Interruption cost: the work does not resume where it stopped

What it is

Erik Altmann and Gregory Trafton set out the memory-for-goals model in Cognitive Science in 2002. Its central claim is that an intention is not stored in a special protected slot; it is an ordinary memory representation subject to the same activation and decay as any other. To act on a goal you must retrieve it, and retrieving it competes with everything else currently active. They tested the model against latency and error data from the Tower of Hanoi, a puzzle that depends heavily on suspending and resuming sub-goals.

The prediction that follows is the useful one for interface design: when you are pulled away from a task, the goal you were holding starts to fade, and coming back is a retrieval problem rather than a resumption of something still running. There is a measurable gap between the interruption ending and useful work restarting, called the resumption lag.

Christopher Monk, Gregory Trafton and Deborah Boehm-Davis measured that lag directly in the Journal of Experimental Psychology: Applied in 2008. Across three experiments using an interleaved-task paradigm, longer interruptions produced longer resumption times, and more mentally demanding interruptions produced longer resumption times, with the two factors interacting. The shape of the effect for interruptions up to a minute supported decay as the mechanism, and the interaction supported the idea that rehearsing the suspended goal is what mitigates it.

The design consequences fall out of that shape. Duration matters, so a long modal is worse than a short one. Demand matters, so an interruption that makes you read, decide or type is worse than one you can wave away. And because rehearsal is what protects the goal, an interruption that prevents rehearsal — by demanding the very attention you would use to hold your place — is the most expensive kind there is.

Search engines have made this a ranking matter as well as a courtesy. Google’s guidance on intrusive interstitials tells site owners not to obscure the entire page with an interstitial and not to redirect users to a separate page for consent or input, recommending a banner occupying only a small fraction of the screen instead, and noting that intrusive dialogs make it harder for search engines to understand the content. Legally mandated screens such as age gates are exempted, with overlaying preferred to redirecting.

The level in this game takes the mechanism to its logical end. It opens with one polite pop-up. Closing it spawns two more, which is a rule the pop-ups cheerfully admit to in their own body text. Closing those spawns four. The dialogs are draggable, and the actual solution is underneath them: a notification tray in the corner of the screen, reachable only once you stop dismissing and start moving things out of the way. The level is unwinnable for exactly as long as you keep engaging with the interruption on its own terms.

Why it works on people

A pop-up exploits the asymmetry between what it costs the person who shipped it and what it costs the person who receives it. Adding one is a single component. Receiving one costs the dismissal, plus the resumption lag, plus whatever the user had in working memory and has now lost — a half-typed sentence, a place in a list, the third of four things they came to do.

Because the cost is in retrieval rather than in the interruption itself, users systematically underestimate it and so do the teams measuring it. Dismissal time is easy to instrument and looks trivial. The lag afterwards is invisible in the same data, and the thing that was forgotten never appears in any metric at all.

Interruptions are worst precisely where they are most commercially tempting. A modal fires during a checkout, a form, a search — moments chosen because the user is engaged, which is another way of saying moments when they are holding the most state. The more valuable the moment to the business, the more expensive the interruption to the person.

A cascade compounds rather than adds. Each successive dialog arrives while the goal is already degraded, and each one costs its own resumption. This is the mechanism behind consent flows that ask a second time, upsells that appear after dismissal, and notification permission prompts that arrive during onboarding: no single one is unreasonable, and the sequence is.

The pattern is self-reinforcing in a way that shapes the product. Users learn that dialogs are noise and start dismissing them without reading — which is habituation, and it is the same mechanism that hollows out consent dialogs. Once that reflex forms, the one dialog that genuinely mattered gets dismissed with all the others, and the usual response is to make the important one harder to dismiss, which accelerates the habituation.

Dismissal is not consent, and treating it as engagement is a measurement error. A user who closes a modal in under a second has told you they wanted it gone, not that they considered and declined the offer. Counting those interactions as impressions makes an intrusive pattern look successful in exactly the reports that would otherwise stop it.

