CRO consulting services: measure the friction, prove the conversion with experiments.

We wrote this guide for marketing, product and ecommerce teams whose sites get steady traffic but who can’t clearly see at which step, and why, visitors give up. The measurement and testing mechanisms we describe are based on the official documentation of Google Analytics, Microsoft Clarity, Firebase and Optimizely, the W3C accessibility standard, a case study published on web.dev, and an announcement from the US Federal Trade Commission (FTC); the sources are listed by number at the bottom of the page. We verified product names and features against the source as of September 2026. Our own field observations are marked separately under the heading “Expert note”.

Our CRO consulting work runs in three layers: measurement that finds where and why visitors drop off; experiments that turn a finding into a testable hypothesis; and implementation that ships the winner and keeps what was learned. The sections below walk through these layers, the calculations we do and the things we won’t do, in that order.

What does CRO (conversion rate optimization) measure, and what does it change?

CRO measures the rate at which visitors complete an action that matters to you, and reduces the friction holding that rate down by changing the site itself rather than adding traffic. The starting point for measurement is deciding which action counts as a “conversion”. In Google Analytics 4, an event that measures an action especially important to your business’s success is called a key event, and any event you collect can be marked as a key event.¹ Google changed this naming because events marked as conversions in Analytics were measured differently from Google Ads conversions, which created inconsistencies between the two platforms; “conversion” now describes the action that measures ad campaign performance.¹ We cover conversion setup on the ad side on our Google Ads consulting page.

In CRO work, we keep four kinds of signal apart:

If one of the intermediate steps becomes the goal, the team chases whichever number is easiest to raise: add-to-carts go up, sales don’t. That’s why the success metric in our tests is always a key event.

How does funnel analysis show where visitors drop off?

Funnel analysis lines up the steps a visitor goes through to complete a task and shows how many leave at each step; in GA4, this is done with funnel exploration, which visualizes success and abandonment step by step.² A funnel can have up to 10 steps; steps are built from events or dimension values, not from metrics.² Whether a funnel is open or closed determines how visitors are counted:²

In both types, visitors are counted only for the steps they complete in the defined order; someone who skips a step drops out of the funnel.² Even if the same person completes the funnel several times within the selected date range, only their first sequence is reported.² The “next action” setting shows the five most common actions after each step, the elapsed time setting shows the average time between steps, and the same funnel can be compared across up to 4 segments.²

We read the funnel first broken down by device category: a step that goes smoothly on desktop but shows a clear drop-off on phones points to a problem with the interface, not the offer.

What do heatmaps and session recordings show?

The funnel shows where visitors drop off; heatmaps and session recordings show why: where the visitor clicked, how far they scrolled, what they expected and couldn’t find. The click map in Microsoft Clarity tracks clicks on desktop and taps on phones and tablets, including clicks on elements that aren’t links.³ From the same screen you can switch to scroll, area, attention and conversion maps. The click types most useful for CRO are:³

• Dead click: places where a click produces no effect or response; elements that look like links in the design come to light this way.
• Rage click: repeated clicks on the same small area within a short time.
• Error click: clicks that come right before a JavaScript error; the first place we look when tracing a broken button.

A heatmap is built from at most 100,000 page views.³ Privacy in session recordings is set during setup: Clarity masks the contents of input boxes in every masking mode, and that behavior can’t be changed; in the default “Balanced” mode, numbers and email addresses are masked too.⁴ Masked content is never uploaded to Clarity.⁴

How do you build a good test hypothesis, and what do we never test?

A good hypothesis rests on an observation, defines a single change and writes down the expected effect before the test; dark patterns that mislead visitors never make it onto our hypothesis list. When ranking hypotheses, we look at three questions: how big is the drop-off, how strong is the evidence, and how expensive is the change? On low-traffic sites this ranking matters even more, because every test runs for weeks.

In a report published in 2022, the FTC addressed dark patterns as design practices that can trick or manipulate consumers into buying a product or service or giving up their privacy.⁵ The tactics in the report include countdown timers on offers that aren’t actually time-limited, subscriptions made hard to cancel, hidden terms and fees, and privacy choices that steer people into sharing more data.⁵ According to the announcement, companies have become able to experiment with these patterns and use the most effective ones.⁵ Since A/B testing is exactly the tool that makes that experimentation possible, we set the limit up front: even if it lifts conversion in the short term, we don’t build a variant that misleads visitors, makes cancellation harder, hides the price or pre-ticks consent.

