A/B testing of pricing pages: method, calculations and pitfalls to avoid

Updated on October 30, 2026

In September 2000, Amazon refunded 6,896 customers, an average of $3.10 each. Their offense? Buying a DVD during a five-day random price test, where discounts varied from 20% to 40% depending on the visitor. Internet users compared their prices, the press spoke of discrimination, and Jeff Bezos published a statement of apology regarding this random price test. Twenty-five years later, the pricing page remains the most profitable page to test and the most dangerous to test incorrectly.

Here's the method: what to test first, how to test the price without legal risk, which indicator to follow, how many visitors to expect.

Why the pricing page isn't like other pages

On a pricing page, the visitor doesn't form an opinion: they decide. It's the last page before the trial, the demo, or the payment.

Three mechanisms are at play: anchoring (the first price seen serves as a reference), loss aversion ("save 2 months" resonates more strongly than "17% off"), and the compromise effect (faced with three offers, many choose the middle one). Feedback from SaaS publishers reports effects of 8 to 15% on pricing pages, compared to 2 to 3% on a traditional landing page.

The downside is that poorly tested pricing can erode customer trust, lead to refunds, or even penalties. The testing method must be more rigorous than for a simple button test.

A/B testing of pricing pages: the short answer

A/B testing your pricing page involves comparing two versions of the page (structure, highlighting of an offer, default billing method, wording, or even price) with two randomly selected groups of visitors, and then retaining the one that generates the most revenue per visitor, and not simply the most clicks or registrations.

We first test the presentation, then the content of the offers, and finally the amount, on new visitors, with an honor rule announced in advance. If you are new to the concepts of variant, control group, or statistical significance, the glossary of A/B testing defines them in one line each.

What to test: the 4-level risk matrix

Rank each idea according to its likely impact on revenue and the risk to trust or compliance.

The practical rule: exhaust levels 1 and 2 before touching level 3, and only go down to level 4 with written legal advice.

Presentation tests to launch first

Level 1 tests launch in minutes, and the offer remains the same for everyone. Here are the most profitable ones, in order of priority.

The default monthly/annual switch

Displaying annual prices (adjusted to a monthly rate) by default lowers the initial price and increases the proportion of annual subscriptions. The downside: some customers feel trapped when they see the total amount billed all at once. Also, track refunds at 30 days.

The featured offer

The "most chosen" badge shifts choices towards the highlighted offer. When placed on the higher-tier offer, it often reduces subscriptions but increases the average order value: only revenue per visitor makes the difference.

The number of offers

Going from five offers to three reduces the comparison effort: level 2 if you merge tiers, level 1 if you only hide the "Enterprise" offer behind "Contact us".

The wording and promise of the button

“Try for 14 days, no credit card required” addresses the main objection, whereas “Sign up” doesn't address any of them. A no-code visual editor is enough for all these changes. To combine the badge and the annual toggle in the same test, first read our comparison between A/B testing and multivariate testing : the second one requires much more traffic.

Testing the price itself: three defensible methods

Testing an amount is possible and legitimate, provided that a method is chosen that creates neither lasting injustice nor opacity. Three approaches are viable.

1. Random testing on new visitors, with an honor rule

Only new visitors are included in the test; existing customers and logged-in visitors are excluded. Everyone keeps the same price for the entire duration of the test (persistent assignment). And you set in advance the rule that Amazon later adopted: any customer who paid the higher price is refunded the difference. Its cost is known from the outset.

2. Testing by market

Different prices across countries are common practice (purchasing power, currency, taxation). Testing a new price in a comparable market provides an indication, not proof, as no two markets are identical. Never go down to the city or postal code level: place of residence is a prohibited criterion for discrimination (see below).

3. The sequential test, before/after

You change the price for everyone on a given date and compare the following four weeks to the previous four. This is the most transparent method, but the least reliable: seasonality and marketing campaigns skew the comparison. It's suitable for a price increase you were planning to implement anyway.

An alternative without any customer being harmed: display a "Pro+" tier which does not yet exist, with a "Be notified" button, to gauge interest before collecting any payment.

