Conversion rate optimisation in 2026: landing pages, A/B testing, benchmarks and a CRO process that works

Conversion rate optimisation is the practice of raising the share of visitors who take the action you want, using research and controlled tests instead of opinion. In 2026 the median landing page converts at 6.6% and the average online store at 2.26%. The reliable wins are still forms, checkout, page speed and message match.

What conversion rate optimisation actually is

A conversion rate is one number divided by another: people who did the thing, over people who could have. Everything contentious about CRO lives in those two numbers. Ask five people in a marketing team for the site conversion rate and you will get five answers, because one counts sessions, one counts users, one excludes paid brand traffic and one counts a newsletter signup.

So fix the definition before you fix the page. Pick one macro conversion that maps to revenue, a purchase, a qualified lead, a booked demo, and count it against a denominator you can defend. Track micro conversions separately. They are diagnostics, not the scoreboard.

CRO is not a set of tricks. It is a loop: measure, research, hypothesise, prioritise, test, decide, then feed the result back into research. The loop matters more than any tactic, because what worked on a SaaS trial page will not transfer to a jewellery product page. Teams that treat CRO as a checklist plateau within a quarter.

One caveat. If traffic is thin the loop still works, but the “test” step becomes qualitative: usability sessions, five-second tests, customer interviews. Statistical testing has a traffic floor, and pretending otherwise produces confident nonsense.

Conversion benchmarks for 2026, and how to use them

Benchmarks are a sanity check, not a target. Use them to answer one question: is my number so far off the market that something is broken? Then go back to your own segments.

The sector spread is the part people underestimate. A 2% conversion rate is a disaster in arts and crafts and a triumph in baby products.

Source: IRP Commerce ecommerce market data, July 2026. The all-sector figure is up roughly 16% on July 2025.

Channel matters as much as sector. Contentsquare’s 2026 benchmark, built on 99 billion sessions across 6,000 sites, puts paid search at 2.8%, AI referred traffic at 1.3% and organic social at 0.7%. Returning visitors convert at 2.9% against 1.7% for new ones. Compare a page fed by paid search with one fed by Instagram and you are comparing two businesses.

Research: finding where the money leaks

Good CRO research pulls from four kinds of evidence, and the value comes from where they disagree.

Quantitative. Funnel and path reports in GA4 or your product analytics tool, segmented by device, channel and new versus returning. Look for a step where one segment drops off far harder than the rest. Set this up first; our marketing analytics guide covers event design and server-side tagging.

Behavioural. Heatmaps, scroll maps, rage clicks and session replay. Replay is the fastest way to see a sticky mobile header covering your button, and the noisiest tool in the kit. Watch replays only for a segment you already suspect from the numbers.

Attitudinal. On-page polls and exit surveys. One question works better than five. “What almost stopped you buying today?” on the confirmation page surfaces objections no analytics tool will show you.

Qualitative. Five moderated usability sessions, plus a read through the last hundred support tickets and twenty sales calls. Sales teams know your objections better than your analytics does.

Turning findings into hypotheses you can rank

A finding is not a test. “Mobile checkout drops off at the shipping step” is a finding. The hypothesis is the sentence that commits you: because mobile users abandon at shipping and exit surveys mention unexpected delivery costs, showing the delivery price on the product page will lift mobile checkout completion.

Write every one in that shape. Because [evidence], changing [element] will move [metric] for [segment]. If you cannot fill the evidence slot, the idea goes to the bottom of the list however much the founder likes it.

Then rank. ICE scoring (impact, confidence, ease, each out of ten) is crude and works fine. The discipline is not the arithmetic, it is that confidence must be justified by research. High impact plus low confidence is a research task, not a test.

Here is my rough ordering of test families by how often they produce a measurable win. Treat it as a starting bias, not a law.

