naturally decline due to functionality settings not being fully compatible. Although continuous optimization and improvement can occur later, I can assure you that the results will never compare to those of the early user group.

So, the question arises: as the user base grows and user quality gradually declines, do they still hold value? Can the product continue to be profitable? More importantly, can the core high-value user group that joined early be retained?

No wonder these early users are often referred to as the "golden group."

User churn is asymmetric. Losing users is extremely easy; in fact, most products lose 90% or more of their users within the first 30 days. Meanwhile, winning back lost users is extremely difficult. This asymmetry between acquisition and churn is the core characteristic of user churn. The reality is often so dire that it is easier to acquire new users than to try to win back old ones.

For this reason, attempting to awaken dormant users through lifecycle marketing by sending discounts or offers is often costly and yields little effect. A more effective approach is to have existing active users awaken dormant users through the product's natural usage scenarios. For example, if a professional tries a new project management tool but fails to continue using it, sending bombarding reminder emails to their inbox is unlikely to win the user back. A more effective approach is to have their colleagues invite that user back to participate in new projects with the tool; that is the effective way. However, this strategy is extremely difficult to implement and is exceptionally complex, typically only applicable to products with network effects (i.e., sharing and collaboration features).

Retention rates are very tricky and difficult to measure. When people talk about retention rates, they often tend to measure the first day, the first week, and the first month, but rarely discuss what happens two years later. This is because, during product development, teams need a sufficiently short time span and easily measurable metrics to make decisions. Therefore, even though annual user churn rates or long-term profitability are extremely important, people often do not measure these and instead focus on the easily measurable metrics at hand. However, this approach has many problems.

Unfortunately, many product categories are strongly affected by seasonal fluctuations. E-commerce, travel, health services, and online dating are typical examples. Even the way businesses use commercial software has cyclical variations. Seasonal factors can interfere with judgment; you may find monthly or quarterly data declining, but is it because the newly launched features are unpopular? Or is it that user behavior patterns differ in this quarter? When retention data is severely lagging, it is indeed difficult to conduct effective assessments.

Similarly, factors such as software bugs, newly conducted tests, or new marketing campaigns can disrupt data. Eventually, you will find yourself constantly reviewing reports that show fluctuations in retention curves, but each piece of data comes with additional explanations because the team needs to verify whether the newly launched Android version has led to unrelated comparisons.

Crazy user growth with terrible retention rates is destined to fail. Many new product developers often focus excessively on new user registrations while completely neglecting user retention. After all, if you only want to see a continuously rising curve, why not just expand the traffic at the top of the funnel to showcase rapid growth? After raising a large amount of venture capital, you can slowly address the user retention issue later, right?

This phenomenon is common in the current industry: a creator promotes their app to millions of fans, or a video post leads to a surge in revenue, and the product experiences a user explosion through TikTok. Although the actual usage rates and user churn are not ideal, this phenomenon continues to occur.

The tech industry has conducted countless experiments like this. The conclusion is the same: products that go viral but have poor user retention will ultimately perish because retention issues are difficult to resolve. When the novelty wears off, user acquisition will slow down, and eventually, you will face a bleak situation where both user acquisition and user retention are dismal; the higher you climb, the harder you fall.

We have witnessed this phenomenon in many scenarios. In the early stages of social networks, many products grew by crazily sending spam emails to acquire user emails and contacts, but ultimately directed users to low-quality products. Sometimes, as long as users can subscribe to certain low-quality ringtone annual services, companies can attempt to monetize and make a profit. But it wasn't until Facebook emerged, with innovations in user experience such as news feeds and real-name systems, that a product was finally created that had both high viral spread characteristics and strong user stickiness. The same situation occurs in the mobile app space; sometimes you see apps that suddenly become popular by forcing SMS invitations, but if the product lacks stickiness, the entire model will quickly collapse.

