For the better part of two decades, the dominant playbook for digital products was simple: make them sticky. Build enough habit, enough convenience, enough switching cost, enough re-engagement that people keep coming back. In many categories, that worked. It produced retention, scale, and revenue.
When activity gets mistaken for value
I would call that a dependency model.
The problem isn’t that organizations optimized for dependency. It’s that they mistook dependency for value. Activity gets mistaken for value when a metric tells us whether people showed up, but not whether the product actually served them.
People may depend on a product because it’s genuinely useful – saving time, reducing effort, or solving a real problem. But dependence can also come from a habit that’s hard to break, high switching costs, or a design that makes leaving harder than staying.
From the dashboard, those situations can look identical. The metrics that defined the last digital era were very good at measuring activity, which in turn justified product decisions without taking the time to disentangle human value from engineered dependency.
This isn’t theoretical. You can see it in the data.
In a 2024 Harris Poll conducted with Jonathan Haidt’s research team, 82% of Gen Z adults said they associate social media with feeling addicted. Not engaged. Not entertained. Addicted.
AnswerLab’s own 2026 loyalty work shows a similar split between what behavior says and what people say. 76% of consumers who are active in loyalty programs don’t consider themselves genuinely loyal to the brands behind them.
Usage can mask ambivalence, which is exactly where relying on behavioral metrics alone can lead product teams in the wrong direction. People can look highly engaged without feeling well served.
Why this moment is different
In the engagement era, when the dependency model took hold, most products sat next to our decisions. They shaped what we saw, what we clicked, what we watched, and what we came back to. But most of them weren’t making the decision itself. Obviously, AI is changing that.
The near-term version looks familiar: an assistant recommending the next step, a system surfacing options, an agent helping us narrow the field. But the trajectory is not toward better recommendations. It is toward action. Not just suggesting the purchase, but making it. Not just flagging the bill, but paying it. Not just surfacing options, but choosing one.
And we can already see the line people are drawing around those actions. In our work on agentic experiences, users let AI help with discovery, comparison, and prep, but they pull it back at the moment of commitment: final purchase confirmation, itemized cost review, data-sharing opt-ins, and account-level changes. This line completely changes the stakes of measurement.
When a product is merely trying to keep you engaged, a weak metric can create a bad experience, a manipulative pattern, or a trust problem that builds over time. These are bad enough. But when a product is acting on your behalf, the cost of getting the metric wrong moves much faster - from a matter of slow-boil philosophical frustration to one of immediate, meaningful consequence.
The tolerance for sloppy proxies shrinks.
Right now, many organizations are still reaching for the same measurement instincts they used before: engagement, usage frequency, session depth, retention. Familiar metrics applied to a fundamentally different relationship.
That is the part I think product leaders need to take seriously. This isn’t just a new product category, it’s a new level of product power. AI systems are being given greater authority to act, more autonomy in how they act, and less human oversight between recommendation and consequence.
And if the last era taught us anything, it's that the metrics organizations normalize early do not stay contained to the dashboard. They become incentives. They become design logic. They become the behavioral environment in which people live.
The question before us isn’t just, "What should we measure now?" It’s, "What kind of human relationship are we quietly building our metrics to produce?"
A red team for your metrics
I don’t have a list of new metrics for us to chase. In fact, I’d argue that simply swapping one KPI set for another isn't the answer at all. Most organizations already have plenty of metrics. The deeper issue is that we do too little interrogation of what those metrics are actually doing.
I think what product leaders need right now is a red team for metric design.
In security, red teams exist to surface the vulnerabilities a system creates before someone else exploits them. Metrics deserve the same treatment.
Before a KPI gets embedded in a roadmap, before it shapes incentives, before teams start optimizing toward it in earnest, someone should be explicitly responsible for pressure-testing what it rewards, what it hides, and what kind of user behavior it is likely to create.
This doesn’t need to be a large approval committee. A small cross-functional group across product, research, design, and data – with engineering, AI, trust and safety, legal, privacy, or operations added as needed – can pressure-test a metric before it enters the roadmap or incentive system.
The goal isn’t another layer of review. It’s a clear decision about what the metric rewards, what it hides, and whether it reflects value or merely activity. Assumptions and risks should be documented, and then a decision made about whether to adopt the metric, revise it, pair it with guardrails, or reject it.
