Conversion optimisation: hypothesis, test, learn, repeat
Traffic is expensive. Turning more of it into inquiries is the cheapest growth most sites have. We find the friction in the data, form a hypothesis, test it properly and ship the fix in your code.
- 2-week diagnostic
- Tests you can trust
- Fixes shipped, not slides
Where visitors leave, and the loop that fixes it
The practice we apply to our own contact form
Our contact form is the example we point to, because it follows every rule below and because we can show you the code. It is also the form that pays our salaries, so we test it.
- Ask only what you need to reply: name, work email, what they are interested in, a message. Budget is optional.
- Visible labels on every field, never placeholder-only. Autocomplete attributes so browsers can fill it.
- Inline validation that says what to fix, next to the field, without clearing the rest.
- A honeypot field and a timestamp for spam instead of a CAPTCHA that punishes humans.
- A status message that confirms what happened and when you will hear back ("we reply within one business day").
- Works without JavaScript: a plain form post and a redirect, so nothing is lost when a script fails.
- One primary call to action per page, phrased as what happens next: "Book a scoping call", not "Submit".
- A secondary, lower-commitment path for people who are not ready: a guide, a tool, a checklist.
| Stage | Common test | What usually moves it |
|---|---|---|
| Visit | Hero headline and sub-line | Saying who the page is for in the first sentence |
| Visit | Page speed | LCP under 2.5s on mobile; no layout shift |
| Engage | CTA placement and wording | A CTA after the first proof point, not only in the footer |
| Engage | Proof | Specific process, real deliverables, named technologies |
| Form start | Field count | Cutting optional fields; moving qualification to the reply |
| Submit | Trust signals near the button | Reply time, no-newsletter promise, privacy link |
| Qualified | Routing and response time | Automated routing to the right person within hours |
Tests you can trust, or no test at all
Most "winning" tests are noise read optimistically. The rules are simple and we hold to them: decide the sample size before the test from the baseline rate and the smallest change worth detecting; run whole weeks so weekday and weekend traffic are both included; do not stop early because the chart looks good; report a confidence interval, not a single lift.
When the traffic is too low for that, we say so and switch methods: session recordings, five-second tests with real prospects, support and sales call summaries, and sequential changes measured across months. A clear qualitative finding beats a fake p-value.
Where AI helps is in finding hypotheses: clustering session recordings by behaviour, summarising hundreds of chat transcripts into the ten questions visitors could not answer, and drafting copy variants for a person to choose between. A well-built assistant on the page can also be a conversion path in its own right, if it answers rather than deflects.
- Weeks 1–2
Diagnostic
Funnel by page and device from your analytics, form analytics, recordings, speed data, and a review of the ten most visited pages. Output: a ranked list of hypotheses with expected value.
- Weeks 3–6
First tests and fixes
Obvious fixes shipped without testing (broken validation, missing labels, slow images). The first one or two tests launched with pre-registered sample sizes.
- Monthly
Cycle
Read results, log the learning, ship the winner, start the next test. The log is reviewed quarterly to update the hypothesis list.
Conversion depends on what comes before and after
Core Web Vitals
Speed is the first conversion fix on most sites.
Fix speedAI Analytics
Funnels, attribution and the dashboards tests are read from.
See the dataCustomer Support AI
Answer the pre-sales questions that block a form fill.
See the solutionMarketing Automation
What happens to the lead in the hour after submit.
Automate follow-upFrequently asked questions
How much traffic do we need to run A/B tests?
Enough that a realistic improvement would be detectable in two to four weeks. For a form converting at a few percent, that usually means thousands of visits per variant. Below that, we use qualitative methods, session recordings and sequential changes measured over longer periods instead of pretending the test is valid.
Do you use an A/B testing tool?
When the traffic justifies it. For many B2B sites we implement tests in code with a server-side or edge split, which avoids the flicker and performance cost of client-side tools and keeps the data in your analytics. The tool matters less than the hypothesis and the statistics.
What is the quickest win on most sites?
Forms. Fewer fields, clear labels, inline validation, a visible privacy note and a status message that confirms what happened. We have seen more gain from a form rebuild than from any headline test. Our own contact form is built this way.
Can AI help with conversion optimisation?
Yes, in specific places: summarising session recordings and support chats to find friction, drafting copy variants for a human to choose from, and clustering visitor segments. It does not replace the test. A model's guess about what converts is still a guess until measured.
How do you report results?
Each test gets a one-page record: hypothesis, variant, sample, duration, result with confidence interval, decision and what we learned. Wins and losses both go in the log. Over time the log becomes the best description of your customers you own.