The Invisible Meter
Alebuild · Ride-Along Sample
Perplexity

The InvisibleMeter

Why the paywall isn’t what lost them.

Duration: 7 minRevenuePsychologyAI products
The finding, in one line

I spent a currency I couldn’t see, on work that never arrived.

Twenty-four minutes, one real research job, two surfaces and a paywall. What follows is what happened, in order, on the clock.

The route

Rode as
A solo design-engineering studio owner, sizing up which AI startups to pitch this quarter. Has been googling it. Was told Perplexity is better. Has twenty minutes. The research was real — a demo query never produces impatience, and impatience is where findings come from.
Setup
Chrome 1280×832, macOS, fresh profile, logged out, Comet not installed. Google SSO auto-authenticated me on arrival — cold by intent, not by identity. Noted because it changes what “logged out” means here, and because the silent sign-in is part of beat 01.
Route
google → home → sign-in → search → results → 2 follow-ups → 3 Computer agents → free limit → pricing → find Comet → download → install → onboarding → attempt to resume
Timing
Leg A 12m · Leg B 4m · Leg C 6m · Leg D 9m
Not covered
Spaces, Discover, mobile, the API, Enterprise, long-term logged-in state.
The numbers
I have no access to Perplexity’s analytics. Every figure here models user motivation — it is not a claim about their funnel. Where I’ve assumed something, I’ve said so.
Conditions
Single session, single run. Perplexity is non-deterministic, so nothing here reviews answer quality — that won’t reproduce. This reviews the experience of getting one, which will.

Where do you think I lost interest?

Empty
100Intent0
Your guess
The wall of text · beat 03
Click or drag to place your marker
The journey · beat 01Beat 12

Most people put it at the paywall. Have a go before you scroll — it costs nothing and it’s more interesting if you commit.

The Intent Curve

Motivation is the thing you already paid to acquire. It’s finite, and every screen either spends it or refills it. Here is twenty-four minutes of it.

Peak 92. End 4. Four climbs, four falls, terminating on the floor. Both rises are worth pausing on: at beat 06 intent goes up because I thought I was getting something free, and at beat 11 because the Comet intro screen is genuinely lovely. Both were spent inside sixty seconds. A rise that isn’t converted is just a further fall arranged in advance.

The scoreboard

The film runs in journey order, because that’s how your user meets the product. This runs in severity order, because that’s how you’ll schedule it.

SeverityQuick winMedium
Critical
  • 09 · Credits documentation
  • 12 · Refund + explain failure
  • 03 · Answer density
  • 06 · Credit meter
Major
  • 04 · Define “Computer”
  • 05 · “Free preview” framing
  • 07 · Limit visibility
  • 10 · Comet placement
  • 12 · Parallel-session cap
Minor
  • 01 · Model picker order
  • 11 · Onboarding ask count
  • —

Ship these three on Monday

All three are cheap, all three are indefensible once named, and none needs a designer.

1

Give credits a unit and a range

“A typical research task costs 3–8 credits; your plan includes 50.” Ranges are fine. Refusing to estimate is not — right now the resource is formally unpriceable, so nobody can decide whether $20/month is worth it. One afternoon of writing. Beat 09.

2

Refund credits on failure, and say why

The highest-value fix here, because it converts the worst moment in the product into a recoverable one. A user who loses a resource silently is gone. Beat 12.

3

One sentence defining “Computer”

A copy change that unlocks the most heavily promoted feature in the product — the one I refused to click because I couldn’t tell what it was. Beat 04.

Where I’d take this next

Perplexity has already solved this problem once and hasn’t noticed. At 13:00 I watched an agent narrate its own work and said the word “nice” out loud — that’s the Labor Illusion, applied correctly, buying patience during a slow operation. Two minutes later I was reading circular documentation trying to find out what any of it cost. The product knows how to show its work. It just doesn’t show its price. Nearly every finding in the middle of this teardown is a version of that one gap.

The second thread isn’t really about Perplexity. On generative products, nobody owns the output. Length, density, failure states, cost variability — all of it is emitted by a model rather than decided by a person, so none of it gets designed. There was no meeting where someone chose to return three screens of prose. There was no spec for what happens when an agent fails. Those are the highest-leverage surfaces in the product and they are, structurally, unowned.

That’s the defining design problem of this product generation, and it’s why an AI company needs a design engineer rather than a designer. Half the fixes above are prompt-and-render decisions that live in a codebase, not in Figma. You can’t hand them off as a mockup, because the thing that needs designing is a constraint on a model, not a screen.

Two ways to work together

Option 1

Focused sprint

2–3 weeks · from $8K

We ship the top three fixes together, in your codebase, plus a short retrospective on what moved.

Best if you want the worst problems gone, fast.
Option 2

Embedded design partner

Monthly · 3-month minimum

I embed with your team two days a week. We ship continuously, and every quarter I run a fresh ride-along so you can watch the curve change.

Best if design is a compounding advantage, not a project.

Want one of these for your product?

Four per month. Free. Delivered in 48 hours.

Request your ride-along →