How to Build a CRO Roadmap When You Have Limited Traffic and Budget

If your site gets a few thousand visitors a month and converts a few per cent of them into enquiries, you almost certainly cannot run statistically valid A/B tests on small changes. Build your CRO roadmap around fixing measurement, then research that needs no test (five-user usability tests, form analytics, session review, message-match checks and customer conversations), then shipping the obvious fixes, and finally making one bold change and reading its effect honestly. A/B testing comes later, and only on pages where the numbers allow it.

This is written for service and B2B sites where the goal is an enquiry, call or demo request. Online stores have more conversions to work with; for them see what to A/B test first on an ecommerce site.

The table shows roughly how many visitors each variant needs to detect a given relative lift, using the standard formula for comparing two proportions. These are calculated values, not benchmarks.

Assumptions: a two-sided test at α = 0.05 (5% chance of a false positive) and 80% power (80% chance of detecting the lift if it is real), two variants with a 50/50 split, and visitors as the unit. Formula: n per variant = [1.96 × √(2p̄(1 − p̄)) + 0.84 × √(p₁(1 − p₁) + p₂(1 − p₂))]² ÷ (p₂ − p₁)², where p₁ is the baseline rate, p₂ = p₁ × (1 + lift) and p̄ is their average.

Worked check: at a 2% baseline and a 20% relative lift, p₂ = 2.4% and p̄ = 2.2%. The formula gives about 21,100 visitors per variant, or about 42,200 in total. That is around 420 enquiries in each arm before you can tell 2.0% from 2.4% with these settings. Online calculators such as Evan Miller's use slightly different formulas, so expect small differences.

Turn the table into time by dividing total visitors needed (two variants) by weekly traffic to the page under test. The arithmetic below assumes all of the site's traffic reaches that page, which is optimistic; most tests run on one page with a share of the traffic.

2,000 visitors a month is about 460 a week (2,000 × 12 ÷ 52). At a 2% baseline, detecting a 20% lift needs 42,200 visitors, which is about 91 weeks. Detecting a 50% lift needs about 7,650 visitors, about 17 weeks.

10,000 visitors a month is about 2,300 a week. At a 2% baseline, a 20% lift takes about 18 weeks and a 50% lift about 3 to 4 weeks. At 5%, a 20% lift takes about 7 weeks.

So a smaller site can test only very large differences, and a mid-sized one can test large differences on its busiest pages. Practitioners disagree on rules of thumb: the testing vendor Convert suggests about 10,000 visitors per variant with around 1,000 conversions, and others quote different thresholds. Your own calculation is more useful than any of them.

This sequence suits a site with a few thousand to a few tens of thousands of monthly visitors, a small team and no dedicated analyst. Durations assume someone can give it a few hours a week; stretch them if not.

Five-user tests. Nielsen's NN/g article argues that testing with five users finds about 85% of usability problems, based on an average 31% chance of a single user hitting each problem, and that several small rounds beat one large study. It also notes that quantitative studies need more participants (around 20). Recruit people who resemble your buyers and ask them to find out whether you can help with a realistic problem, then to get in touch. Our usability testing guide covers the mechanics.

Form analytics. Track form starts, the field where people abandon and which errors appear. A form that is started often and finished rarely is a better lead than any page-level metric. For what to change, see improving enquiry forms without attracting low-quality leads.

Session review. Watch recordings filtered to people who reached the enquiry page but did not submit. Look for patterns, not anecdotes: three people stuck in the same place matters; one is noise. Check what your tool records and that form inputs are masked.

Message-match audit. For each top ad, campaign and search landing page, put the promise people clicked on next to the page headline and first screen. Mismatches are common and cheap to fix.

Customer and sales-call interviews. Ask recent customers what nearly stopped them and what they compared you with, and ask sales what prospects repeatedly misunderstand. This is where offer and positioning problems show up.

A before-and-after comparison is weaker than a controlled test, so plan for its limits rather than ignoring them.

Once a page can reach the sample in roughly a month or two, testing becomes worthwhile. The free option that many small sites relied on has gone: Google says Optimize and Optimize 360 are no longer available as of 30 September 2023, and it pointed to GA4 integrations with AB Tasty, Optimizely and VWO. Budget for a tool before planning a programme.

On lead-generation sites, test the things most likely to move results by a large margin, because small effects are too expensive to detect: the offer itself (a call versus an audit versus a guide), the headline and value proposition on the main landing page, the number and type of form fields, the call to action and what it promises, and the proof shown near the form. Measure qualified leads as well as submissions where volume allows, since a variant can win on fills and lose on quality. Ecommerce teams have different priorities; see how stores prioritise experiments.

Hypothetical example: a firm gets about 6,000 visits a month and around 90 enquiries, a 1.5% rate. A sales lead tags last quarter's enquiries and finds half are poor fits. The sample-size sum shows a 20% lift would need about 28,300 visitors per variant, roughly 56,600 in total, which is more than nine months of the whole site's traffic, so testing is off the table.

Weeks 1 and 2 fix a contact form that silently failed on one mobile browser and move generate_lead to fire after server confirmation. Five-user tests in weeks 3 to 5 show visitors cannot tell which services are core. Weeks 6 to 9 rewrite the services page and add banded qualifying questions. In week 10 the firm replaces its generic "Contact us" with a defined discovery offer, and compares qualified enquiries over the next eight weeks against both the prior period and the same period last year, noting a trade show that happened in between.

If your site has plenty of traffic and a steady flow of conversions, this roadmap is too cautious; run experiments, and read why tests mislead before you start. If the problem is that too few of the right people arrive, CRO will not fix it; start with diagnosing traffic that does not produce leads. And if you are about to rebuild the whole site anyway, fold the research phase into the redesign rather than optimising pages you will replace.

Counting proxy events such as form starts or button clicks gives you more data sooner, but those events may not track qualified leads. Use them to find friction, not to declare winners.

Google: Optimize sunset notice; Google: GA4 recommended events reference; NN/g: why you only need to test with 5 users; Evan Miller: sample size calculator; Convert: not enough traffic to A/B test (vendor view).

Sample sizes and week conversions are our own calculations using the stated formula and assumptions. Accessed 14 October 2026. Tool availability and analytics features change; check current versions.

With limited traffic, the most valuable CRO work is not experimentation. Count leads properly, find the biggest problems through research, fix what is clearly wrong, and make one meaningful change at a time. Run the sample-size sum for each page; when a page passes it, start testing there. A research-led CRO audit can compress the first two phases if you lack the time to run them yourself.