Lovable, not just usable: the 2026 product bar

Usable means a competent person can complete the task. Lovable means they’d be annoyed to lose the product. In 2026, AI raised the floor on usable so high that any competent team can ship something that works. The remaining hard problems are viable (will people pay, is the market real) and lovable (will they stay, return after churning, and tell others). Interviewers at frontier AI companies test whether you can tell those two apart with a real example, not just recite the definition.

Why the bar shifted

The old PM triangle held three hard problems: viable, feasible, usable. Feasibility is now close to free: AI can build almost anything a PM can spec. Usability has a floor: AI-generated interfaces are competent by default. That collapses the triangle from three corners to one genuine challenge. The interview question is no longer “did you design for users?” It’s “do you know the difference between a product that functions and one that earns loyalty?”

The Kano model helps explain why this is a moving target, not a one-time achievement. Features that were delighters in 2022 and performance features in 2024 are now invisible must-haves: AI-powered search within apps, smart autocomplete in text editors, one-click document summarization, personalized recommendations. Candidates who can name specific examples of this decay show actual product taste. The shelf life of a delighter has compressed. What earns love in 2026 must be something AI can’t trivially replicate for a competitor next quarter.

What makes a product lovable

Three properties, in order of how often they’re missed:

• Meets people where they work. Fits the existing workflow and tools rather than demanding a new habit. Notion calendar sync is lovable because it didn’t ask you to leave Notion. A standalone AI calendar app, however clean, asks for a behavior change.
• Anticipates needs without becoming obnoxious. Proactive is good; imposed is not. The senior judgment is knowing the line between the two (more on this below).
• Gets the steps right. The interactions actually complete the job with no dead ends or busywork. “Close but you have to finish it manually” is usable. Finishing the job is lovable.

The Kano decay problem

You don’t achieve lovable once. The a16z AI retention benchmarks (2025-2026) show the strongest AI products have a “smiling” retention curve: users churn in months 1-3 and return later as the product improves. That return is the behavioral signature of a lovable product. They didn’t find something better; they came back to the thing that already had their identity invested in it. The PM who designed that product understood that lovable is earned continuously, not shipped at v1.

Kano delighters that have already decayed to must-haves by 2026: AI chat autocomplete, smart document summarization, personalized content feeds, semantic search within productivity apps. None of these differentiate anymore. The next candidates are: agents that handle multi-step tasks without check-ins, voice interfaces that remember context, and AI that proactively surfaces the thing you didn’t know to ask for. Interviewers notice candidates who can name a specific Kano migration and explain what it means for what to build next.

The obnoxious line

The most common way AI products fail love is over-eagerness. Clippy is the canonical failure: it interrupted, made wrong inferences, and couldn’t be quiet when quiet was right. Early Gemini sidebar integrations repeated this: they surfaced context nobody asked for, mid-task. Notification-spam onboarding flows in AI tools repeat it again.

Interviewers at frontier AI companies now explicitly probe whether candidates know when not to act. The prep community calls this the “quiet tool vs. eager assistant” distinction. A quiet tool does exactly what you asked, then stops. An eager assistant guesses what you might want next and acts on the guess. At the wrong moment, eager reads as obnoxious. The senior signal is understanding that restraint is a product decision, not an omission.

Metrics that signal lovable vs. usable

Standard DAU/MAU metrics misread AI products because agent activity inflates them. Userpilot’s 2026 adoption research found that up to 80% of usage in some accounts is AI agents, not humans. Measuring lovability requires different signals:

• Human return rate after a 30-day gap. Usable products don’t pull people back after they’ve drifted; lovable ones do.
• NPS from power users vs. average users. A large gap means lovable moments exist but aren’t broadly delivered. A consistently high floor means the baseline experience is genuinely strong.
• Referral rate. Users refer products they want to be seen using, not just ones that work.
• Sean Ellis test rate. “I would be very disappointed if this product went away” at 40%+ is the classic PMF signal; it’s also a lovable signal. Below 20% is usable at best.

When usable is enough

Not every product should try to be lovable. Internal B2B workflow tools, compliance dashboards, and admin surfaces often shouldn’t. Trying to make a payroll tool lovable is scope creep in the wrong dimension. The PM judgment is knowing which category you’re in: utility (optimize for completion rate, error rate, support tickets) vs. consumer or competitive SaaS (optimize for retention, referral, and emotional attachment). Showing this judgment in an interview is itself a signal.

How interviewers probe this

Exponent’s 2026 product sense guide notes that a structured answer used to be a reasonable signal; now it’s a baseline. The lovability layer is the differentiator. Meta’s AI product sense round now includes a 30-minute live prototyping session where interviewers watch whether candidates bias toward function (usable) or toward the right user moment (lovable).

The favorite product question is the most common probe. Here’s what separates a weak from a strong answer.

Name the usable version, name the lovable version, and identify one thing the product chose not to do. That structure answers the question and signals product taste in a single move.

In the loop

When you answer product-sense questions and AI design questions, name the lovable version explicitly: “the usable version does X; the lovable version anticipates Y without getting in the way.” That distinction separates a hire from a strong hire in 2026. The viable/lovable frame connects directly to feasibility is free and to the obnoxious AI antipatterns that reveal the same judgment from the other side.