A/B testing statistics without the maths degree

You need four numbers before a test starts, not after. Baseline conversion rate. Minimum detectable effect, the smallest lift you would care about. Significance level, conventionally 5%. Power, conventionally 80%. Feed those into any sample size calculator and it returns visitors per variation. That number is your commitment.

The arithmetic is unforgiving in a way people find upsetting. At a 2% baseline, detecting a relative lift of 10% at 80% power needs roughly 30,000 visitors per variation. A 5% lift needs about four times that. So a store doing 20,000 monthly sessions should be making bigger, bolder changes rather than tuning a button.

Three rules save more tests than any tool feature. Run in whole weeks, because Tuesday behaviour is not Sunday behaviour. Do not stop the moment significance flashes green, since repeated checking inflates false positives badly. And run a sample ratio mismatch check: if a 50/50 split arrives as 52/48, the test is broken, whatever the p value says.

Two more traps. Novelty effects flatter changes for a fortnight with returning visitors before fading. And the winner’s curse means a barely significant, underpowered test overstates its own lift. If a test claims a 40% improvement on 300 conversions, it did not find one.

Landing page anatomy that holds up in 2026

The median landing page converts at 6.6%, but that median hides a spread from about 3.8% in SaaS to 12.3% in events and entertainment, per Unbounce’s benchmark study. The pages at the top of any category tend to share the same skeleton.

Message match first. The headline should repeat the promise of the ad, email or search result that brought the person there, in their words. It is the most common defect I see on client accounts and free to fix. If your search ad says “same day boiler repair in Leeds” and the page says “welcome to our website”, the auction cost you money for nothing.

One page, one job. A landing page with a navigation bar, a chat widget, a newsletter popup and three competing calls to action has not decided what it wants. Strip the exits on paid pages. Keep them on organic pages, where people came to research.

Then the four things a visitor needs before acting: what it is, what it costs, why you rather than the alternative, and what happens after they click. Put proof beside the decision, not in a testimonials section at the bottom. A named customer with a number beats five anonymous five star quotes. Case studies and comparison pages do real conversion work here; our content marketing guide covers building that library.

Design for the thumb. Mobile is 69.9% of visits and converts far worse than desktop, and much of that gap is mechanical: tap targets, keyboard types on inputs, sticky bars covering buttons, hero images that push the price three scrolls down.

Forms and checkout, where the cheapest wins are

Baymard Institute’s benchmark found the average checkout asks for 11.3 form fields when most sites need no more than eight (Baymard). Field count matters more than step count, which surprises teams who spend weeks arguing about one page versus three page checkouts.

The stated reasons for abandonment are stable across 50 studies, and mostly commercial rather than cosmetic.

Source: Baymard Institute, aggregated from 50 studies. Note that 42% of people abandon because they were browsing, which no checkout change will recover.

For lead generation the same logic applies to your enquiry form: every field is a tax on volume, and anything you can enrich later (company size, industry) goes first. And recover what you lose rather than only preventing it. Abandoned cart and browse flows are the highest revenue per send messages most stores have; the mechanics are in our email and SMS marketing guide.

Speed, Core Web Vitals and conversion

Google’s study with Deloitte and 55, covering 37 brands and more than 30 million sessions, found a 0.1 second improvement in mobile load speed raised retail conversion by 8.4% and average order value by 9.2%. Travel conversion rose 10.1% and lead generation form progression improved 21.6% (web.dev). It runs on 2019 data, so treat the multipliers as indicative. The direction has never been seriously contested.

In 2026 the metric to watch for conversion work is Interaction to Next Paint, which measures responsiveness across every click, tap and keypress in a visit rather than only the first. Google’s thresholds are 200 milliseconds or under for good, up to 500 milliseconds for needs improvement, and above that poor (web.dev). INP is where heavy tag managers, chat widgets and client side testing scripts do their damage, which is a genuine tension: the tool you installed to raise conversion can lower it.