The right indicator: revenue per visitor

The conversion rate alone is misleading: a lower price almost always converts better, but may generate less revenue. The primary metric should be revenue per visitor (RPV): revenue generated divided by the number of visitors exposed.

A numerical example. Version A: 1,000 visitors, 5% subscriptions at €40 per month, resulting in €2,000 in recurring monthly revenue and a RPV of €2.00. Version B, price reduced to €30: the rate climbs to 6% (+20%), but revenue falls to €1,800 and the RPV to €1.80. A conversion-focused dashboard declares B the winner; it costs 10% of revenue.

Add at least two safeguard indicators:

• Retention at 30 or 60 days : a low price can attract customers who leave at the first payment. A case study published by Statsig cites a test that increased sign-ups by 50% while doubling the churn rate.
• The offer mix : more subscriptions to the intermediate offer may mean less revenue.

Declare these objectives before launch: configure the payment event and its value in the module goals and tracking, then read the breakdown by offer in the test analysis dashboard.

How much traffic is really needed?

The required traffic depends primarily on the stage being measured: a test judged on trials requires approximately four times fewer visitors than a test judged on payments. Our calculations, with 95% confidence and 80% statistical power:

In practical terms: with 3,000 monthly visits to your pricing page, a presentation test judged on trial rates (25% effect) would conclude in approximately three months. A pricing test judged on payment rates with a 15% effect would take nearly three years: unrealistic.

Three levers for a low-traffic website:

• test outright changes (a tier removed, a price 30% higher) rather than €2 tweaks;
• measure a more frequent event (trial, click on "Choose this offer") while keeping payment as a safeguard;
• pooling traffic from pages that display the same pricing grid.

If you are following several tests at once, a tool like the Autopilot from BlastOptimize automates part of the monitoring and prevents you from leaving a test running that has become unnecessary.

Legal and ethical pitfalls: what the law says

Displaying different prices to different visitors is not prohibited per se in France, but it is regulated by consumer information, non-discrimination, and data protection laws. The following is a guideline, not legal advice: have any level 3 tests validated by your legal counsel.

The obligation to provide information in the case of personalized pricing

The European Directive 2019/2161, known as the "Omnibus" Directive, requires businesses to inform consumers when the price offered has been personalized based on automated decision-making. It was transposed into French law by an ordinance of December 22, 2021, applicable since May 28, 2022. Its recital 45 clarifies that simple dynamic pricing, which varies according to demand or time without individual profiling, is not covered.

An A/B price test is a gray area. A purely random allocation doesn't rely on any profile; but as soon as the price depends on a visitor-specific criterion (country, device, source, history), it approaches personalization. The prudent approach is to inform. A statement like "Our prices are being tested; the lowest price tested will be applied" costs little and mitigates most of the risk.

Prohibited criteria for discrimination

As the National Consumer Institute points out in its analysis of the challenges of price personalization, a price cannot vary according to a criterion protected by Article 225-1 of the Penal Code: origin, sex, age, family status, place of residence, state of health, etc. Segmentation by postal code falls directly under this text: hence level 4 is prohibited.

The sanctions

The Omnibus Directive raises the maximum fine for large-scale infringements affecting multiple Member States to at least 4% of annual turnover in the countries concerned. For an SME, the most immediate risk remains reputational: a screenshot comparing two prices circulates quickly.

Crossed-out prices: a trap specific to tests

For price reduction advertisements covered by Article L. 112-1-1 of the French Consumer Code, the crossed-out reference price must be the lowest price charged during the previous 30 days. If your variant B displayed €29 during the test, subsequently advertising "€39 instead of €49" becomes questionable. Archive the tested prices and their dates.

B2B, B2C: who is affected?

Consumer law protects consumers: a strictly B2B SaaS is less exposed. But a freelancer who subscribes often behaves like a consumer. Apply the same rules everywhere; it's simpler and more defensible.

Cookie consent

Assigning a variant relies on an identifier stored in the browser, which may fall under the prior consent required by the CNIL (French Data Protection Authority) depending on your configuration. A high refusal rate reduces usable traffic accordingly: factor this into the duration calculation.

Technical and statistical pitfalls

On a pricing grid, a technical defect does not go unnoticed: it creates suspicion.

