The average conversion rate optimisation (CRO) agency retainer sits around $16,000 a month, with a low end near $2,000 and premium programmes at $30,000, according to Convert's 2026 roundup of agency pricing. For that money, most of the month is research, analysis and test building. Only about 12% of experiments produce a statistically significant win on their primary metric, in Optimizely's analysis of 173,000 experiments, so the work you are paying for is mostly the process that finds the few winners.
This guide covers what a CRO agency does week by week, what it costs, how much traffic a programme needs before testing makes sense, and the signs that a programme is real rather than decoration. The numbers come from Convert, Optimizely, Evan Miller's statistics writing, Microsoft Research and agencies' own published prices.
What a CRO agency is, and what the retainer buys
People searching "cro agency meaning" usually want a plain answer. Optimizely defines conversion rate optimisation as the process of increasing the percentage of conversions from a website or app. A CRO agency is an outside team paid to run that process: find where visitors drop off, form a hypothesis about why, test a change against the current version, and keep the changes that win.
The market is smaller and more fragmented than the search results suggest. Convert's 2025 agency report identified 237 experimentation agencies worldwide. Of those, 55% are dedicated experimentation agencies and 43% are full or multi-service shops. Over half have one to five people, and only 7% have more than 30. Nearly a third run 10 or fewer experiments a month across all their clients, and about 60% manage two or fewer tests per person per month.
If a small agency promises six tests a month on your account alone, ask who builds them: that is above what most of the market delivers per head.
The same report shows the work shifting. 70% of agencies say they are moving from tactical testing towards programme design and enablement, and 65% report clients bringing experimentation in-house, rising to 87% among large agencies. A good agency today is often building your team's ability to test as much as it is running tests for you.
A typical month, week by week
A real programme has a rhythm. The exact split varies, but the four kinds of work below should all appear every month, and tests should be running in the background throughout.
| Week | Main work | What you should receive |
|---|---|---|
| 1 | Research: analytics review, funnel drop-offs, session recordings, heatmaps, surveys, user tests | A short research note with the problems found and where |
| 2 | Hypotheses and prioritisation: turning problems into testable changes, scored against traffic and effort | An updated backlog with a score and a reason for each item |
| 3 | Design, build and QA of the next test, including tracking checks across devices | A test brief: hypothesis, primary metric, sample size, planned end date |
| 4 | Analysis of finished tests, reporting, rollout of winners | A results write-up that includes losers and inconclusive tests |
Source: Octo synthesis from Convert, Optimizely and Speero practice guidance
Research. This is the part most often cut, and the part that decides the win rate. Optimizely recommends problem-first hypotheses: name the problem, the change, and the result you expect. Qualitative research is cheap to start. Jakob Nielsen's long-standing Nielsen Norman Group rule is that five users find about 85% of usability problems, and that three rounds of five, with fixes between, beat one round of fifteen. Microsoft Clarity offers heatmaps and session recordings free, with no traffic limits, so tooling is not an excuse for skipping it.
Prioritisation. The backlog should be scored with a written method, not ordered by whoever shouted last.
Build and QA. A test brief written before launch is the single best protection against a bad result. It fixes the sample size and end date in advance, which matters for reasons covered below.
Reporting. A monthly report that lists only winners is a sales document. Optimizely's data says a healthy programme reaches a conclusive result 35% to 40% of the time, so most tests in an honest report will be inconclusive or losing.
What a CRO agency costs in 2026
Pricing is rarely published, and the figures in circulation are mostly compiled by agencies and tool vendors. Read them as ranges, not quotes.
| Model | Low | Average | Premium | Source |
|---|---|---|---|---|
| Monthly retainer | $2,000 | $16,000 | $30,000 | Convert, 2026 roundup |
| Hourly project work | $200 | $500 | $2,500 | Convert, 2026 roundup |
| One-off conversion audit | $2,500 | $3,500 | $4,500 | Convert, 2026 roundup |
| Invesp's tier estimates, as reported | $2,000 to $5,000 | $6,000 to $8,000 mid tier | $10,000 and up, most around $16,000 | Convert, quoting Invesp |
| Agency retainer tiers, vendor-compiled | $2,000 to $5,000 for 2 to 4 tests | $5,000 to $15,000 for 6 to 10 tests | $15,000 to $50,000 and up | Clicks Geek, 2026, no source given |
Sources: Convert, Clicks Geek
Convert does not say how its summary table was gathered; it draws on Invesp's estimates, older Quora answers and agency pages. The Clicks Geek tiers come from an agency selling CRO and cite no data. Both point the same way: under about $5,000 a month buys a few tests, and a full research-led programme costs five figures.
Individual agencies' starting prices, as compiled in an October 2026 comparison by Web Tonic (an agency we build sites for, which also sells CRO), include KlientBoost from $1,500 a month, SiteTuners from $3,000, Spiralyze and SplitBase from $5,000, Invesp from $7,000 and Conversion Rate Experts from $7,500. Treat these as list prices, not what a scoped programme ends at.
