Price testing: the complete guide for software founders (2026)

Price testing for software founders: why revenue per visitor beats conversion rate, four methods compared, how to test safely, and a 10-step checklist.

Short answer

Price testing shows different visitors different prices for the same product and measures which price earns the most. Judge each price by revenue per visitor: $49 bought by 2.2% of visitors earns $1.08 per visitor and beats $29 bought by 3% ($0.87). Keep your current price as the control, assign prices on the server and charge exactly what you show.

  • Revenue per visitor is price times the share of visitors who buy, so $49 at 2.2% beats $29 at 3% by 23.9%.
  • Moving from $29 to $49 earns more unless it costs you more than 40.8% of your buyers.
  • A fixed 50/50 test needs about 13,800 visitors per price to confirm that 23.9% gap at 95% confidence and 80% power.
  • Before/after comparisons mix the price change with launches and seasons; a split test runs every price at the same time.
  • Keep crawlers, link previews and existing customers on your standard price.

If 3% of your visitors buy at $29 and 2.2% buy at $49, the higher price earns 24% more from the same traffic, even though it loses more than a quarter of your buyers. No spreadsheet can tell you which of those worlds you live in. A price test finds out with your own visitors and your own checkout.

The stakes add up quickly. At 10,000 pricing-page visitors a month, the gap between $0.87 and $1.08 per visitor is $2,080 every month, before a single new customer arrives.

What is price testing?

Price testing is a controlled experiment: visitors are randomly assigned one of two or more prices for the same product, at the same time, and you compare how much money each price brings in per visitor. The product, the page and the checkout stay the same. Only the number changes.

Random assignment is what makes it a test. A visitor's price doesn't depend on who they are, where they come from or which device they use, so the groups differ only in the price they saw. Running the prices at the same time cancels out launches, holidays and ad campaigns, because those hit every group equally. And each visitor keeps one price for the whole test, so nobody sees $29 on Monday and $39 on Tuesday.

That randomness also separates a price test from personalized pricing, where a seller picks a price for a specific person based on data about them. A test exists to find one price to charge everyone afterwards.

Which metric should a price test optimize?

Revenue per visitor. It is the money a price group brought in, divided by the number of visitors who saw that price, which for a single purchase works out to price × conversion rate. Conversion rate on its own almost always favors the cheapest price, because a lower price usually sells to more people.

Example: two prices, 10,000 visitors each.

$29$49
Visitors who saw the price10,00010,000
Conversion rate3.0%2.2%
Buyers300220
Revenue$8,700$10,780
Revenue per visitor$0.87$1.08

The $49 price loses 26.7% of the buyers and still earns 23.9% more: $2,080 extra for every 10,000 visitors. A test judged on conversion rate would have crowned $29.

A quick sanity check before any test: moving from an old price to a new one earns more as long as you lose fewer than 1 − old ÷ new of your buyers. From $29 to $49 that is 1 − 29/49 = 40.8%. The price increase calculator runs this for any pair of prices.

Illustration: conversion falls as the price goes up, but revenue per visitor peaks in the middle. A price test looks for that peak.

Before you trust the raw number, adjust it for fees, refunds and, if you sell subscriptions, retention.

Fees take a bigger bite from cheap prices. At Stripe's US rate of 2.9% + 30¢ for domestic cards (Stripe pricing, as of November 2026), $29 nets $27.86 and $49 nets $47.28, so the gap in the example widens from 23.9% to 24.4% once fees are counted. If you sell a product with real unit costs, compare profit per visitor instead.

Refunds belong to the price that caused them. A price that attracts impulse buyers who ask for their money back a week later looks better than it is until the refunds land, so subtract them from the group that earned the sale.

For subscriptions, the first payment is a proxy. Before you call a winner permanent, compare cancellations and upgrades by price group over the next billing cycles. A higher price that loses customers faster can still lose on lifetime value.

What are the four ways to test a price?

There are four: before/after testing, a fixed A/B split, a multi-armed bandit and a willingness-to-pay survey. The first three measure real purchases; the survey measures what people say they would pay. They answer different questions, so the right one depends on your traffic and on what you need to know.

MethodHow it worksTraffic neededMain weaknessBest for
Before/afterEveryone sees price A, then everyone sees price BAnySeasons, launches and ads change between periodsProducts where two prices can't run at once
Fixed A/B splitRandom split held constant until a sample size set in advanceHighMany visitors see the worse price until the endA precise estimate of the difference
Multi-armed banditStarts even, then shifts visitors toward the price earning moreMediumRougher estimate of how bad the losers areEarning while you test three or more prices
Survey (Van Westendorp)Ask people which prices feel cheap or too expensiveNoneStated willingness to pay runs highPicking the range before a live test

Before/after (sequential) testing

You change the price for everyone on a date and compare the period before with the period after. It needs no tooling and works at any traffic level. Its weakness is that everything else changes too.

