Hook model (Nir Eyal): apply it, audit it, know when to walk away

The Hook Model is a four-stage loop for designing products that build habitual use without requiring paid advertising or aggressive re-engagement campaigns. Nir Eyal published it in Hooked (2014), synthesizing B.F. Skinner’s variable-ratio reinforcement schedules with BJ Fogg’s behavioral model (Motivation x Ability x Trigger). In interviews, the skill is applying the loop to a real product, connecting each stage to a measurable signal, stating when the model is the wrong tool, and naming the ethical test on the spot. Reciting the four stages with Instagram likes as the only example is the failure mode: interviewers at top companies have heard it hundreds of times and it signals pattern-matching, not product thinking.

The four stages

Trigger. The cue that initiates the behavior. External triggers are notifications, emails, and badges. Internal triggers are emotional states: boredom fires TikTok; uncertainty fires Google; loneliness fires iMessage. The goal is to anchor an external trigger to an existing internal one until the internal trigger fires on its own. In an interview, always identify the internal trigger for the product in question. A product with no internal trigger anchor has no hook.

Action. The simplest behavior the user can perform in anticipation of reward. Fogg’s equation governs this stage: behavior happens when motivation, ability, and trigger align. The PM’s job is to maximize ability (reduce friction) and time the trigger to a moment of high motivation. Duolingo’s five-minute lesson is a friction reduction as much as a product decision; it keeps ability high even when motivation is marginal.

Variable reward. Unpredictable payoffs reinforce behavior more durably than predictable ones. Eyal defines three types:

Rewards of the Tribe: social, unpredictable validation (likes, views, comments, follower counts). Instagram Reels drove a reported 30% increase in time-on-app after launch because the social reward is both unpredictable and tribe-facing.
Rewards of the Hunt: resource-seeking where the outcome is uncertain (LinkedIn job alerts, feed scrolling, email triage). The variability is in what you find, not whether you will find something.
Rewards of the Self: mastery and completion (Duolingo streaks, Wordle, Linear’s keyboard-shortcut feedback). The reward is intrinsic, tied to a sense of capability or progress.

Spotify Wrapped is worth understanding structurally: it is a single annual investment-trigger event, not a daily loop. It works because the investment (listening history) is deep and the reward (the reveal) is high-surprise. It was shared over 225 million times after the 2023 edition. It is not a hook in Eyal’s sense; it is a one-shot variable reward that drives an annual external trigger. That distinction matters in an interview.

Investment. The user puts something in: data, content, connections, preferences. Investment loads the next trigger and makes future reward more personally relevant. A Spotify user who has curated 40 playlists gets a Wrapped that means more than one who streamed passively. Investment that compounds user capability is the good version. Investment that only increases switching cost without improving user output is the exploitative version.

The Manipulation Matrix: Eyal’s own ethical test

Eyal defines a two-axis test for whether habit design is ethical. Axis one: does this product materially improve the user’s life? Axis two: would the creator use it themselves?

Yes on both: facilitator. Build the hook.
No on both: exploiter. Stop.
One yes, one no: gray zone requiring explicit PM judgment.

This is not a complete ethics framework, but it is a fast on-the-spot test that signals depth. State it explicitly when any interview question has an engagement or retention angle: “I’d apply the Manipulation Matrix. Does this mechanic improve the user’s life, and would I use it myself?” Then give your actual answer for the product in question, not a generic disclaimer.

The regulatory environment has sharpened this from a theoretical concern into a viability constraint. Belgium and the Netherlands ruled loot boxes (a direct variable-reward mechanic) as gambling under existing law. The EU Digital Services Act curbs addictive design patterns for platforms with 45 million or more EU users. Consumer PMs at companies with EU exposure have to factor compliance costs into retention design decisions.

When not to use it

Low-frequency B2B utilities. A tool used once a month for a specific workflow does not need a hook; it needs reliability and speed. Habit design applied here produces dark patterns, not retention.

Safety-critical products. An alert system or medical device should not have variable reward mechanics. Predictability is the feature.

AI agents designed for task delegation. If the product’s job is to act on the user’s behalf so they do not have to be present, dependency is a design failure. A hook that makes users return compulsively to an agent contradicts the agent’s purpose. The product should be working even when the user is not watching.

