Start High and Drift Down, or Just Test It?
You're launching something. The product is ready, the page is written, and one question is still open: what do we charge?
There are two respectable answers, and they are not interchangeable.
Approach one: start high, move down
Launch at a premium price. Sell to the people who want it most and will pay most. Then, over months, come down — through official price cuts, or promotions, or cheaper tiers — and pick up progressively more price-sensitive buyers on the way.
This is price skimming, and it's the standard playbook for new consumer electronics, new drugs coming off patent protection, and hardcover-then-paperback publishing. It works because a demand curve is a queue: the people at the front will pay a lot, and if you open at a low price you have sold to them for less than they'd have given you.
What it buys you. High margins early, when you most need cash. A price that signals quality before anyone has formed an opinion. And a genuine anchor: a later discount reads as a deal because there's a number to compare it against.
What it costs you. Volume early, which matters enormously if your product has network effects and not at all if it doesn't. Brand damage if the discounting becomes a pattern people learn to wait for. And you learn surprisingly little — you find out that some people will pay your opening price, which is not the same as knowing what the curve looks like.
Approach two: test it
Show price A to half your traffic and price B to the other half. Measure which makes more money. Keep the winner.
What it buys you. An actual measurement rather than a theory. Low commitment — you can find out a $49 price outperforms a $39 one without ever announcing a price change. And it generalises: the same machinery tests packaging, copy, and feature bundles.
What it costs you. Traffic, mostly. And time, which people consistently underestimate.
The part everyone underestimates
Here is the thing that decides whether testing is even available to you.
How long before you can actually call it
A simulated test of two prices. The band is the 95% confidence interval on the difference in conversion rate — you have a result when the whole band clears zero.
version A
the effect that genuinely exists
split evenly between A and B
Visitors needed per variant
17,920
5% significance, 80% power
Days to a callable result
59.7
that is a long time to hold a price
This run first cleared zero on day
12
and slipped back — only safe from day 30
Interval on day 108
0.57 pts
± 0.32 pts
- Observed difference, B − A
- 95% confidence interval
- The true lift
Drag the traffic down and the band stays fat for weeks — with a small audience the honest answer is “we cannot tell yet”, and any number you read off it before then is noise. Drag the true lift down and watch the same thing happen: the smaller the real effect, the more evidence it takes to see it. Notice that this run's band first cleared zero on day 12 and then slipped back, only settling from day 30. A team checking daily would have called it on day 12 and been lucky rather than right.
Show the numbers
| Day | Visitors per variant | Observed difference | 95% interval |
|---|---|---|---|
| 11 | 3,300 | 0.91 pts | -0.07 pts to 1.89 pts |
| 22 | 6,600 | 0.61 pts | -0.09 pts to 1.30 pts |
| 33 | 9,900 | 0.78 pts | 0.21 pts to 1.35 pts |
| 44 | 13,200 | 0.78 pts | 0.29 pts to 1.27 pts |
| 55 | 16,500 | 0.87 pts | 0.43 pts to 1.31 pts |
| 66 | 19,800 | 0.85 pts | 0.45 pts to 1.26 pts |
| 77 | 23,100 | 0.78 pts | 0.40 pts to 1.16 pts |
| 88 | 26,400 | 0.75 pts | 0.40 pts to 1.10 pts |
| 99 | 29,700 | 0.60 pts | 0.27 pts to 0.93 pts |
Set the baseline to 4%, the true lift to 15%, and the traffic to a few hundred visitors a day, and the honest answer is that you'll be waiting weeks. Not because the tooling is slow — because the effect is small relative to the noise, and separating them takes the number of observations it takes.
Two things fall out of that, and they're the difference between a test and a ritual.
Small effects need enormous samples. Halving the effect you're looking for roughly quadruples the sample you need. A 30% lift is findable on modest traffic. A 3% lift, on the same traffic, is not findable at all in any timeframe you care about — and the number your dashboard shows you on day four is noise wearing a decimal point.
Stopping when it looks good is how you get fooled. If you check every day and stop the moment the result crosses into significance, you will "find" effects that aren't there — because with enough looks, the interval wanders across zero by chance eventually. Pick the sample size before you start, and let it run.
Choosing between them
Skim when you're launching something genuinely new, you have no traffic to test with, the product carries a quality signal, and your buyers arrive in waves rather than a steady stream. You are not really choosing skimming over testing here — you're acknowledging that testing isn't available yet.
Test when you have volume, the decision is reversible, and you can wait out the sample. E-commerce, SaaS signup flows, anything with a checkout and real daily traffic.
And when you have neither traffic nor a story, both of these are the wrong question. Go and talk to twenty customers, look at what competitors charge, and price off your read of the value you deliver. That's not a cop-out — value-based pricing from qualitative evidence beats a statistically meaningless test every time.
The version most teams should run
The two approaches stack, and the sequencing is the useful bit:
- Launch high, on judgement, to the people who want it most. You get margin and you get an anchor.
- Use that period to build the traffic that makes testing possible at all.
- Then test the things around the price — bundles, tiers, trial lengths, annual-versus-monthly. These often have bigger effects than the headline number and are far less loaded.
- Come down deliberately, with a reason attached. A cheaper tier, a student plan, a regional price. Not a sale that teaches people to wait.
The mistake isn't picking the wrong one of these. It's running a test you don't have the traffic to conclude, and then acting on it with total confidence.