Example: you raise Pro from $29 to $39 on March 1. February brought 12,000 visitors and 360 sales, or $0.87 per visitor. March brought 15,000 visitors and 330 sales, or $0.858 per visitor, which reads as a 1.4% drop. But March also had a launch that sent 4,000 low-intent visitors, of whom 20 bought. Without them, March's regular traffic converted at 2.82% and earned $1.10 per visitor, a 26.3% gain over February. The same data supports opposite conclusions, depending on a traffic mix you may not even see.

If before/after is your only option, alternate the periods: run A, then B, then A again, with equal lengths, the same weekdays and no launches or promotions in any of them. If the second A period matches the first, the B result is more believable.

Fixed A/B split testing

A fixed split assigns visitors at random, usually 50/50, and keeps the split constant until you reach a sample size you chose before starting. It gives the cleanest estimate of how much better one price is, and it is the textbook A/B test.

The costs are real. Half of your visitors keep seeing the worse price until the test ends. And you can only read the result once: checking every day and stopping at the first good-looking number inflates false winners. In Evan Miller's simulation, a test checked after every observation (up to 150) and stopped at the first significant result declared a winner 26.1% of the time when the change did nothing, more than five times the 5% error rate it claimed.

For the $29 versus $49 example above, a fixed split needs about 13,800 visitors per price to confirm the 23.9% gap at 95% confidence and 80% power. The formula is in the section on test length below, and bandit vs A/B test compares the two designs in depth.

Multi-armed bandit testing

A bandit starts by splitting traffic evenly, then keeps moving visitors toward the price that is earning more, while still sending some traffic to every price in case early numbers mislead. One well-studied way to do this, Thompson sampling, goes back to a 1933 paper by William R. Thompson: show each price in proportion to the probability that it is the best one.

Simulation: we ran the production Leverprice engine against three prices ($19, $29 and $49), with $29 converting 2% of visitors and $49 earning 20% more per visitor. At 300 sales a month it reached a decision in a median of 23 days and picked $49 in 28 of 30 runs. A fixed three-way split sized for 95% confidence and 80% power needs 176 days on the same traffic. Across all 960 simulated runs (four effect sizes, eight traffic levels, 30 runs each), the engine never called the worst price the winner.

Simulation of the production engine at 300 sales a month, with $49 earning 10% more per visitor than $29: by day 40 it sends 89% of visitors to $49.

The trade-off is precision about the losers. Because the losing prices get less traffic, a bandit tells you less about exactly how much worse they are. If you need that number for a forecast or a board deck, use a fixed split.

This is the method Leverprice uses: a Thompson-sampling bandit scored on revenue per visitor that recomputes traffic weights every few minutes, keeps at least 5% of visitors on every price, and calls a winner when one price is at least 95% likely to be the best.

Surveys: Van Westendorp and other willingness-to-pay questions

A survey asks people about prices instead of charging them. The best-known format, the price sensitivity meter that Dutch economist Peter van Westendorp introduced in 1976 (background), asks four questions: at what price would this product be so cheap you'd doubt its quality, a bargain, getting expensive, and too expensive to consider. Plotting the answers gives a range of acceptable prices.

It needs no traffic and takes days, which makes it the right tool before launch. The catch is that people overstate what they would pay: a meta-analysis of 77 studies found that hypothetical willingness to pay was on average 21% higher than real willingness to pay. Use the survey to pick the range, then test inside it with real checkouts. The Van Westendorp calculator turns survey answers into the curves.

Which method should you use?

Match the method to your traffic and to the question you need answered.

Your situationUse
No product or almost no traffic yetA survey to choose the range
Fewer than about 50 sales a monthOne bold change at a time, with a long bandit or alternating before/after periods
You need a precise estimate of the differenceA fixed split with a sample size set in advance
You want to test three or more prices and lose little revenue doing itA bandit
You sell one-off contracts or can't show two prices at onceAlternating before/after periods

The 50-sales line comes from the same simulation: at 30 sales a month, even a price earning 20% more per visitor took a median of 230 days to call, and 12 of 30 runs had no answer within a year. Price testing with low traffic covers what to do below that line.

What should you test first?

Start with the price of your most popular plan, with wide gaps between the test prices. Then move on to the annual price, the gaps between tiers and discount depth. Change one of these per test.