How is A/B testing set up now that Google Optimize has shut down?

Since Google Optimize shut down, A/B tests are set up in three ways: third-party testing tools integrated with GA4, Firebase A/B Testing, and server-side tests where the variant is chosen on the server. Optimize and Optimize 360 have been unavailable since September 30, 2023; experiments still running on that date ended as well.⁶ Google says it has been working on integrations with AB Tasty, Optimizely and VWO, and has opened up its APIs so anyone can connect their own testing tool to Google Analytics.⁶ Firebase A/B Testing continues; experiments started on or after October 23, 2023 are calculated with frequentist inference, without relying on Optimize infrastructure.⁶ Firebase evaluates results at a 0.05 significance level: it uses a z-test for proportions on binary metrics such as conversion, and a t-test with unequal variances on continuous metrics such as revenue.⁷

According to Optimizely’s documentation, a client-side testing tool works through a one-line snippet added to the page, while server-side experiments work through software development kits (SDKs) and code deployment; server-side experiments can be more efficient than in-browser experiments, which can affect load time or cause a slight flicker.⁸

How many visitors does a test need?

The number of visitors you need is calculated from four inputs: today’s conversion rate, the smallest difference you want to detect, the false positive risk you accept, and the probability of catching a real difference (statistical power). Optimizely’s fixed horizon test documentation likewise calculates the sample from the baseline metric value, the minimum detectable effect (MDE), the significance level, the variance and the number of variants.⁹ In a test comparing two proportions, the sample per variant is written with the common approximate formula as follows:

1. p₁: today’s conversion rate (the baseline rate); it comes from funnel and key event data.
2. p₂: the rate you want to be able to detect if the variant reaches it; with a relative MDE, p₁ × (1 + MDE).
3. α: the significance level, meaning the risk of finding a difference when there isn’t one; in a two-tailed test, z ≈ 1.96 for α = 0.05.
4. 1 − β: statistical power, meaning the probability of catching a real difference when one exists; the higher the power, the larger the sample.

The formula’s denominator is the square of the difference: halving the difference to be detected roughly quadruples the sample you need. And as the baseline rate falls, you need more visitors to see the same relative difference. In the example in Optimizely’s documentation, a 5% baseline rate, a 10% relative lift, a 95% confidence level (α = 0.05) and two variants require 34,363 visitors per variant.⁹

How long does a test run, and when can you read the result?

Test duration is found by dividing the total sample needed by the daily visitors entering the test, and rounded up so it covers at least one full week. Optimizely recommends running every test for at least one business cycle, meaning seven days, to capture variation in behavior such as the difference between weekday and weekend visitors.⁹ Firebase, too, recommends two weeks as the minimum for a typical Remote Config experiment.⁷ When the result can be read depends on the statistical method:¹⁰

In a fixed horizon test, looking at the result before the sample is complete (peeking) increases the probability of a false positive; you risk mistaking a chance difference for a winner and stopping the test early.⁹ With the sequential method, results can be monitored continuously; Optimizely says this method keeps results valid by controlling the false discovery rate.¹⁰

We read the result not as a single percentage, but with its confidence interval. Firebase reports a 95% confidence interval for the difference; if the interval includes zero, no significant difference was detected between the variant and the baseline.⁷ The same documentation recommends looking at secondary metrics and the possible loss at the lower end of the confidence interval before rolling out the leading variant.⁷

How do you simplify form and checkout steps?

Simplifying form and checkout steps means removing unnecessary fields and making every remaining field easy to fill in correctly the first time, with its label, error message and autofill. Our yardstick is the W3C’s Web Content Accessibility Guidelines (WCAG 2.2); the criteria were written for users with disabilities, but the problems they address are friction for every visitor. The success criteria we check:¹¹

1. Labels or instructions (3.3.2, A): content that requires input has a label or instructions; we don’t leave the label to placeholder text.
2. Error identification (3.3.1, A): when an input error is detected automatically, the item in error is identified and the error is described in text.
3. Error suggestion (3.3.3, AA): if a correction is known, it’s suggested to the user unless that would compromise security or the purpose of the content.
4. Identify input purpose (1.3.5, AA): the purpose of fields that collect information about the user can be programmatically determined; in practice, we use the autocomplete attribute, which turns on autofill, for name, email, phone and address fields.
5. Redundant entry (3.3.7, A): if information entered earlier in the same process is needed again, it’s auto-populated or available to select; for example, letting the billing address be selected from the shipping address.
6. Target size (2.5.8, AA): targets selected with a pointer are at least 24 × 24 CSS pixels, apart from defined exceptions.

