Why AI apps feel fast (even when they're not)

The first time I came across the concept of "perceived performance" was in a post by Julie Zhuo. It was fascinating. Something happening right under my nose — something I had experienced a thousand times — yet not something I had ever consciously designed for (yet). And once you see it, you can’t un-see it.

So what is perceived performance?
While some states are obvious to optimize for: loading screens, onboarding, waiting indicators, the unseen opportunity lies in designing the invisible, the system that the user feels more than sees.

• Actual performance = the time it takes to actually compute, fetch, or generate.
• Perceived performance = how fast the system feels to a human.

The thing is, users don’t experience time linearly.

And great AI products exploit this difference masterfully.

---

1. Speed is psychological

It's not a new concept. The three dots that you see when someone is "typing", initiated at IBM in 1997, has the same purpose. It reduces the perceived wait time and fills the gap while the other person hits send, making the system feel faster and more responsive.

The skeleton loading state that you see in digital products has the same purpose.

---

2. Continuous change feels faster

Instead of waiting idly, seeing the system doing something make the wait feel faster. It's the same logic as going to a restaurant with 10 people and waiting for your ALL of the order to arrive together. The wait is just much much longer (along with the growing thirst in your eyes as each waiter passes by).

What to generally look for —

• Partial answers that grow over time
• Streaming text instead of full blocks
• Skeleton placeholders that resolve incrementally

Users mentally “co-create” the output as it appears.

Let's breakdown Perplexity's micro-animations.

What all animations are happening in the original?

• Shimmer effect through colours
• All sources are broken down instead of showing one long list of sources
• Countdown of the no. of sources
• Rotating logo with a controlled speed so it doesn't appear slow nor too fast
• Streaming text

All of this working together shows continuous change giving the illusion of a faster application.

What if all this didn't exist?

v2 - No micro-animations, but same breakdown of sources and content
v3 - No breakdown of sources (most common implementation of AI apps)
v4 - Everything mentioned in one step, no breakdown into sections
v5 - Only showing the list of sources, no reasoning

Each demo loads the output in 13 seconds, but each feels different in terms of the time:

v2 - You keep doubting if the application is still working or not
v3 - Wait feels never ending
v4 - Good to see the breakdown but the wait feels a little longer than original
v5 - "Okay how many sources is it going to go through?"

---

3. Reveal before you’re done

In Chronicle, they are quick to load the first slide, and then rest of slides load in their time. This is equivalent to "...typing" or basic skeleton pages that we saw earlier. You see the vision first, making you excited for the output, then wait patiently. IKEA uses a similar psychology to get you to buy furniture — they show you how it would look for your home — and once you can see it, you hold it dearly.

Many image generation products, like Midjourney and ChatGPT, load a rough image of the output and then slowly load rest of the image.

It's even better if at this stage you can personalize: with names, preferences, prior actions shown early. This gains the user's trust in the output, thus increasing the patience.

---

4. Reframing time

Instead of showing the loading state as a generic "thinking", you can add multiple layers of thinking like perplexity or use specific terms like:

• “Analyzing…”
• “Synthesizing insights…”
• “Personalizing your results…”

Or you can position it as fun lil thing.

Chronicle calls it "musings", to position it as fun raw thoughts. In fact, from my conversation with Chronicle's Head of Design, she said that it also comes from how our team talk to each other, we often say “here’s a musing” or “here’s my musings on how to approach this feature”.

Claude code does something similar: "Galvanising.."

There's a common saying from the design lords above — a 5-second wait feels shorter if it looks expensive.

---

5. Deploying an escape

Sometimes, it's just better to add an escape for the user. Maybe they decide to not wait any longer or they accidentally chose the "Intense Thinking Mode" when they just wanted a one word answer.

You don't know what the user's mindset it.
Maybe they prefer something now over perfect later. Many just want iterable answers instead of one thought out long answer.

ChatGPT's answer now is another example.

"Draft mode" in Midjourney output isn't going to be good, but its going to be fast.

So before taking the user into your rabbit hole, consider asking them? Were they expecting this? What if they've changed their mind?

---

6. Tiny cues, big impacts

A lot of the experiential magic lies in the tiny details, the micro-interactions.

Circling back to perplexity's animations, the brain treats responsiveness as speed. Even fake immediacy reduces frustration.

---

In the end, while it might seem that perceived performance is about tricking the users, it's actually about respecting how humans experience time.

Many of the best AI products don't compute faster, they just feel faster and intentional.

And if great design has taught us anything it's that difference lives in the parts no one explicitly asks for — but everyone remembers.