Use field data from the Chrome UX Report rather than a lab score, and fix the slowest revenue template first, usually product detail or checkout. Speed is also a ranking input, so the work pays twice; see the SEO guide for how Core Web Vitals sit alongside the rest of Google’s 2026 signals.

Testing tools in 2026, after Google Optimize

Google Optimize and Optimize 360 were switched off on 30 September 2023, with Google pointing users at AB Tasty, Optimizely and VWO and opening its APIs so third party tools could integrate with GA4 (Google). Three years of consolidation followed, and it has narrowed the market noticeably.

Sources: Convert, OpenAI and VWO.

Practically, you are choosing between three shapes. Client side visual editors (VWO, AB Tasty, Convert, Optimizely Web) are fast for marketers and carry a flicker and INP cost. Server side and feature flag platforms (GrowthBook, LaunchDarkly, Amplitude Experiment, Statsig) need engineering time but test anything, including pricing and logic, with no flicker. Open source is the honest answer for teams who were never going to afford enterprise pricing.

Every vendor now ships AI features that propose test ideas and write variations. Useful for the blank page problem, dangerous as a decision maker. Keep a human on what a result means, because a model that has not read your support tickets does not know why the test won. More in our guide to AI in digital marketing.

Optimising for AI referred visitors and agents

This is the newest section of any CRO plan, and the data is now good enough to act on. Adobe Analytics, working from more than a trillion visits to US retail sites, reported that in July 2026 traffic arriving from AI assistants converted 60% better than other traffic, the eleventh consecutive month of outperformance. Those visitors also generated 53% more revenue per visit, bounced 33% less and added 28% more items to cart (Digital Commerce 360). Contentsquare saw the same shape in a different dataset: AI referred conversion up 55% year on year, volume up 632%.

The volumes are still small, which is why this matters now. AI referred sessions sit at 1.3% in Contentsquare’s data, below paid search, but they arrive pre researched. A visitor who has already compared you with two competitors inside a chat window does not need your comparison page. They need price, stock, delivery date and a fast path to buy.

Three practical changes. Keep price, availability and delivery promise identical across your structured data, your feed and your page, because assistants read the first and shoppers verify the third. Do not gate the information an assistant needs behind a form. And check your consent banner and bot rules do not break the landing experience.

Agentic checkout, meanwhile, has cooled. OpenAI moved away from Instant Checkout on product listings in March 2026 and pushed buying into merchant branded apps, citing the complexity of stock, tax and pricing updates (Digital Commerce 360). My read is that discovery in AI and purchase on your own site is the pattern to optimise for through 2026, which puts your checkout back at the centre of the plan. The ecommerce marketing guide tracks the protocols, and the September ecommerce news roundup has the latest moves.

A 90-day CRO programme, and what stalls it

Days 1 to 20: fix measurement. Agree one macro conversion, audit the tracking, confirm consent handling, build segmented funnel views by device and channel. Nothing else counts until the numbers are trustworthy.

Days 21 to 40: research. Run the four evidence types above, writing findings as one line each with a screenshot or number attached. Aim for thirty findings, not three.

Days 41 to 50: hypotheses and prioritisation. Convert findings into hypotheses, score them, pick the three you will run first, and calculate the sample size for each now. If two are not testable at your traffic, replace them with ship-and-measure changes or usability work.

Days 51 to 90: run. Two or three concurrent tests on different templates, each for whole weeks, with a written decision at the end. Also ship the obvious defects without testing them. Nobody needs an experiment to prove a broken postcode validation should be fixed.

What stalls programmes, in order. Testing pages with too little traffic. Calling tests early. Running one test a quarter and expecting compound gains. Optimising for a conversion the sales team later rejects as unqualified, which is why lead quality belongs in your primary metric. And treating CRO as a project with an end date rather than a standing function, the mistake that quietly costs most. For how CRO fits the channels feeding your funnel, start at our complete digital marketing guide, and the weekly news roundup covers conversion-affecting changes as they land.

Last researched and updated: 7 September 2026.