Where you meet it

• Newsletter and discount modals timed to fire a few seconds after arrival, or on exit intent, while the visitor is mid-sentence.
• App install interstitials covering a mobile page reached from a search result — the specific case Google’s interstitial guidance addresses.
• Permission prompts for notifications, location or contacts requested on first launch, before the user has any context for deciding.
• Rating and review requests triggered by a timer or a session count rather than by the user completing something worth being pleased about.
• Onboarding tours that open modally over an interface the user is already trying to use, with no way to leave without stepping through.
• Chat widgets that auto-open and take focus, moving the caret out of the form the user was filling in.
• Cookie or consent dialogs that reappear on every navigation because the choice was never persisted, converting one interruption into one per page.
• “Are you sure you want to leave?” dialogs on navigation away, which interrupt at the one moment the user has definitively finished deciding.
• Upsells shown immediately after dismissing a previous upsell, the pattern this level implements literally.

Designing around it

• Interrupt at a boundary, not in the middle. The cheapest moment is one where the user has just finished something and is not holding state — after a save, after a completed order, on returning to a list. The most expensive is mid-form, mid-search, mid-sentence. This single choice matters more than anything about the dialog’s design.
• Prefer a non-modal surface. A banner occupying a small fraction of the screen, an inline notice, or a badge lets the user finish their thought and attend to it when they choose. Google’s guidance recommends exactly this over covering the page, so the accessible option and the search-visible option are the same option.
• Make the dialog cheap to leave and honest about it. One obvious dismissal, reachable by mouse, touch and the Escape key, with a real label rather than a guilt-worded one. The wording of a decline is its own well-documented problem — see the page on confirmshaming.
• Never let a dismissal produce another prompt. Dismissal is a decision. Answering it with a second dialog teaches users that the only way out is to leave the product, and it is the exact behaviour this level satirises.
• Persist every choice, including the negative ones. A “no” that is forgotten on the next page load is not a choice, it is a toll, and it converts a single interruption into one per navigation.
• Help the user rehearse. Because rehearsal is what protects a suspended goal, an interruption that leaves the underlying context visible costs less than one that covers it. Keep the page behind the dialog legible where you can, preserve scroll position and form state absolutely, and return focus to precisely where it was.
• Restore state on return, and treat losing it as a defect. Preserve the caret, the scroll offset, the partially completed field, the open disclosure. Every element you restore is a piece of the goal the user does not have to reconstruct.
• Count dismissals as refusals in your reporting, and record time-to-dismiss. A modal with a median dismissal under a second is not being read, and the metric that makes it look successful is measuring reflex rather than interest.
• Budget interruptions across the whole product rather than per feature. Every team’s prompt is individually justifiable; the total is what the user experiences, and nobody owns the total unless someone is made to.

Questions

What is the real cost of a pop-up?

Far more than the time it is on screen. The expensive part is the resumption lag — the delay after the interruption ends before useful work restarts, while the user reconstructs what they were doing. Monk, Trafton and Boehm-Davis measured this directly in 2008 and found it grows with both how long the interruption lasted and how mentally demanding it was.

Why is it so hard to get back to what I was doing?

Because the goal you were holding is an ordinary memory representation, not a saved state. Altmann and Trafton’s memory-for-goals model treats an intention as subject to the same activation and decay as any other memory, so while your attention is elsewhere it fades, and coming back means retrieving it in competition with whatever the interruption put in your head.

Do pop-ups hurt SEO?

Intrusive ones can. Google’s guidance tells site owners not to obscure the entire page with an interstitial and not to redirect users to a separate page for consent or input, recommending a banner that uses only a small fraction of the screen, and notes that intrusive dialogs make content harder for search engines to understand. Legally required screens such as age gates are exempt.

When is the least damaging moment to interrupt someone?

At a task boundary — just after something is finished, when the user is not holding state. The worst moment is mid-form, mid-search or mid-sentence, which is unfortunately when engagement-triggered modals usually fire. Choosing the moment matters more than anything about how the dialog itself is designed.

Why is a sequence of pop-ups so much worse than one?

Because each one arrives while the goal is already degraded and adds its own resumption cost, so the damage compounds rather than adds. It also trains users to dismiss dialogs without reading them, which is habituation — and once that reflex forms, the one message that genuinely mattered gets swept away with the rest.

Related patterns

• Confirmshaming and the confirmation gauntlet: when “are you sure?” is the trap
• Cookie consent dark patterns: double negatives and asymmetric choice
• Forced consent: unreadable terms and the ritual of agreeing