What to testExample pricesMeasureWatch for
Price of one plan$29, $39 and $49 per monthRevenue per visitor on that planSteps under 10% need enormous samples
Annual price$290 vs $249 a year, monthly fixed at $29Revenue per visitor over 12 monthsUpfront annual cash flatters the short term
Tier pricesPro at $29 vs $39, Starter and Team fixedRevenue per visitor across all plansBuyers moving to the plan next door
Discount depth20% vs 40% off the first yearRevenue per visitor after the discountTeaching customers to wait for sales

Narrow tests are expensive. Telling $29 at 3% apart from $35 at 2.6%, a 4.6% difference in revenue per visitor, takes about 272,000 visitors per price at 95% confidence and 80% power. A wide test settles faster and teaches you more about the shape of demand, and you can narrow the range in the next round.

For annual plans, count the full year: a $290 annual purchase is worth 12 months of a $29 plan minus the discount, and judging it on day-one cash makes any annual price look great. Annual vs monthly discount covers how to size the yearly discount before you test it.

For tier prices, measure revenue per visitor across the whole pricing page. Raising Pro pushes some buyers down to Starter and some up to Team, and a Pro-only metric misses both moves.

What not to test at the same time

  • Price and anything else on the pricing page. If you change copy, layout or plan features in the same test, a change in revenue per visitor has no single cause.
  • Two experiments on the same visitors. A page redesign test running alongside a price test splits each price group again and muddies both.
  • Prices for existing customers. Test on new visitors only; a change for current subscribers is a separate decision with its own notice period.
  • Changes to the range mid-test. A price added halfway through saw different weeks than the others. Restart instead.

How do you run a price test without losing customers' trust?

Treat every visitor as someone who might compare notes with a friend. These seven rules cover the failures customers actually notice:

  1. Keep your current price as the control. It is the baseline every other price is judged against, and it means the worst case of the test is the business you already have.
  2. Assign each visitor once and keep it. A visitor who sees $29, reloads and sees $39 stops trusting both numbers. Store the assignment in a signed first-party cookie and key signed-in users by their account.
  3. Decide the price on the server. A price chosen by a script in the browser flickers, disappears when an ad blocker removes the script, and can be edited before checkout. Server-side price testing covers the mechanics.
  4. Show crawlers and link previews your standard price. Googlebot generally doesn't support cookies (Google's testing guide), so an enrolled crawler could see a different price on every visit, and one Slack preview of a test price reaches a whole channel. If your current price is in the test, crawlers see exactly what your control group sees.
  5. Charge exactly the price shown. Compute the checkout amount on the server from the same assignment that rendered the page, and never read the amount from the browser.
  6. Decide your refund policy before launch. After a random DVD price test in 2000, Amazon refunded 6,896 customers an average of $3.10 and promised that future test buyers would automatically get the lowest test price (Amazon's statement). Your version can be simpler: anyone who asks during the test gets the difference back.
  7. Leave existing customers alone. Run the test on new visitors and keep current subscribers on the price they signed up for. How to raise SaaS prices covers changing prices for the people already paying you.

How long does a price test take?

Long enough for each price to collect the visitors needed to detect the gap you care about, and never less than a full week, so weekday and weekend buyers are both in the sample. Small differences need far more traffic than big ones.

For a fixed split, the visitors you need per price are roughly:

n = (1.96 + 0.84)² × (σ²A + σ²B) ÷ (RPV_B − RPV_A)²

where 1.96 and 0.84 set 95% confidence and 80% power, and σ² = price² × c × (1 − c) for a price with conversion rate c (each visitor pays either nothing or the price). For $29 at 3% and $49 at 2.2%, that is 7.85 × (24.47 + 51.66) ÷ 0.208² ≈ 13,800 visitors per price. At 500 pricing-page visitors a day split two ways, the test takes 55 days.

A bandit usually reaches a decision sooner, because it concentrates traffic on the prices still in contention and stops once picking the leader would cost almost nothing. Simulation: the same engine as above, three prices, with sales counted at the $29 control.

Sales a monthBest price ahead byBandit: median days to decideFixed three-way split: days needed
10010%1131,989
10020%68527
30010%40663
30020%23176
1,00010%14199
1,00020%853

The fixed split is sized to detect the gap between the best price and the control at 95% confidence and 80% power, with three prices sharing traffic. The bandit figure is the median of 30 simulated runs; at 100 sales a month and a 10% gap, 4 of the 30 runs had not decided within a year.

How long should a price test run? breaks these results down further, and the sample size calculator gives the number for your own traffic, prices and conversion rate.

Does price testing hurt SEO?

