Don’t guess what to change. Know what to test.

120 conversion experiments for product designers. Each one has an annotated before and after, a clear hypothesis, measurement guidance and sourced examples.

Get the library₹1,499Explore a free experiment
ControlWhat people see now
The button the treatment changes
TreatmentOne change
Secondary action demoted, not deleted
✳ 120 experiments written for the library✳ 120 illustrative before-and-afters✳ 141 public examples, 109 organizations✳ Honest about limits

You spotted the friction. Now what?

Signups are down. Can you improve this page?

IllustrativeA message most designers know.

A visible change is not yet an experiment.

Guesswork

Try a brighter button.

No reason. No evidence. No way to tell.

Experiment thinking

Visitors may face competing choices. Test the hierarchy, measure signups, watch sales enquiries.

  • Where is the struggle?
  • What is the evidence?
  • Which change fits?
  • What stays constant?
  • Which outcome should improve?
  • What could get worse?

More than a before and after. A brief you can defend.

Every experiment opens like this (PR-03, pricing): the change, the reasoning, what to check, how to measure it and what could go wrong.

PR-03PricingPlausible mechanism
Control
Treatment

IllustrativeFictional product, not GoodScool, not observed data.

  • One honest recommendationOnly the plan that fits most buyers.
  • Billing cycle unchangedPrices and plans are identical.
  • Revenue per visitorPlan share is only a diagnostic.
  • Mix can shift down tooWatch revenue, not just conversion.
The idea to rememberA design change becomes an experiment when it has a hypothesis, a primary metric and a guardrail.

Pricing-page visitors spread across the plans with no clear pattern, a large share pick the cheapest plan or leave, and replays show them hovering over all three plans before deciding.

With several plans and no guidance, a buyer must work out which one fits. A visible recommendation may act as a default: some buyers will accept it and others will use it as a reference point. If the recommendation is not honest, buyers who notice may lose trust.

Add one "Recommended" label to the plan that genuinely suits most buyers, leaving prices, features and layout unchanged.

Held constant

  • Prices, billing cycle and plan contents
  • Descriptor under each plan name
  • Card styling and button on every plan
  • Headline and call to action
  • Order of the plans

If the plan that fits most buyers carries an honest "Recommended" label, then first-period revenue per pricing-page visitor may increase because buyers who would stall or default to the cheapest plan get a reasoned starting point.

  • Customer data (for example retention and feature use by plan) shows which plan suits most buyers; the label is not just the highest-margin plan.
  • If you want to say "Most popular", that plan is currently the most chosen among paying customers and you can show it on request.
  • Tier mix, paid conversion, revenue per visitor and early downgrades are measured at baseline, because mix shift can move revenue in either direction.
  • Replays or sales notes show visitors stalling or choosing the cheapest plan out of caution.

Primary outcome
First-period revenue per pricing-page visitor
Numerator
Subscription revenue from first-period charges of assigned visitors, net of refunds issued within 30 days
Denominator
Visitors for whom the pricing page rendered with an assigned variant
Window
From first exposure to 30 days after purchase (matured for refunds); run whole weekly cycles

Diagnostics explain why the primary moved

  • Share of paid starts on the labeled plan
  • Tier mix among paid starts
  • Plan-card click rate by tier
  • Pricing-page exit rate

Guardrails must not get worse

  • Paid conversion rate per pricing-page visitor. Revenue per visitor can rise on fewer, larger purchases; total paid conversion must not fall beyond tolerance.
  • Refund and downgrade rate in the first 30 days. Steering buyers to a plan that does not fit may raise early downgrades and refunds.
  • Complaints or contacts citing a misleading plan label. An unearned recommendation is a trust and compliance risk.

Decision rule

Support
Treat as supported only if revenue per visitor rises by at least the pre-registered minimum effect under the pre-registered analysis (which handles heavy-tailed revenue), with paid conversion and refund guardrails holding and the sample-ratio check passed.
Hold
If revenue per visitor is inconclusive but the labeled plan's share rose, record that as a mix change without a revenue result and do not ship on mix alone.
Stop
Stop and investigate if paid conversion or refunds breach pre-registered tolerance, allocation is imbalanced, or the label becomes untrue.

What could get worse

  • The label can pull buyers down from a higher-value plan as easily as up from the cheapest; revenue can fall even when conversion rises.
  • Steering buyers to a pricier plan than they need may lift first invoices but raise early downgrades and refunds.
  • A label can make the other plans look inferior and add pressure that some buyers resent.

Avoid when

  • You cannot show that the labeled plan is the best fit for most buyers; choosing a favourite for margin is manipulation.
  • There are only two clearly different plans; a label adds little.
  • Paid purchase volume is too low to detect a revenue-per-visitor effect (a heavy-tailed metric); use buyer interviews instead.
  • The labeled plan's price or contents are about to change.
Plausible mechanism

No public example is linked to this experiment, and the record says so.

One small change. One hypothesis worth investigating.

The real LP-02 brief. Switch versions, then see the reasoning.

Sample, fictional product

Proposed testFictional product. No measured winner.

Plausible mechanism

A reason to test, not evidence of an effect. 3 public examples are linked: 2 reported helping, 1 mixed. See them.