The flashing of the old price

If the tool loads the variant after the original page, the visitor sees price A for a moment, then price B: visual proof that two prices coexist. A variant delivery with no flicker is a requirement here, not a nicety.

The inconsistency between the page and the payment

The variant displays €39, checkout charges €49. The tested price must carry through to the invoice: check the full journey of each variant before going live.

The visitor who changes groups

A prospect checks your prices on mobile on Monday and on a computer on Thursday: without an assignment tied to their account or email, they see two prices. The honor rule covers this case.

The premature stop

Checking the results daily and stopping as soon as one variant takes the lead is the most common mistake. This multiplies false winners, a phenomenon detailed in our article on false positives in A/B testing. Set the duration in advance, in full weeks, excluding promotional periods. See also our list of common mistakes in A/B testing.

The bias of visitors who are already convinced

Prospects who return after a demo already know the price: including them in a price test will skew the results. Exclude them. For an online store, product pages and the shopping cart also display the price: our page dedicated to A/B testing for e-commerce details these cases.

Checklist before launching the test

Twelve points to check; for a price test, a single missing item is enough to postpone it.

1. The test is classified in the matrix (level 1, 2 or 3; never 4 without legal advice).
2. A written hypothesis: "By displaying X, we expect Y because Z."
3. The main indicator is revenue per visitor; retention and offer mix are stated as safeguards.
4. The sample size and duration are calculated and noted before launch.
5. The duration covers full weeks and avoids promotional periods.
6. Existing customers and visitors already in the sales cycle are excluded from the amount tests.
7. The assignment persists from one visit to the next.
8. The tested price is transmitted until payment, confirmation, and invoicing.
9. No blinking: the variant is displayed before the first rendering.
10. The honor rule (lowest price applied) is written and validated internally.
11. The information notice is visible if the test involves the amount.
12. The tested prices and their dates are archived for the 30-day rule on crossed-out prices.

In terms of tools, compare what each solution offers in these areas before choosing: our comparisons BlastOptimize vs AB Tasty and BlastOptimize vs Kameleoon review display, targeting and revenue tracking. And for an overview of our own plans, see the BlastOptimize pricing page.

FAQ

Is it legal to display different prices to two visitors?

Yes, in principle, provided that the difference is not based on a discriminatory criterion (place of residence, age, origin, etc.) and that the consumer is informed when the price is personalized by an automated decision. A short, random test on new visitors, with an honor rule, is the most justifiable approach.

Is it better to test the price or the layout of the pricing table?

Start with the presentation: default annual switch, featured offer, number of tiers. These tests are quicker to complete and raise no fairness issues. Save the amount test for a more strategic decision.

What should be done with the customers who paid the highest price during the test?

Announce the rule before launch: at the end of the test, each customer who subscribed at the higher price is refunded the difference or switched to the selected price if it is lower. Amazon did this in 2000, but after the fact.

My site has little traffic: can I still test my pricing page?

Yes, by targeting clear changes and frequent events. With fewer than 5,000 monthly visits, an amount test judged on payments is rarely conclusive; a presentation test judged on trials is conclusive in one to three months.

How long should a test on a pricing page last?

A minimum of two full weeks, plus the time needed to reach the initially calculated sample size. With a free trial, extend the tracking to a full billing cycle to measure actual payments.

Should visitors be informed that a test is in progress?

For a presentation test, no: the offer is identical for everyone. For a price test, this is the most prudent practice with regard to the Omnibus Directive. A sentence below the pricing grid is sufficient.

Key takeaways

• Classify each idea in the matrix: presentation, packaging, price for new visitors, price based on profile. Exhaust the first two levels; never go to the fourth without legal advice.
• Judge by the revenue per visitor: a variant that converts 20% better may bring in 10% less.
• Calculate traffic before launching: approximately 4,400 visitors per variant for a 25% effect on trials, more than 49,000 for a 15% effect on payments.
• For a price test: random and persistent allocation, existing customers excluded, announced honor rule, information notice, archiving of tested prices.
• No display flicker tolerated: on a pricing grid, it reveals the test to the visitor.