The 2026 AI inversion

In 2026, any team can ship a trigger system and a variable reward loop in a weekend. The mechanics of the Hook Model are engineering table stakes, not a competitive advantage. The question is no longer “can we build a hook?” but “does this hook attach to a real job worth doing?”

The specific failure mode for AI products is accidental habit formation around cognitive offloading. A June 2025 study by Gerlich found that heavier AI tool reliance correlated with measurable cognitive offloading and weaker critical thinking. A variable-reward slot-machine pattern in an AI assistant can make users return compulsively while their own judgment and skills atrophy. They are hooked, but the hook is attached to a craving the product manufactured, not a problem they genuinely have.

Eyal’s own follow-up book Indistractable (2019) argues that the same psychological levers used to hook users can and should be given back to users as self-management tools. His pivot from “build better hooks” to “build for user agency” is worth citing at companies with a stated user-welfare position (Anthropic, Duolingo, any health tech).

The opposite design posture is the compounding model: reliable utility that improves with use, without manufactured variability. Mind the Product reported in June 2026 that some AI products deliberately design for restraint, refusing to grab attention when a task is done. For AI-native products, the compounding model is often a stronger design position than a hook. Lovable, not just usable in 2026 means investment compounds user capability, not switching cost. The viable test is whether users return because the product solves a real problem, not because a notification manufactured a craving.

Habit-formation KPIs

If you apply the Hook Model, you need metrics for whether it is working:

D1/D7/D30 retention curves. Is the gap between D7 and D30 narrowing across cohorts? Compression means the habit is forming faster with each product iteration.
DAU/MAU ratio. Above 0.5 is a strong habit signal for consumer apps. Below 0.2 on a product that needs daily engagement means the hook is not landing.
Median session interval compression. Are users returning more frequently over their first 30 days? This is the clearest behavioral signal that an internal trigger is forming.
Investment depth score. Connections made, content uploaded, preferences set, integrations configured. A proxy for how much future reward is loaded into the product for this user and a leading indicator of churn risk.

Use it in an interview, do not recite it

strong

"I'll walk the Hook Model for Duolingo, then tell you when I'd use a different model. The internal trigger is mild guilt or identity attachment to being someone who learns languages. The action is one five-minute lesson, friction-reduced to the point where ability is never the barrier. The variable reward is the weekly streak league: your rank against other learners shifts unpredictably based on their effort, which is a Reward of the Tribe with a Hunt element. The investment is the streak count and vocabulary history, both visible and both gone if you churn.

Before extending that model I'd apply the Manipulation Matrix: does the streak mechanic improve the user's language ability, or does it optimize for return visits at the expense of learning depth? The research on spaced repetition suggests short daily sessions do improve retention over long infrequent ones, so the mechanic aligns with the user's goal here. I'd draw the line at notifications timed to exploit loss aversion past a certain frequency, because at that point the hook is manufacturing anxiety rather than serving the learner.

Metrics: D7/D30 retention delta to track habit consolidation, DAU/MAU ratio targeting above 0.4, session interval compression in the first 30 days as a leading indicator, and investment depth via lesson history and streak count.

Where I'd walk away from this model entirely: an AI agent designed for delegation. If the job is to take work off the user's plate so they are not present, dependency is the failure mode. I'd design for compounding utility instead: the agent gets more capable as it learns context, users return because it keeps working, not because a variable reward loop pulls them back."

weak

"The Hook Model has four steps: Trigger, Action, Variable Reward, and Investment. A good example is Instagram. You get a notification, open the app, see an unpredictable number of likes, and then post more content to keep the loop going. It's important to consider the ethics of this model because it can be addictive." This fails because it uses the example cited in every article on the topic, names no metrics, has no opinion on when the model is wrong or harmful, cannot connect variable reward type to the product's specific dynamic, and the ethics acknowledgment has no substance. The interviewer has heard this exact answer hundreds of times. It signals you read a summary, not that you have product judgment.

The viable and lovable standard applies here directly. Variable reward is necessary but not sufficient: it must reward a real job, not manufacture a craving around a non-problem. Investment must compound user capability, not just switching cost. In 2026, the failure mode is shipping a polished hook that drives DAU/MAU while users’ actual outcomes stagnate or decline. That is a retention metric that will not survive a viability audit and it will not survive an interviewer who asks what success actually looked like for users.