Control: two buttons of equal weight

  • The button the treatment changesIt looks as important as the main button.
  • Headline unchangedSame in both versions.
  • Primary button unchangedSame label, style and position.

Seen both?

Treatment: one dominant action

  • Secondary action demoted, not deletedStill reachable, as a quiet line.
  • Headline unchangedSame in both versions.
  • Primary button unchangedSame label, style and position.

Seen both?

Why this is worth testing

If the secondary hero action is demoted from a button to a text line, then signup completion per exposed visitor may increase without a fall in sales-contact requests because one dominant action could reduce the effort of choosing at the decision point.
Primary metric
Signup completion rate per exposed visitor
Guardrail
Sales-contact requests per exposed visitor
Also check
Accessibility audit of the demoted action (keyboard reach, text contrast)

A second button becomes a quiet text line.

Headline, primary button, destinations, layout and traffic.

Signups per exposed visitor rise by the minimum effect set before launch, with sales requests inside tolerance.

A higher share of clicks on the primary button alone.

Get all 120 experiments

Start with the problem. Find the experiment.

Describe what you are seeing. The library finds matching experiments. It searches its own text, not your site.

People abandon checkout when the final cost appears
Try another:
CK-06Checkout

Show shipping and the all-in total at the cart, not at the end

Shoppers see shipping and fees only at the last step and may leave when the total jumps.

Effort MediumRisk MediumLiterature-informed2 public examples
CK-12Checkout

Make the final order summary itemized and complete

At the last step shoppers may hesitate when they can't quickly verify what they're buying and when it arrives.

Effort MediumRisk LowPlausible mechanismNo public example
CK-02Checkout

Show real step progress in a multi-step checkout

In a multi-step checkout, shoppers can't tell how many steps remain and may quit midway.

Effort LowRisk LowPlausible mechanismNo public example
CK-07Checkout

State the real returns terms beside the order total

First-time buyers may hesitate to pay because they are unsure they can return the order.

Effort LowRisk MediumLiterature-informedNo public example

Good experiments don’t always win. Good designers learn either way.

141 public examples from 109 organizations, including the ones that did not work.

How the 141 examples came outDated 2007 to 2026, sources checked 2026-10-12. Not a win rate: companies tend to publish wins.
  • Reported helping85
  • Mixed31
  • No detectable effect10
  • Reported hurting10
  • Not stated5

The Guardian

Reported as helping

Dropped a second button on a first screen

80%reported uplift in registrations and sign-ins

One team, one test. A reason to try it, not a number to expect.

Caveats
  • Self-reported by the agency
  • Several changes at once
  • Sample not stated

Media, around 2022. Source grade B, vendor case study.

ConversionLinked to LP-02

Focus on the Family

Reported a mixed result

Replaced a carousel with one static hero

1.9% to 3.9%click rate. Donations flat: 239 against 232, no significant difference.

The metric you choose decides whether it looks like a win.

Caveats
  • Run by a testing firm for its client
  • Mobile not reported separately
  • No revenue gain shown

Nonprofit, 7 May to 1 June 2026, 40,756 visitors. Source grade B.

NextAfterLinked to LP-02

StubHub

Reported as hurting

Showed all-in prices from the first listing

Almost 21%more revenue when fees appeared only at checkout. Not advice to hide costs.

When a cost appears is a trade-off, not a free win.

Caveats
  • Randomization glitches, users excluded
  • Co-author financial tie disclosed
  • Listings, not a cart

Ticket marketplace, August 2015, several million users. Source grade A, peer-reviewed.

National Bureau of Economic Research (Marketing Science, 2021)Linked to CK-06

Reviewed by an AI editor. No human expert has reviewed these records yet.

75 of 120 experiments have a linked example. The rest say so.

Evidence labels are not proof.

Get the library. Here is exactly what’s included.

  • 120 structured conversion experiment briefs
  • 120 illustrative before-and-after comparisons
  • A clear hypothesis and the evidence to check first, for every experiment
  • Primary metrics, diagnostics and guardrails
  • Trade-offs, and the situations where an experiment may backfire
  • Access to the public example catalog: 141 sourced records with caveats
  • A searchable, filterable library that runs in the browser
  • Briefs you can copy as Markdown, download or print

What it does not do

  • It does not run A/B tests or connect to your website.
  • It does not calculate live uplift or predict winners.
  • It does not guarantee that any change will improve conversion.

Fair questions. Straight answers.

Product designers who own conversion-sensitive journeys. CRO practitioners can use it as a reference of design interventions. It is not a statistics course.

No, but be comfortable with metrics. Every brief names the metric, guardrails and events, so you can hand it to an analyst.

No. It helps you identify, frame and plan an experiment. You run it with your own tools.

No. They are illustrative screens of fictional products, labelled as such.

No. Each is a hypothesis to test in your context, and every record shows an evidence label. Some changes could lower conversion, so every brief lists trade-offs.

141 public records from 109 organizations, each with its source, what was reported and what is uncertain. 75 of 120 experiments have one.

Yes: copy as Markdown, download or print.

No. It runs in your browser through your GLE account and stores nothing. Its content was written and checked by an AI editing process, with no human expert review yet.

Stop collecting CRO tricks. Start framing better tests.

120 experiments, 141 sourced examples and an honest account of what they cannot tell you.

Get the library₹1,499Try the free experiment