3.3.7 and 2.5.8 are among the criteria added in WCAG 2.2.¹¹

How do page speed and the testing tool affect conversion?

Both page speed and the testing tool itself can affect conversion; the effect of speed can be measured with a controlled A/B test, and the Vodafone case published on web.dev is a documented example.¹² Half of the paid media traffic was sent to a page optimized for Web Vitals (Google’s user experience metrics), and half to the baseline page; the two versions were visually and functionally identical.¹² In the optimized version, rendering of one component was moved to the server, critical HTML was generated on the server, the main image was resized, and images not yet visible were kept from loading.¹² A 31% improvement in LCP (Largest Contentful Paint, the time it takes to render the largest content element) brought 8% more sales, a 15% improvement in the lead-to-visit rate and an 11% improvement in the cart-to-visit rate.¹²

We share this case as an example of method, not as a promise: the effect of a speed improvement on your own site should also be measured with a controlled test. We cover speed work in detail on our website speed optimization page. The testing tool itself can affect speed too; in-browser experiments can lengthen load time or cause flicker.⁸ That’s why we don’t leave the winning variant in the testing tool; we build it into the site’s code and remove the test script.

How do our CRO services work in the first 30 days?

The work of the first 30 days isn’t launching tests; it’s verifying measurement, finding where and why visitors drop off, and drawing up the first hypothesis queue with sample size calculations. We go in this order.

1. Week 1 · Measurement audit: GA4 key events, funnel steps, the Clarity setup and masking settings are audited; missing measurement is fixed.
2. Week 2 · Diagnosis: funnel drop-off is read by device and channel; heatmaps and session recordings for the step where visitors drop off are reviewed.
3. Week 3 · Hypotheses and calculations: findings are turned into hypotheses, and the sample size and duration are calculated for each; obvious bugs are set aside for direct fixing.
4. Week 4 · First test and plan: the first test that traffic can support goes live; the next month’s test queue and reporting rhythm are set together.

How to choose a CRO agency

Who runs your conversion rate optimization shapes how far you can trust every test: two teams using the same tools can work very differently, one sizing the sample before launch, the other calling a winner on the first good day. The three common models compare like this:

Whichever model you pick, ask for concrete answers to these questions: which key event defines a conversion, how do you calculate sample size and duration before a test, how do you report losing tests, how do you mask personal data in session recordings, and do you use dark patterns? We’d also check the following:

1. Account ownership: the GA4, Clarity and testing tool accounts are opened in your name, with the partner added as an authorized user; the test history and learning archive stay with you as well.
2. Sample size up front: each test’s sample, duration and decision rule are written down before launch; “let’s look in a few days” is how false positives get shipped.
3. Losing tests in the report: inconclusive and losing tests appear in the report with their confidence intervals, not just the winners.
4. Implementation: it’s agreed from the start whether winners stay in the testing tool or get built into the site’s code, and who does it.
5. Guarantee language: a proposal that promises a specific conversion lift before any test has run is claiming to know the result without measuring it.

What determines CRO pricing?

The cost of CRO services depends less on the number of pages than on the flows to be reviewed and the monthly testing pace. Traffic volume plays a part too, because it decides how many tests can finish soundly in a month.

We work out the scope together in the first meeting, then send a written proposal; any testing tool license is separate from our service fee.

CRO consulting pricing: estimated ranges

Estimated range, excluding VAT; rows marked “/ month” are monthly, the others are one-off. Scope is confirmed in a written proposal after the first meeting. As of October 2026.

If you’d like to read together at which step your site loses visitors, pick a meeting day that suits you below; in the first meeting, we clarify your measurement status, traffic volume and priorities together.