• Конверсія · Google Analytics 4

A/B testing: a hunch can be checked with figures only when there are enough figures

The temptation is understandable: instead of arguing about the button colour, show half the visitors one version and half the other, and see where there are more orders. The problem is that «more» still has to be proven. If you get twenty orders a month, the difference between twelve and eight is not a result but a coincidence that is easy to pass off as a victory. So we start not with the test but with a calculation of whether there will be anything to compare at all.

See how it works
Price
after a free audit
Guarantee
30 days after project sign-off
Focus
measuring and verifying changes on a site
after a free audit
Price
30
Guarantee

days after project sign-off

measuring and verifying changes on a site
Focus
10
Timeline

to 40 working days per test

a free calculation of whether the data is sufficient
Start
30
Warranty

days from the date the acceptance act is signed

within 2 hours
Reply
under a written contract
Terms
What's included

Complete list of work and what you get as a result

  • a calculation before the start: how long it will take for a difference to become provable
  • an honest answer that testing is not worth it at your volumes, if that is the case
  • checking one assumption at a time — otherwise it is unclear what exactly worked
  • an even and random split of visitors between the versions
  • recording the expected result before launch rather than interpreting the figures afterwards
  • a stopping rule: when there is enough data and when the test must be halted early
  • the conclusion «there is no difference», if there is none — that is a full result too
  • a list of what is pointless to test at your volumes, with the reason why
When this service isn't right

What's not included — so there are no surprises at delivery

  • promises of conversion growth as a result of the test
  • testing where there are not enough visitors for any conclusion at all
  • endless testing of trifles instead of fixing obvious faults
  • rebuilding pages and design — adjacent work with a separate estimate
  • conclusions from a test stopped early because the figure looked good
Who it's for

Situations where this service delivers results

Scenario 1 of 4

A store with a noticeable flow of orders

Hundreds of orders a month is the volume at which a test makes sense and produces a conclusion in a reasonable time rather than in six months.

We'll review your situation in a free audit
Free calculation of whether the data is sufficient

We calculate how long a test would take at your flow. The answer often closes the question without any work at all.

What we measure

  • How many visitors and target actionsThe key figure. The length of a test depends not on wishes but on how many orders you get per week.
  • How big a difference you are looking forA third more is proven quickly. Two per cent more takes months, and is usually not worth the effort.
  • Whether target actions are measuredA test without reliable measurement of an order will show nothing. If analytics is broken, that is where to start.
  • What exactly you want to testA small change produces a small effect that is hard to prove. The best candidates are things that change a person's path, not the colour.
  • Whether there are obvious faultsIf the form does not work on a phone, testing the heading is pointless. The obvious comes first, and we will say so outright.
  • Who will make the decisionIf the result will be overridden by the manager's taste anyway, the test becomes an expensive formality. Better to establish that before the start.

What you get

  • a calculation of the test's duration on your figures
  • an answer on whether testing makes sense in your case
  • a list of what is worth fixing without tests
  • an estimate of the work, if a test is appropriate after all

Timeline: 1–3 working days

Why is it free

The calculation takes hours, and the answer is often that the data will not be enough. Selling a test that will produce no conclusion is simply spending someone else's money the long way round.

What's next

You get the calculation and the list. Then you decide for yourself — the document stays with you.

Short form: your contact and site URL

  • Contract, act and 30-day warranty

    Every project gets a written contract: scope, deadlines, amount, acceptance procedure. After delivery — act and invoice, then 30 calendar days of warranty.

  • Sole proprietor & bank transfer

    The contractor is a registered sole proprietor. Payment by invoice with closing documents.

  • Rights & access — yours

    Code, design and materials transfer to you after full payment. Domain, hosting, repository and analytics are registered to you.

  • Client portal instead of email chains

    During the project you get access to a portal: contracts, invoices, acts and project status in one place.

  • European clients

    Among our work — projects for Norway, Bulgaria, Moldova and Spain.

  • Verifiable numbers

    Every case in the portfolio comes with a link to a live site and a technical measurement.

  • Audit first, then pricing

    There is no price list on the site intentionally: the scope of the same work differs multiples between clients.

  • We say "no" when unsure

    If the task isn't ours or the deadline is unrealistic — we tell you upfront.

What affects the price

Why two seemingly identical tasks are priced differently

  • State of the measurementIf target actions are measured reliably, we build on that. If analytics is broken or absent, it has to be fixed first.
  • Complexity of the changeA different heading takes an hour. A different order of checkout steps is already a second working version of the page that has to be maintained.
  • Number of testsOne test is a one-off job. A standing queue of hypotheses means monthly management with analysis of results.
  • Technical baseOn a page assembled on the server, showing two versions is harder than on one drawn in the browser — and that affects the speed.
  • Depth of the conclusionComparing the share of orders is simple. Proving that the change did not spoil something else — returns, average order value — is a separate analysis.
  • DurationOn a small flow a test runs for weeks, and all that time it has to be watched so nothing breaks. Time is a cost too.
Cases

Tasks and results in numbers — all metrics measured by us

a free public service: a converter of KVED-2010 codes into NACE 2.1-UA for entrepreneurs

Task
In 2026 Ukraine is moving to the NACE 2.1-UA classifier, and every sole trader has to work out their new code. There is no automatic transition, and the official correspondence tables are inconvenient for a person. The task was to let an entrepreneur find their new code from the old one — or simply upload an extract from the state register — and get an answer in a minute, free of charge.
Solution
An application on Next.js: a converter by the old code, the full correspondence table, code recognition straight from an uploaded document, separate pages for every code, and explanatory material on which single-tax groups are available to whom and how the transition works. Every code got its own page, so a person finds the answer straight from search rather than through the home page. Indexing is open, including rules for the crawlers of AI assistants.
Result
Measured 02.08.2026: the service works, first server response 244 ms, the sitemap holds 1,366 addresses — 616 activity codes plus supporting pages. For the period 1 July – 2 August 2026: 349 active users, 346 of them new, 2.3 thousand events, average engagement time 1 min 45 s. The converter's home page had 514 views and 288 active users.