I’d suggest starting with three questions:
1. What behavior does this metric actually reward?
Not the behavior you hope it reflects. The behavior it will actually incentivize once you start optimizing for it. Because this is where metrics become most dangerous.
Watch time was not created to encourage compulsive scrolling. It was created as a plausible proxy for value. If people keep watching, the thinking went, the content must be working. But the behavior it reliably rewarded was not meaningful engagement. It was continued consumption. Content built to prevent stopping. Interfaces built to collapse reflection. Systems optimized for persistence, not necessarily benefit.
That pattern is not unique to media. Any time you set a metric, you are setting off an incentive chain. Internal teams optimize for it, product decisions get made to move it, and users adapt to the resulting experience.
The question isn’t whether that chain exists – it always does.
The question is whether you have traced it honestly enough to know what behavior you would actually be creating in people’s lives if you succeeded. If we hit this number, what will we have actually built?
2. What is this metric a proxy for, and have we ever measured what's actually underneath it?
Most product metrics are proxies.
- Engagement stands in for value.
- Enrollment stands in for loyalty.
- Frequency stands in for usefulness.
- Retention stands in for satisfaction.
- Reliance stands in for trust.
In nearly all situations, this is reasonable, because the thing you actually care about is almost always harder to measure than the behavior standing in for it. It’s a useful shortcut - until that shortcut becomes the destination.
When enrollment becomes loyalty by default because it is the only thing being tracked. When usage becomes usefulness, because nobody has asked the user whether it was actually useful. When reliance becomes endorsement, because the product was used repeatedly.
The problem starts when the stand-in becomes the definition. Once that happens, teams stop asking whether the product is creating value and start managing to the metric instead. The way to correct this issue is not through measurement perfection. It is the discipline to look underneath it, using research to understand whether the behavior you're measuring reflects the value people are actually experiencing.
When’s the last time you checked whether the metric is actually measuring what users themselves say they are getting?
That is the part too many teams skip. The proxy is convenient, continuously measured, and already embedded in how performance gets discussed. But if you don’t regularly look below it, you can’t be measuring with confidence. You’re just perpetuating assumptions.
3. Would people still choose the behavior this metric is driving if they paused to think about it?
This is the hardest question, but probably the most important one.
A behavioral metric can indicate whether someone is using a system, delegates to it, or stops checking its output. But behavior alone doesn’t tell you whether the person would defend that pattern if you held it up in front of them and asked, plainly: Is the relationship you have with this product actually good for you? Is this the relationship you want?
That distinction matters.
It is possible to rely on something without endorsing it. To use something often without feeling well served by it. To hand something over to a product because it is easy, familiar, or hard to escape, not because it meaningfully improves your life.
That is the blind spot in a lot of product measurement.
It captures what the user did. It doesn't always capture how they make sense of what the product is doing for them. Or to them. And in the AI era, where systems are increasingly positioned to act on our behalf, that difference will matter even more.
A person can let a system do something and still feel uneasy about the broader pattern it is creating. They can benefit from the convenience while distrusting the system. They can accept the delegation while resenting the shape of the dependence.
If you are not checking for that tension, your metrics may be overstating alignment at the exact moment it matters most.
What this changes for product leaders
None of this will make metric-setting easier. But it will make the stakes clearer.
The tension here, of course, is that human outcomes and business outcomes aren’t always aligned. Products that are harder to leave are often commercially attractive. Friction that protects users can suppress conversion. Slower trust-building can look less impressive than aggressive automation - until it breaks.
Those are real tradeoffs that should be treated as decisions, not defaults.
I won’t promise that a red-team approach to metrics will deliver a perfect KPI, but it will provide visibility into the value system you’re embedding before it disappears into execution.
That matters because teams are giving AI more authority before they can accurately judge whether its outcomes are actually good for the people on the other side.
The leaders who get this right will get clearer about the human outcome they’re trying to create, look past easy proxies, and be honest about whether their metrics capture value or just activity.
That’s the difference I think will matter in the AI era. Not whether people keep coming back, but whether they’d still choose the relationship with your technology if they could see what it was becoming. The products that win in the long term won’t be the ones people feel stuck with – they’ll be the ones people trust enough to depend on.