Tools on top. The testing platform is often a separate line. Convert, a vendor, publishes its Growth plan at $299 a month billed yearly and Pro at $420, each with 100,000 tested users a month. Ask whether the retainer includes the platform, and whose account it sits in. The account should be yours, so your test history stays with you if the agency goes.
Contract length. Experts quoted by Convert suggest a minimum of six to twelve months (Riccardo Vandra of Uplifted Consulting) and scoping retainers to a fixed number of tests, such as one to three a month (Lucia van den Brink of Speero). A three-month trial is rarely long enough to see a programme compound.
The traffic test: can your site support a programme?
Before buying a testing retainer, work out whether your traffic can finish a test. Evan Miller's rule of thumb for sample size per variation is 16 times the variance divided by the square of the smallest effect you want to detect, where the variance of a conversion rate p is p times (1 minus p).
| Baseline conversion rate | Absolute change detected | Visitors per variation | Two-variant test, total |
|---|---|---|---|
| 2% | 0.4 points | About 19,600 | About 39,200 |
| 3% | 0.6 points | About 12,900 | About 25,900 |
| 5% | 1.0 point | About 7,600 | About 15,200 |
| 10% | 2.0 points | About 3,600 | About 7,200 |
Source: Octo calculation using Evan Miller's n = 16σ²/δ² rule of thumb
A B2B landing page with 5,000 visitors a month and a 2% conversion rate needs about eight months to finish one two-variant test at that sensitivity. No agency can change that arithmetic. What it can change is what you test: bigger, bolder changes that produce larger effects, tests on the highest-traffic pages, and research-led fixes that do not need a test at all.
Web Tonic's comparison suggests a similar line: below roughly 10,000 monthly visitors, or about 250 conversions a month, buy research, analytics and UX work rather than a testing retainer. Our posts on A/B testing tools for low-traffic B2B sites and on landing page testing tools cover the options in that range.
Real programme or decoration: the signs
Decoration is a programme that produces activity, reports and confident lift numbers without changing the business. These are the signs we would look for.
| Area | Real programme | Decoration |
|---|---|---|
| Win rate claims | Near 10% to 20%, with the definition stated | 30% or more, with no denominator |
| Stopping rule | Sample size and end date fixed before launch | Tests called when the dashboard turns green |
| QA | Written QA step and a sample ratio check on every test | No QA, results taken at face value |
| Backlog | Scored, with research behind each item | A list of button colour ideas |
| Reporting | Losers and inconclusive tests listed with what was learned | Winners only, lifts quoted as revenue |
| Lift sizes | Mostly under 20% | Doublings and triplings |
| Ownership | Test history and tool account in your name | Everything in the agency's account |
Sources: Optimizely, Evan Miller, Microsoft Research, Speero via Convert
The stopping rule. Evan Miller's 2010 article is still the clearest explanation of why peeking ruins tests. In his worst-case example, checking after every visitor and stopping at the first significant reading pushed the false positive rate to 26.1%, more than five times the nominal 5%. A programme should either fix the sample size in advance or use a sequential or Bayesian method designed for early looks.
Sample ratio mismatch. If a test meant to split traffic 50/50 ends up at 47/53, something is broken: a redirect, a bot filter, a tracking bug. Aleksander Fabijan's 2019 Microsoft Research paper describes sample ratio mismatch as a symptom, like a fever, of many data quality problems, based on experiments across four companies and over 25 products. A real programme checks for it on every test.
QA and prioritisation. Speero's 2025 maturity report, cited by Convert, found 52% of companies have no process to QA an experiment before launch, 58% have no clear prioritisation framework, and 25% have exactly one person responsible for the whole programme. An agency should fill those gaps, not inherit them.
Current tooling. Google Optimize stopped working on 30 September 2023. A programme whose last result dates from then has not been running.
Questions to ask before you sign
- What is your win rate, how do you define a win, and over how many tests?
- Show us the sample size calculation for our first test, using our numbers.
- What research happens before the first test, and what do we receive from it?
- How do you prioritise the backlog? Show us a scored backlog from another client, anonymised.
- What is your QA checklist, and do you check sample ratios?
- How many tests per month will run on our account, and who builds them?
- Whose account holds the testing tool and the test history?
- What does a monthly report look like when nothing won?
Convert's experts suggest a programme should aim to return several times its cost; Andra Baragan of ONTRACK Digital uses three to five times as a target. Ask the agency how it would measure that for your funnel, given that Optimizely puts the average revenue-focused experiment at a 0.4% incremental lift.
What this means for choosing a CRO agency
Start with the traffic test. If your key pages cannot finish a test in six to eight weeks, pay for research and better pages first, and buy a testing retainer later. If they can, choose the agency whose process you can audit: a written test brief, a fixed stopping rule, a QA step, honest reports and a researched backlog. When the problem is the page itself rather than the test plan, our marketing websites page explains how we scope a rebuild in Webflow.
