How Web and Mobile App Teams Choose Performance Goals Users Feel Without Slowing Delivery

Here's the thing most teams get wrong: they chase averages. But nobody actually experiences an average. Your user lives inside one specific moment, and if that moment stutters, they don't think "this is slow" — they think "wait, did I break something?" Slowness reads as doubt, not as lost seconds. That's the real cost, and it never shows up on a median load-time chart.
So we quit optimizing for the median and started obsessing over our angriest 5%. For us the moment of truth is time-to-first-audio — that half second between hitting play and actually hearing a voice. We put a hard budget on it, measured at the 95th percentile on real, beat-up phones and spotty cafe wifi, not the shiny laptops we build on. The gap between those two setups is exactly where churn likes to hide.
The rule that keeps features moving: every new feature has to pay its own way. If it slows the moment of truth, it doesn't get blocked — it just has to fund the speed-up somewhere else in that same flow. Performance and features stop being enemies fighting over the roadmap and start trading in one currency. Engineers actually like it, because there's no vague "make it faster" cloud hanging over them, just a number tied to one screen they can point at.
Funny enough, this made us ship faster, not slower. A clear budget ends more arguments than it starts.

I'm Runbo Li, Co-founder & CEO at Magic Hour.
The only performance goal that matters is the one where a user decides to stay or leave. Everything else is vanity. We don't set abstract latency targets pulled from Google's Web Vitals documentation. We watch where users drop off, and we treat that drop-off point as a hard deadline our product has to beat.
Here's the concrete decision rule we live by: if a performance issue doesn't show up in our retention or completion funnels, it doesn't get prioritized over feature work. Period. We call it "funnel-gated optimization." If users are completing their video renders and coming back the next day, I'm not going to pull resources to shave 200ms off a page transition nobody's bouncing on.
Early on, we noticed that users who waited more than about 90 seconds for their first AI video render had a significantly higher chance of never coming back. That was the number that mattered. So we invested heavily in perceived speed for that specific moment, adding progress indicators, preview frames, and intermediate outputs that gave users something to engage with while the model ran. We didn't make the model faster overnight, but we made the wait feel purposeful. Retention on first-time users jumped noticeably.
The trap most teams fall into is treating performance as a separate workstream with its own roadmap. That's how you end up with engineers spending weeks on infrastructure refactors that users never feel. Instead, we bake performance into the feature spec itself. When we ship a new template or workflow, the spec includes the acceptable wait time at each step, informed by where we see users abandon similar flows. If the feature can't hit that threshold, we scope it down or change the UX to mask the latency.
Ship the feature, watch the funnel, fix what's bleeding. Performance work that isn't tied to a real user behavior metric is just engineering vanity dressed up as discipline.

I choose performance goals by asking two simple questions: where does the data live and how fast do users need the answer. That framing identifies which user journeys need strict, low-latency targets and which can tolerate laxer timings so teams can focus speed work only where it affects perception. My decision rule is this: if the network is unreliable or the user experience is sensitive to delay, place computation locally; if the network is reliable and latency is less critical, use cloud processing. Applying that rule keeps real user speed high on critical flows without forcing a full rewrite or stalling other feature work.

When setting performance goals for our autonomous support agents at AGO, we focus almost entirely on perceived latency. If a customer is trying to process an order or get a refund, they don't care how complex the backend database query is--they just feel the seconds ticking by while the system "thinks."
The one decision rule that helps us improve real user speed without slowing down our feature releases is implementing a strict multi-model routing layer. Instead of passing every new capability through our heaviest, most capable language model, we route basic transactional validation to smaller, much faster models. We save the heavy compute only for complex reasoning.
I saw this dynamic clearly when scaling AI products for millions of users at Leboncoin. If you tie every new feature directly to your most resource-intensive infrastructure, your feature work stalls because optimization takes too long. By decoupling the quick triage from the deep processing, our engineers can push out new capabilities quickly on the heavy models, knowing the initial user experience remains instantly responsive on the lighter ones.

The decision rule that made performance stop being a negotiation and start being a release gate was simple: we picked one user-facing metric that users can feel, made it non-negotiable, and refused to ship anything that broke it.
For us, that metric is time-to-swap. How long between the user tapping "swap" and seeing the confirmation that the transaction executed. Not time-to-load. Not time-to-render. Time-to-swap. Because that's the moment that decides whether a user trusts the app enough to come back.
We set the threshold at under three seconds for 95% of swaps. Not an average. A percentile. Because averages hide the tail, and the tail is where users churn. If your swap takes eight seconds once, you don't remember the four swaps that took two seconds. You remember the eight.
That threshold became an acceptance criterion for every sprint. If a feature we were building would break time-to-swap, the feature didn't ship until we fixed the performance regression it introduced. No exceptions. No "we'll fix it next sprint." The feature waits.
The reason this worked when quarterly performance initiatives didn't is that it removed the negotiation. Performance wasn't a separate workstream we had to advocate for. It was built into the definition of done for every feature. Engineers knew that before they opened a pull request. Stakeholders knew it before they scoped a feature. There was no room for "let's ship this now and optimize later" because later never came.
The constraint this created was that we had to route more than we built. We couldn't ship perpetuals if building the matching engine in-house would have broken our time-to-swap threshold. So we routed perpetuals to Hyperliquid through builder codes. We got matching-engine parity with best-in-class perps without building the matching engine. Same pattern for prediction markets through Polymarket. The performance constraint forced better architectural decisions than we would have made without it.
The result is that performance stopped being something we talked about improving and became something we refused to break. That shift is what made it durable.

Most performance work optimizes things users never feel. You can shave 200ms off a load nobody was waiting on and call it a win, while the one screen people actually sit and wait for stays slow. So the first move is to stop measuring averages. Averages hide the pain — they blend the fast paths that were never a problem with the slow ones that are. Look at the worst cases on the flows people use most, because that's the speed users actually experience.

My rule is to pick a small number of moments that have to feel instant — the ones tied to the core thing people came to do — and set a hard budget on those, then leave everything else alone. On Pitch.ac that's dealing into a game and playing a card; if those feel laggy, nothing else matters. A menu loading 100ms slower, nobody notices.

That's how you keep it from stalling feature work: you're not chasing a global "make everything fast" mandate, which never ends and blocks everything. You're protecting three or four moments and ignoring the rest on purpose. Feature work continues everywhere outside the budgeted paths. The discipline is in what you decide not to optimize.