Did not find your case?

Describe how it works on your side — we will tell you whether “A/B testing on the site” fits and what it means in your situation. No brief and no call: one question, one answer.

Process steps

Transparent stages with approval at every step

Total duration:10–40 days

  1. The sufficiency calculation

    1–3 working days

    We calculate on your figures how long the test would take. If it comes out at half a year, we say so straight away — and the work often ends there.

  2. Choosing the assumption

    1–3 working days

    We choose what to test. What works are changes that touch a person's path, not the styling of trifles: only those produce an effect large enough to prove.

  3. Preparation

    3–10 working days

    We build the second version and check that the measurement of target actions works identically in both. A mistake here devalues the whole test.

  4. Launch and observation

    10–30 calendar days

    We launch with an even split and leave it alone. The most common mistake is stopping on the third day after seeing a good figure: it is almost always random.

  5. The conclusion

    2–4 working days

    We compare the result against the expectation recorded before launch. If there is no difference, we say so: that is a normal and useful outcome.

  6. Decision and next step

    1–2 working days

    Either the new version stays, or the old one comes back, or we call the test inconclusive. Then the 30-day warranty from the date the acceptance act is signed applies.

Technologies & integrations

What we build on and what it connects to

Stack

  • Google Analytics 4
  • Google Tag Manager
  • Google Optimize
  • JavaScript
  • TypeScript
  • HTML
  • CSS
  • Next.js
  • PHP
  • Looker Studio
  • Python

Integrations

  • Google Analytics 4
  • Google Tag Manager
  • Looker Studio
  • Hotjar
  • Clarity
  • CRM
A test or a change made blind

How this option differs from the alternative

Where the difference came fromvisible, because identical periods are compared
If it got worsevisible on a portion of visitors
Requirements on the flowhundreds of target actions needed
Speedweeks of waiting
When the simple path is betterpointless at dozens of orders a month
What we need from you

We can't start without this — best to prepare in advance

  1. how many visitors and target actions per month
  2. what exactly you want to test and what you expect
  3. whether the measurement of orders or enquiries is set up
  4. what size of difference you would consider worth acting on
  5. who makes the decision based on the result
  6. who on your side accepts the work

If something is missing — let us know, we'll help you gather it or do it as a separate task.

FAQ

Most frequently asked questions — with concrete answers

How many visitors are needed for a test to make sense?

The question is not about visitors but about target actions. As a guide: to prove a relatively large difference you need hundreds of orders in each version. At twenty orders a month the test would run for years, and we will say so during the calculation — free of charge and before any work.

And if there is not enough data — what then?

Fix the obvious without tests. A form that does not work on a phone, a hidden delivery cost, compulsory registration — these are not hypotheses but faults, and they need no test. An analysis of the buyer's path will give more here than any testing.

Why can a test not be stopped when an advantage is visible?

Because that is exactly how false conclusions are produced. On small numbers the difference jumps both ways, and if you stop at the moment the figure looks good, you will reliably find a «victory» even between two identical versions. The stopping rule is fixed before launch.

Can several changes be tested at once?

Technically yes, in practice no, if the flow is small. When three things change and the result is better, you do not know what worked and cannot repeat it. One change at a time takes longer but produces knowledge rather than a coincidence.

What if no difference turns up?

That is a normal result, and we name it plainly. It means the change is not worth the effort — and that saves the money that would have gone on rebuilding the whole site in the same direction. Tests that «always find something» usually find noise.

Do you guarantee conversion growth?

No. A test is a way of learning the truth, not a way of improving it. Some hypotheses will not be confirmed, and that is inherent to the method. Anyone who guarantees growth from a test either does not understand the method or is selling something else.

How long does one test take?

10 to 40 working days, and most of it is waiting rather than work. The duration is set by your flow: the fewer target actions per week, the longer you have to wait for the difference to stop being random.

Is this the same as a usability audit?

No, and the two complement each other. An analysis of the buyer's path shows where people stop and produces a list of assumptions. A test answers whether a specific change actually does anything. It is almost always worth starting with the first: it is cheaper and does not require a large flow.

Let us calculate whether you have enough data for a test

A duration calculation on your figures — and an honest answer if testing is not worth it. Free of charge, we reply within 2 hours.

From measured casesMeasured 02.08.2026: the service works, first server response 244 ms, the sitemap holds 1,366 addresses

View cases
  • Reply within 2 hours
  • No commitment
  • We work under a contract

There is no price list on the site on purpose: the same work differs several times over between two clients, and a “from” figure explains nothing in that case. First a free audit — we count your pages, duplicates and speed — then we name the sum and the deadline and fix both in the contract.