Landing Page Optimization: A Repeatable Testing Method

Landing page optimization is the work of raising one page’s conversion rate with a repeatable loop: diagnose where visitors drop with a page funnel, prioritize the hypotheses, size each test against the traffic you actually have, run it, and read the result honestly. It is not a list of tips. The single most useful fact about it is that your traffic budget, not your list of ideas, decides which changes you can afford to test. This guide is part of our conversion rate optimization guide and walks through the method end to end with one illustrative lead generation page, real calculator outputs and a value estimate you can defend in a budget meeting.

Many guides on this topic hand you a long list of things to change. The problem is not a shortage of ideas. It is that a page with 6,000 visitors a week can run maybe a dozen honest tests a year, and a list with no order and no sizing will spend that budget on changes too small to ever be measured. The sections below follow the order in which the decisions actually happen.

What landing page optimization is, and what it is not

A landing page is the page a visitor lands on after clicking an ad, an email or a link, built around one main action: submit a form, start a trial, buy a product, book a call. Landing page optimization (LPO) is conversion rate optimization narrowed to that single page and that single action.

That narrowing matters in three ways:

• Against the CRO pillar. CRO covers the whole journey: navigation, pricing, checkout, onboarding, retention emails. LPO takes one page and asks a sharper question: of the people who arrive here, how many complete the action, and where do the others leave?
• Against SEO. SEO is about getting visitors to a page; LPO is about what happens once they are on it. Many paid landing pages are not even meant to rank. The two meet on speed and relevance, and paid search makes the overlap explicit: in Google Ads, Quality Score is calculated from three components, and one of them is landing page experience, defined as how relevant and useful your landing page is to people who click your ad.
• Macro and micro conversions. The macro conversion is the action the page exists for, here a submitted form. Micro conversions are the steps on the way: scrolling past the first screen, clicking the call to action, starting the form. You optimize the macro conversion and you diagnose with the micro ones. A test that lifts CTA clicks while submissions stay flat has not improved the page.

The loop has five steps and they always run in the same order: diagnose, prioritize, size, test, read. Skip the diagnosis and you test at random. Skip the sizing and you run tests that cannot finish. Skip the honest reading and you ship noise and call it a win.

Diagnose with the page funnel

Start with the micro conversions. For the rest of this guide we use one illustrative page: a lead generation landing page with 6,000 visitors a week (about 26,000 a month) and a baseline conversion, form submitted divided by visits, of 4.00 percent. Over four weeks, the funnel looks like this.

step	count	step rate	lost at this step	ceiling of relative gain at this step
visits	24,000
scrolled past first screen	15,600	65.00%	8,400	+54% (to 100%)
clicked CTA	2,880	18.46% of scrollers	12,720	very large
started form	2,160	75.00% of clickers	720	+33% (to 100%)
submitted form	960	44.44% of starters	1,200	+125% (to 100%)

You can rebuild this table for your own page with the conversion funnel calculator.

The biggest leak is not the criterion

Here is the part most audits get wrong. The end to end rate is the product of the step rates: 65.00% times 18.46% times 75.00% times 44.44% equals 4.00%. In a product, a 10 percent relative improvement in any factor becomes a 10 percent relative improvement in the result. It does not matter whether that factor is the leakiest step or the healthiest one.

So the absolute loss column, which makes the CTA step look like the emergency, is the wrong ranking signal. Two other things decide where to work:

1. Headroom. Taking form starts from 75.00 to 82.50 percent consumes a large share of what is left before 100 percent; the ceiling is a 33 percent gain. Taking submissions from 44.44 to 48.89 percent is a small step for a step that could, in principle, more than double.
2. Cost per unit of movement. Some steps respond to one well understood change. A form that asks 7 questions and loses more than half of the people who start it has a known, cheap lever: fewer fields. You can model the field count effect on your own numbers with the form fields conversion calculator.

The scroll step deserves one more note. Nielsen Norman Group’s eyetracking research found that users spent about 57 percent of their page viewing time above the fold and 74 percent in the first two screenfuls. A 65 percent scroll rate is therefore not a disaster to fix; it is a reminder that whatever must be read to convert has to live near the top.

The elements, and where each one gets its own guide

Every landing page is built from the same handful of parts. Each of them has its own article on this blog with its own test design, so here they get one paragraph each: what the element changes in the funnel, and where to go deeper.

Headline and message match. The headline is where a visitor checks that they arrived where the ad promised. A mismatch costs at the scroll and CTA steps, and it is also the part of the page that the landing page experience component of Quality Score is about. See ad to landing page message match.

