When it comes to proving their solidity and success, many startups use an indicator that is supposed to cut short any discussion: Annual Recurring Revenue, or ARR.
With business models shifting towards subscriptions, it has the advantage of talking about the present while providing guarantees for the future: customers are engaged and know that it is easier to subscribe than to cancel, so on a given day, you can say how much revenue you will earn at a minimum over a year.
This is more secure than selling individual units.
Take software, for example. In the past, you would buy software for, say, $200, but you could never upgrade to the next version or skip a few. Today, you pay $10 a month and get automatic updates… as long as you keep paying. If you stop, you don’t keep an old version that, ultimately, does the job.
That’s the beauty of Software as a Service.
But the same goes for music. You can try Spotify and then switch to Apple Music.
Depending on the series you want to watch, you might subscribe to Netflix one month and Prime the next.
So we have a figure that is supposed to summarize the strength, traction, and recurrence of the business model, but it’s a misleading figure.
If business models have evolved towards a “painless” model that encourages purchasing and makes us forget that we can/should cancel our subscription, consumers have adapted and also changed their consumption behaviors.
- ARR reflects recurring revenue based on subscriptions, but assumes increasingly illusory customer stability.
- No-engagement models have changed behavior: users test, switch, and cancel quickly, reducing the predictive value of ARR.
- Even in B2B, actual usage and loyalty are no longer guaranteed despite annual contracts.
- Other indicators such as retention, active usage, or NRR offer a more reliable view of sustainability.
- In a volatile market, a model aligned with actual usage is better than a high but misleading ARR.
The illusion of recurrence
ARR is based on one and only one assumption: stable, predictable income spread out over time. This assumption no longer holds true as models have evolved toward greater flexibility.
In the beginning, when billing was monthly, the engagement was usually annual, which made ARR meaningful.
But as we know, the easier it is to disengage, the easier it is to engage, and so we have seen the emergence of models without engagement. Here again, the idea is that the price is low enough that the bill is lost in the vastness of recurring monthly payments, to the point that you forget about it and never think to cancel it even if you no longer need it.
Add up all your supposedly painless monthly subscriptions and you’re sure to be surprised.
But consumers have adapted. Faced with a plethora of offers, they test, switch from one tool to another, compare several tools for the same need, mix free and paid offers, and subscribe on an ad hoc basis for a specific need.
And, above all, they cancel. They cancel a tool to try out a competitor, subscribe for a specific need for a month, and cancel preventively as soon as they subscribe.
This phenomenon could reach fever pitch with generative AI. There is a plethora of solutions available to meet every need. And for the same need, who can say which is the best model based on the specific expectations of each user and the specific use case, when new products are released every week? No one. So we’re feeling our way in the dark.
When I talk about this with people around me, it’s not uncommon to hear about someone canceling their ChatGPT subscription to try Claude or Mistral, then backtracking, signing up for a one-month Midjourney subscription just for a one-time need, and so on.
If at any given moment the revenue is certain, recurrence no longer means anything in a market that is not consolidated. You can announce $1 billion in ARR one day, and if the next day a competitor creates a buzz with something more relevant, simpler, and cheaper, you can find yourself at $700 million the following month.
In terms of predictability, we’ve seen better.
ARR no longer provides a reliable means of projecting into the future and is increasingly becoming a snapshot. It is an accounting reality on a given day, but not proof of customer loyalty and engagement.
Often flattering, sometimes misleading.
And at a time when the AI market is hyper-dynamic and everyone is testing everything and anything before moving on to the next thing, I wouldn’t be surprised to see a churn phenomenon never before seen on the market when the market stabilizes and consolidates a little.
Even in B2B, loyalty is no longer guaranteed
In my opinion, this phenomenon is very strong in B2C, but we shouldn’t think that B2B is completely immune.
In B2B, annual contracts are still common, but they no longer guarantee active use or automatic renewal. Here too, behaviors have changed (the famous consumerization of the business world) and, except for truly critical applications, customers disengage well before the contract expires or reduce their scope at renewal time based on actual usage.
In fact, today many businesses are conducting pilot programs, most often on a paid basis, in the field of AI, using competing technologies that they compare before making a choice. There is no doubt that most of them consider these potential customers as acquired and include the future revenue in their ARR as if the trial had been converted. At the end of the pilot programs, there will probably be only one winner, with the impact on ARR that one might imagine.
But ultimately, it’s not all about behavior: it’s the business model itself that encourages such behavior.
Durability vs. recurrence
ARR provides a surface-level view of the business: it highlights growth but can mask the volatility of the customer base.
Once recurrence is no longer a given, especially in an immature market, it may be necessary to look for what could be interpreted as signs of durability.
The good news is that these indicators are well known; all that remains is to use them or reevaluate their importance.
• Cohort retention: which customers remain after 3, 6, or 12 months?
• Active usage: are customers really using what they are paying for?
• NRR (Net Revenue Retention): one of the few indicators to include expansion, churn, and contraction.
• Retention cost: how much does it cost to keep each customer? Sometimes more than acquisition.
• Compatibility between the pricing model and perceived value: a pay-per-use model may be healthier than a poorly priced subscription.
These indicators are of course monitored in most businesses (at least I hope so), but are rarely used for external communication, unlike ARR. Perhaps this is not surprising…
Is ARR the world of yesterday?
ARR is a relevant metric in certain contexts, such as when a sector is growing rapidly, rates are low, and fundraising is easy. By giving the illusion of stability and predictability, it is perfect for reassuring and attracting investors.
However, in a more constrained environment where profitability becomes a priority, it may be less relevant.
A healthy startup, at least in the current context, is not necessarily one with a high ARR, but one with a model aligned with actual usage and perceived value that allows it to survive short cycles, budgetary trade-offs, and behavioral volatility.
It’s when the tide goes out that you see who’s been swimming naked, and when the frenzy around AI calms down, either because the market matures or because individual and business customers have to arbitrate their spending more strictly, we may be in for some surprises.
Bottom line
ARR remains a useful indicator, but it is only one indicator among many and is not the most reliable for measuring the resilience of a startup in an uncertain environment.
When customers are exploring and testing solutions in an immature market, when every customer is watching every dollar they spend, value creation is more about retention than acquisition, and it may be worth looking at metrics that are closer to reality.
Today, ARR tells us about the present, but it is more difficult to use to predict the future.
Image credit: Image generated by artificial intelligence via ChatGPT (OpenAI)