Not if the test is temporary and crawlers see a price your real visitors also see. Google's guide to website testing asks you not to cloak test pages (show Googlebot one set of URLs and humans another), to use temporary 302 redirects rather than 301s if a test redirects between URLs, and to point any variant URLs at the original with rel=canonical. It also warns that a site running an experiment for an unnecessarily long time may be treated as trying to deceive search engines (Google).

Two checks are specific to prices. Keep any price in your structured data equal to the price on the page, because Google requires structured data to truly represent the page's content (structured data policies). And keep test prices out of page titles and meta descriptions, which search results can keep showing after the test ends. Does A/B testing prices hurt SEO? covers cloaking, snippets and markup in depth.

The rules written so far focus on prices set from personal data and on disclosing them. In the EU, the Consumer Rights Directive, as amended by Directive (EU) 2019/2161, requires sellers to tell consumers, where applicable, that a price was personalised on the basis of automated decision-making (Directive 2019/2161, Article 4).

New York's Algorithmic Pricing Disclosure Act, in force since November 10, 2025 (Kelley Drye), requires the line "THIS PRICE WAS SET BY AN ALGORITHM USING YOUR PERSONAL DATA" when a price is set by an algorithm that uses personal data, which the law defines as data that identifies or can reasonably be linked to a specific consumer or device (N.Y. Gen. Bus. Law § 349-a).

A random test that ignores who the visitor is differs from a price computed from someone's data, but whether a disclosure rule reaches a test that keys assignment to a cookie or an account is a question for a lawyer in your market. This is not legal advice. Is price testing legal? walks through the rules with sources.

Which tools can run a price test?

Four kinds of tools can run one, and they differ mostly in where the price is decided and where the revenue numbers come from.

Tool typeExamplesWhere the price is decidedWhere revenue comes from
Experimentation platformsOptimizely, GrowthBook, PostHogA flag plus your own pricing codeEvents you send, or your data warehouse
Stripe price-testing tools1Price, PricingFastA script on your pricing pageStripe
App paywall platformsRevenueCat, Superwall, AdaptyThe paywall or offeringApp store and billing transactions
Server-side price enginesLeverpriceYour server, through an SDKSigned webhooks from Stripe, Polar, Lemon Squeezy or Dodo Payments

If you already run GrowthBook or PostHog, a flag plus your own checkout code is a reasonable way to start: you build the assignment, the checkout mapping and the revenue pipeline yourself. How to A/B test prices shows both that DIY version and the SDK version step by step, and price testing software compares the options in detail.

The 10-step price testing checklist

  1. Write down your current price, conversion rate and revenue per visitor for the last 30 days.
  2. Pick one thing to test: one plan's price, the annual price, a tier gap or a discount.
  3. Choose one or two test prices around the control, at least 20 to 30% apart, inside the range a survey or the market supports.
  4. Estimate the sample size and duration. If the test would run longer than about three months, test wider prices or fewer of them.
  5. Choose the design: a fixed split for a precise estimate, a bandit to earn while you test.
  6. Assign prices on the server, store each assignment in a signed cookie, and key signed-in users by account.
  7. Give crawlers, link previews and existing customers your standard price.
  8. Charge exactly the price shown, and count purchases from signed payment-provider webhooks, net of refunds.
  9. Write the support reply and refund policy for "I saw a different price" before launch.
  10. Stop at the planned sample size or decision rule, make the winner the price for new visitors, update structured data and ads, and plan the next test around the new price.

Frequently asked questions

Is price testing the same as price discrimination?

Not in the usual sense. Price discrimination means charging different customers different prices on purpose, as a lasting strategy, based on who they are or what they will pay. A price test assigns prices at random for a limited time, to learn which single price to charge everyone afterwards.

How many prices should I test at once?

Two or three with a fixed split, up to four or five with a bandit. Every extra price divides your traffic further, so add prices only when you have the visitors to support them.

Should I test prices on existing customers?

No. Test on new visitors and keep current subscribers on the price they signed up for. Changing what existing customers pay is a price increase that deserves advance notice and a plan of its own.

What if a customer finds out someone else paid less?

Tell them the truth: you test prices on new visitors, prices are assigned at random, and you'll refund the difference if they paid more. Decide that policy before the test starts so support doesn't have to improvise.

Do I need a developer to run a price test?

For a server-side test, yes, but only for a few lines: one call where the price is shown and one where the checkout is created. Sites without a backend can route the Buy button through an assignment link that redirects to a checkout per price, at the cost of not showing the tested price on the page itself.

When should I run my next price test?

Once the winner has been live for a full billing cycle and refunds and cancellations look normal. Then test a narrower range around the new price, or move on to the annual price or the gaps between tiers.