Hero visual and video. A hero video changes content, load time and the first screenful at once, which makes it one of the most confounded tests on a page. See landing page video A/B testing.

Social proof and trust badges. Proof answers “does this work for people like me”, badges answer “is it safe to hand over my data”. Both act mostly near the CTA and the form. See social proof A/B testing and trust badges A/B testing.

Form. For a lead page the form is the last step and often the cheapest one to move, because the main lever, the number and kind of fields, is well understood. It is the first hypothesis in the example below.

Body copy length. Long against short is usually a package test that changes several things at once; it is legitimate if the report does not pretend otherwise. See long copy vs short copy.

Speed. Not an element but a property of all of them. Web Vitals set the targets for a good experience: LCP within 2.5 seconds, INP of 200 milliseconds or less and CLS of 0.1 or less, measured at the 75th percentile of page loads. See page speed and conversion.

For software products specifically, where the page usually sells a trial or a demo, B2B SaaS landing page A/B testing covers the differences. For a quick heuristic pass over a live page before you build the funnel, the landing page grader gives a structured first read.

Prioritize with ICE

With the funnel in hand, write hypotheses that name the step they expect to move (the format is in how to write an A/B test hypothesis) and score them. ICE rates each idea from 1 to 10 on Impact (how much it could move the macro conversion), Confidence (how much evidence supports it) and Ease (how cheap it is to build and run). The score below is the simple average of the three, the same calculation used in the PIE and ICE score calculator.

hypothesis	Impact	Confidence	Ease	ICE
Cut the form from 7 to 4 fields (form start to submit sits at 44.44%)	8	7	8	7.67
Headline that repeats the ad’s promise (message match)	7	6	9	7.33
Social proof next to the CTA	5	5	8	6.00
Change the CTA button colour	2	3	10	5.00
Video in the hero	5	3	4	4.00

Three observations about the scores, because the numbers are only as good as the reasoning behind them:

• The form hypothesis gets a high Confidence because the funnel points at it: more than half of the people who start the form abandon it, and the step has the largest ceiling of the four. That is evidence, not intuition.
• The button colour has the best Ease on the list and still lands fourth. A trivial build with a trivial expected effect is a cheap way to burn weeks of traffic, as the next section shows.
• The hero video has a plausible Impact but a low Confidence and a low Ease: it is expensive to produce and it confounds speed with content.

ICE is a sorting device, not a forecast. Its only job is to put the most promising, best supported and cheapest ideas at the top of a queue whose length is set by traffic.

The traffic budget: what your page can afford to detect

Before launching anything, ask the question most teams skip: with my traffic, what is the smallest effect this test can reliably detect? That number is the minimum detectable effect (MDE), covered in depth in minimum detectable effect. For our page, at a 4 percent baseline, 95 percent confidence, 80 percent power, a two-sided test, two variants and 6,000 visitors a week:

relative MDE	visitors per variant	days	tests per year (14-day floor)
5%	154,304	361	about 1.0
10%	39,475	93	about 3.9
15%	17,943	42	about 8.7
20%	10,317	25	about 14.6
30%	4,783	12, run as 14 (calendar floor)	about 26.1

Put the sample size calculator to work on your own page. With the inputs current conversion rate 4, minimum detectable effect 20 relative, confidence 95, power 80, 6,000 visitors per week and a two-sided test, the screen shows 10,317 visitors per variation, a 20,634 total for two variations and an estimated duration of 25 days.

Current conversion rate%Minimum detectable effectrelative (%)absolute (pp)Confidence909599%Power8090%Visitors per week (total)Testtwo-sidedone-sided

-Visitors per variation

-Total (2 variations)

-Estimated duration

Two-proportion normal approximation, 2 variations (50/50). Tweak the inputs and watch it update live.

Run the arithmetic backward and you get the page’s real budget: four full weeks give 12,000 visitors per arm, and at that size the MDE is 17.72 percent relative, or 0.71 percentage points on a 4.00 percent baseline. The conclusion is blunt. At this traffic, only big swings, around 18 to 20 percent or more, are testable in a month. A button colour test, whose honest expected effect is a few percent at best, would need 361 days, nearly a year, even at a 5 percent lift, and in practice never finishes. That is not a reason to give up testing. It is a reason to test bold changes to the steps with the most headroom, and to treat small polish as something you ship on judgment. Strategies for even smaller sites are in CRO for low traffic sites.