High retention rates are almost magical. You may feel a bit frustrated after reading this article; I know that launching projects can sometimes be tough. But when a product truly works, that feeling is unparalleled. When you witness a product achieving a 50% 30-day retention rate (which I see once every few years), the shock is indescribable. I gradually realized that these fleeting successful products are not because the creators have a systematic A/B testing methodology or achieve goals through high-speed iteration processes; the real key is that they need that little bit of magical insight. This magic comes from breakthrough insights into market or customer needs, which, although seemingly obvious in hindsight, allows the product to achieve extremely high retention rates by being the first to realize this understanding. Today, we evaluate video conferencing software, ephemeral photo features, or magical AI that can respond to any topic in the same way; this magic cannot be achieved solely through iteration and metric-driven testing.

The Real Question

You may still have a big question after reading all of the above: So how do you achieve high retention rates? (If I could answer this question with certainty, my job as a startup investor would be so much easier, wouldn't it?)

But let's do our best. In my previous points, there are actually some clues buried: ideas are really important.

If you want a product with a high retention rate, you need to choose a category that inherently has a high retention rate.

You need to choose a product category that you are already using existing products in every day.

You will build a product that competes directly with it.

If you win, then you will stop using that product and switch to using your own product.

This is a high bar, but I think it's a good start to think this through.

Of course, if the product you create directly competes with existing products, you may wonder, "Isn't it really difficult to get users to switch camps?" Indeed, it is. Therefore, at this point, you need to decide to take on enough market risk, but it must be moderate risk, by launching a novel and unique product that redefines the core interaction model. However, the innovation referred to here is more likely to be a 20% improvement rather than an 80% disruptive innovation. Ideally, you should be able to help users quickly and intuitively understand this innovation within the first minute of using the product.

At this moment, you cannot avoid one of the most frequently asked and hardest to answer questions from investors: "Why is this feasible now?" Because your answer must point out that a new trend has emerged in the industry, such as a general technology like large language models, or social changes like the saturation of social media, making your innovative idea timely.

This can help you quickly capture the existing market and is more likely to achieve excellent user retention in the early stages. Timing is crucial. If you miss the timing, enter a low-attention field, and your product differentiation is not prominent enough, you will find that you have merely transformed the user retention problem into a user acquisition challenge. The difficulty in developing a new web browser lies in the fact that once successful, user stickiness will be extremely high. However, people are already very satisfied with existing browsers, and getting users to try a new product requires costly and complex efforts.

This is why I do not blame those who propose ideas like "the Cursor of a certain field" or "the Figma of a certain industry," just like the past concepts of "the Uber of a certain vertical." They are trying to leverage existing markets and behavior patterns to avoid huge market risks.

If you can accurately grasp the differentiation advantages, hit the market timing, align with massive user needs, and pinpoint the core product positioning, then this model can indeed succeed.

How to Open New Markets?

The natural opposing viewpoint is that new markets are often more exciting than existing markets. Shouldn't the tech industry be about building entirely new things rather than innovating 20% on old ones? Of course, this is true, but I believe that such products account for an extremely small portion.

My counterpoint to this is that, in fact, most products inherit some "old things," even if those predecessor products are quickly forgotten.

Before Instagram, there was Hipstamatic, which early on ranked at the top of the paid photography apps in the App Store, confirming the huge market potential of filter features. Just as Google was not the first search engine (it was actually the tenth entrant, following platforms like Lycos, Excite, and Infoseek), these cases both prove the strong demand for search functionality and reveal the commercialization dilemmas of early search engines. Tesla was not the pioneer of electric vehicles, and the iPhone was not the first smartphone. History repeatedly proves that the true market landscape is often determined by the tenth-generation innovators. This phenomenon is known as the "latecomer advantage," and I find this perspective very enlightening.

However, sometimes true innovation does happen. The birth of Uber transformed the existing offline taxi behavior into an online application, rather than being based on a previously successful ride-hailing app (at the time, Lyft was still just a quirky bus booking service). Looking at ChatGPT, OpenAI took five years from concept to the third version's true rise, during which there was no existing blueprint to refer to. This kind of innovative journey is extraordinary and is the driving force behind the thriving tech industry, as it creates entirely new product categories at the cost of taking on real risks.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。