AI Products Threw Out 40 Years of Empty-State Research. The Retention Curve Shows It.

Open ChatGPT for the first time. You see a blank text box. Open Claude. Blank text box. Open Gemini, Copilot, Cursor, Perplexity, Grok, Le Chat. Eight category leaders, eight blank text boxes, eight prompt strings of suggestion chips that vanish the moment you start typing. The product opens and asks the user to invent the value proposition.

This is not a stylistic preference. It is the consumer-facing collapse of a discipline that took four decades to build. Adi Leviim’s analysis in UX Collective documents the discipline that died and the retention curve that resulted. We have written before about design systems as governance for the agent era, about design tools that bypass the constraint layer, and about governance of AI output inside design surfaces. Those essays argued the producer side. This one is the empirical receipt: when you remove the empty state, users do not stay long enough to discover what your product is for.

The Discipline That Was Thrown Out

The empty state was never just a screen. It was the place where Don Norman’s signifiers, Jakob Nielsen’s heuristics, Steve Krug’s “don’t make me think,” Bret Victor’s direct manipulation, Kathy Sierra’s Suck Threshold, and Erika Hall’s content strategy converged into a single product surface. The empty state told the new user three things at once: what this product does, what to do first, and what success will look like. It was the moment where the design system did its hardest work.

Forty years of accumulated research went into making that moment legible. Then a single product category replaced all of it with the same artifact: a text input, a placeholder string, and four suggestion chips that disappear on first keystroke.

The replacement happened so fast and so uniformly that nobody flagged it as a design choice. It read as inevitable. The model is the product, the prompt is the surface, and the suggestion chips are the only signifier. That is not an evolution of the empty state. That is its abolition.

The Retention Curve Is the Receipt

Leviim publishes the first-party retention numbers from a Chrome extension launch, and they are the cleanest indictment available:

100% of users install
30% open a second session
12% reach a fifth session
4% become weekly active

Roughly 70% of new installs never return for a second session. The product never gets a chance to teach the user anything because the user closed it before a single round trip with the model finished.

Curves like this are not new. SaaS dashboards have produced them for years. What is new is that every category leader in AI ships the same first-screen experience and gets the same drop-off. ChatGPT, Claude, Gemini, Copilot, Cursor, Perplexity, Grok, Le Chat. Eight different teams, eight different funding stories, one identical empty state, one identical leak between install and habit. When eight competitors converge on the same anti-pattern and the same failure curve, the cause is not any single team’s onboarding choices. The cause is upstream of any of them.

Why the Suggestion Chip Is Not a Signifier

Defenders of the blank-prompt design point to the suggestion chips. The chips, they argue, do the work the empty state used to do. They show what the product can do. They invite a first action.

The chips are not signifiers in Norman’s sense. Signifiers are persistent. They sit on the surface and remain available as the user reads, considers, and chooses. The AI suggestion chip is the opposite: it appears, the user reads it, the user starts typing, the chip vanishes, and the user is now staring at a blank text box again with no memory of what the product was capable of three seconds ago. The chip is decoration that auto-removes on contact.

The empty state’s job was to scaffold the user’s first action by reducing the space of plausible inputs to a small, legible set, with examples that survived as the user worked. The chip’s job is to make the empty space look less empty. Those are not the same job. The product team confused decoration for affordance, and the retention curve recorded the cost.

The Producer-Side Argument, Now With Consumer-Side Data

We argued earlier that design systems are the governance layer for AI-era products. The thesis: when models can render anything, the design system stops being a component library and becomes a constraint layer that determines what gets shipped. We argued that design tools that bypass that constraint layer destabilize teams who depended on the design system to encode discipline. We argued that AI output without governance produces interfaces that look novel and feel hostile.

The empty state collapse is what those arguments look like from the user’s side of the screen. The constraint layer was the empty state. The design system used to make sure the new user never landed on a blank canvas without scaffolding. AI product teams shipped without that constraint, and 70% of installs walked back out.

Note who is implicated. Not startups working out of a coffee shop. Not first-time founders without design budget. The category leaders. Companies with millions in annual design spend, mature design systems, named design VPs, published design guidelines. The constraint layer was thrown out by the teams most equipped to defend it. That is the data point worth dwelling on.

What the Constraint Should Reclaim

The empty state needs to come back as a governed surface, with the same discipline a design system applies to color, type, and component composition. Concretely:

Show the product’s capability surface, persistently. Not as chips that vanish. As a structured legend that survives the user’s first interaction. The user typing into the box should still be able to see, in their peripheral vision, the eight things this product is good at and the two it is not.

Show the user’s history, even on day one. A new user has no history. Show them sample conversations from the team that built the product. Show them other users’ patterns, anonymized. Show them the most common first prompts in their organization. Empty does not have to mean empty.

Show the cost of the next action. The user is about to spend tokens, time, or both. The empty state used to communicate the unit of value. The blank prompt box communicates nothing. A governance-respecting empty state shows a credit balance, a token estimate, a privacy mode indicator, the model that will answer.

Show the way out. Every empty state in the pre-AI era told the user how to undo, how to restart, how to escape. The blank prompt offers a single submit button and an implicit “good luck.” Restoring the way out is the cheapest, most obvious win.

None of these are speculative. All of them existed in production software a decade before transformer models did. The teams shipping AI products have the institutional memory to reintroduce them. They have not, and the retention curve is the cost they are absorbing in lost users every day.

A Note on the Source

Leviim’s piece is sharp on the discipline and clean on the data, and the author’s commercial interests in third-party AI UX gap-fillers are disclosed in the original. We use the data and the canonical citations, not the product recommendations.

Do This Now

Pull your AI product’s first-session retention numbers. If the install-to-second-session rate is below 50%, the empty state is the suspect, and the next 30 minutes of design work has a higher expected return than the next 30 days of model tuning. Look at the first-session screen with the design system in hand and ask: which of the four properties above is present? Which got cut on the way to launch?

The teams that win the next two years of AI product growth are not the ones with the smartest models. They are the ones whose first screen passes a 1985 usability test.

This analysis synthesizes The Death of the Empty State in AI Products (UX Collective, May 2026).

Victorino Group helps product and design leaders reclaim the empty state as a governance surface, not a blank prompt. Let’s talk.