Worked example: the four-field form, end to end

Hypothesis 1 wins the queue. The plan, written before launch:

• Change: the form goes from 7 to 4 fields; nothing else on the page changes.
• Primary metric: form submitted divided by visits, the macro conversion.
• Expected step: form start to submit, currently 44.44 percent.
• Size: MDE of 20 percent relative, which needs 25 days; rounded up to 28 days, four full weeks, so every weekday appears four times.
• Stopping rule: no decision before day 28, whatever the dashboard shows on the way.
• Guardrail: LCP of both arms stays within 2.5 seconds at the 75th percentile.

After 28 days the test closes with this data:

arm	visitors	conversions	conversion rate
A (7-field form)	12,000	480	4.00%
B (4-field form)	12,000	571	4.76%

Control (A)

VisitorsConversions

Variation (B)

VisitorsConversions

Confidence90%95%99%

Control (A) · Rate-

Variation (B) · Rate-

Relative lift-

p-value-

95% CI of the difference-

Two-sided two-proportion z-test. "Not significant" almost always means not enough sample, not that the versions are equal.

Enter 12000 and 480 for Control (A) and 12000 and 571 for Variation (B). The calculator shows a rate of 4.00% for A and 4.76% for B, a relative lift of +19.0%, a p-value of 0.0041, a 95% CI of the difference of +0.2% … +1.3% (pp), and the verdict Significant winner · B wins.

Behind the rounded screen, the exact figures from the same formula are a rate of 4.7583 percent for B, an absolute lift of 0.7583 percentage points, a relative lift of 18.96 percent, a p-value of 0.004098 and an interval from 0.2406 to 1.2760 percentage points. The z statistic, 2.87, is not shown on the calculator; it comes from the formula itself. The observed lift of about 19 percent sits just above the 17.72 percent the page could detect in four weeks, which is exactly the kind of effect this budget was sized to see.

What the winner is worth

Now translate it into money, with an illustrative value of US$ 300 per lead:

estimate	extra leads per month (26,000 visits)	value per month at US$ 300
lower bound of the CI (0.2406 pp)	about 62.6	about US$ 18,770
point estimate (0.7583 pp)	about 197	about US$ 59,150
upper bound of the CI (1.2760 pp)	about 332	about US$ 99,530

The point estimate is the number that ends up on a slide, and it is the one to distrust. When you pick a winner because it cleared a significance bar, the effect you observed is, on average, larger than the true effect: the tests that happened to draw a lucky high reading are exactly the ones that pass. That is the winner’s curse. Plan with the lower bound, US$ 18,770 a month, and treat anything above it as upside. A forecast built on US$ 59,150 will disappoint even if the change is genuinely good.

Reading the result without fooling yourself

The worked example gave a clean answer. Most tests do not, and the reading rules matter most when the result is ambiguous.

Read the interval, not the point. A lift of +19.0 percent with an interval from +0.2 to +1.3 percentage points says “almost certainly positive, size uncertain by more than a factor of five”. Reports that show only the point estimate hide that uncertainty.

Do not stop early, and do not peek and stop. Checking the result repeatedly and stopping at the first significant reading inflates false positives far beyond the nominal 5 percent. Evan Miller’s classic example of a test stopped at the first 5 percent significance reading, or after 150 observations, shows a false positive rate of 26.1 percent when the change does nothing. The fix is the one he gives: decide the sample size in advance and wait until the experiment is over. See the peeking problem.

One main metric. The test was declared on form submissions over visits. If you look at ten metrics after the fact, one of them will “win” by chance. Micro conversions stay as diagnostics: they explain why the macro moved, they do not decide.

Guardrails. Declare in advance what must not get worse. For a landing page the obvious one is speed: if the variant pushes LCP past 2.5 seconds at the 75th percentile, a win on conversion may be borrowing from paid search quality and from mobile visitors. Lead quality is a second guardrail for forms: fewer fields can raise volume while lowering qualification, so check what sales does with the extra leads before you call it a final win.

Whole weeks. Even when the sample is reached early, close the test on a full weekly cycle so Monday traffic and Saturday traffic are both represented in the same proportion as in real life.

Common mistakes

1. Ranking steps by absolute loss. In a multiplicative funnel the biggest hole is not automatically the best target; headroom and cost to move decide.
2. Testing changes too small for the traffic. If the expected effect is below your four-week MDE, the test cannot end with an answer.
3. Optimizing a micro conversion. More CTA clicks with the same number of submissions is not a better page.
4. Changing several things and reporting one. A package test is fine; crediting the whole lift to the headline is not.
5. Stopping on a good day. Peeking turns a 5 percent error rate into something several times larger.
6. Forecasting with the point estimate. The winner’s curse makes it optimistic by construction.
7. Ignoring speed. A heavier variant can win on desktop and lose on mobile, and the average hides it.
8. Comparing your rate to someone else’s average. Benchmarks give context, not targets; see conversion rate benchmarks for how to use them without being misled.

The ecommerce and SaaS versions

The method does not change with the business model; the macro conversion, the funnel steps and the guardrails do.

page type	macro conversion	typical micro steps	guardrail worth declaring
lead generation	form submitted	scroll, CTA click, form start	lead quality, LCP
ecommerce product page	purchase	add to cart, checkout start	revenue per visitor, refund rate
SaaS trial or demo	trial started or demo booked	pricing view, signup start	activation in the first week

In ecommerce, the primary metric is better expressed as revenue per visitor than as conversion rate, because a variant can sell more units of cheaper items. In SaaS, the landing page win is only real if those trials activate at a similar rate, which is why activation belongs among the guardrails.

Landing page optimization checklist before launch

Copy this list into your test plan and fill in every line before the variant goes live.

1. Macro conversion defined in one sentence, with the exact event that counts.
2. Funnel built for the last four weeks, with counts and step rates for each micro conversion.
3. Target step chosen by headroom and cost to move, not by absolute loss alone.
4. Hypothesis written with the change, the step it should move and why.
5. ICE score recorded, with the evidence behind the Confidence rating.
6. MDE checked against traffic: the expected effect is at or above what four weeks can detect.
7. Sample size and duration fixed in whole weeks, never shorter than one full week.
8. Stopping rule written: no decision before the planned date.
9. One primary metric declared; micro conversions listed as diagnostics.
10. Guardrails declared with thresholds: LCP within 2.5 seconds at the 75th percentile, plus lead quality or revenue per visitor.
11. Variant checked on a real phone on a mobile connection.
12. Value plan prepared with the lower bound of the interval, not the point estimate.

Automate this with Donnu

The painful part of landing page optimization is not having ideas. It is running each idea as a clean experiment: building the variant without waiting on a developer, splitting traffic correctly, tracking the macro conversion instead of a proxy, and getting a report that does not invite you to stop on a lucky day.

Donnu runs A/B tests and Split URL tests on web pages. The visual editor lets you build a variant like the shorter form or the message-matched headline directly on the live page, and Split URL covers the cases where the variant is a whole different page. Conversion goals can be a click, a form submission or a page visit, so the test is decided on the action the page exists for. Reports are Bayesian, and the product is built PT-BR and LGPD first.

What stays with you is the thinking in this guide: the funnel, the queue, the honest sizing. The sample size calculator and the significance calculator reproduce every number above on your own data, free.

References

• web.dev (Google). Web Vitals. Article read in full. The passages used: “LCP should occur within 2.5 seconds of when the page first starts loading”, “pages should have a INP of 200 milliseconds or less”, “pages should maintain a CLS of 0.1. or less”, and “a good threshold to measure is the 75th percentile of page loads, segmented across mobile and desktop devices”. Supports the speed targets used in the elements section, the guardrail and the checklist. Checked on October 20, 2026. web.dev/articles/vitals.
• Google Ads Help. About Quality Score for Search campaigns. Page read. The passages used: “Quality Score is calculated based on the combined performance of 3 components: Expected clickthrough rate (CTR) … Ad relevance … Landing page experience: How relevant and useful your landing page is to people who click your ad.” Supports the link between landing page relevance and paid search in the definition section. Checked on October 20, 2026. support.google.com.
• Nielsen Norman Group. Scrolling and Attention. Article read in full. The passages used: “users spent about 57% of their page-viewing time above the fold” and “74% of the viewing time was spent in the first two screenfuls”. Supports the reading of the scroll step and the fold annotation in the anatomy diagram. Checked on October 20, 2026. nngroup.com.
• Miller, E. How Not To Run an A/B Test. Article read in full. The passages used: the scenario in which “You stop the experiment as soon as there is 5% significance, or you call off the experiment after 150 observations” and the logo “actually does nothing”, with the answer “Try 26.1%”, and the advice to “Decide on a sample size in advance and wait until the experiment is over”. Supports the peeking paragraph. Checked on October 20, 2026. evanmiller.org.
• The funnel, ICE scores, sample sizes, test result and value estimates in this guide are illustrative and were computed with the same formulas that power the calculators on this blog (two-proportion normal approximation and two-sided z-test).

Read next: Conversion rate optimization guide · Ad to landing page message match · Minimum detectable effect · The winner’s curse · The peeking problem · How to write an A/B test hypothesis · Leia em português