Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

Conversion Rate Optimisation › Conversion Research

Conversion Research Methods

The conversion research methods and when to use each: analytics and funnel data, session recordings and heatmaps, surveys and polls, and user interviews.
Guideintermediate12 min readUpdated

What research has to find

Before choosing a method, be clear about what the research is for. A finding is useful to a conversion programme only when it has four properties. It has a location: a page, a step in the funnel (Funnel), a device, a traffic source or a segment of visitors. It has a size: how many people pass through that point and how many are lost, so that its cost can be estimated and compared with other findings. It has a cause, or at least a plausible one, that visitors have described or recordings have shown, because a location without a cause produces a guess rather than a hypothesis. And it has evidence from more than one direction, because every method has a blind spot.

"The checkout converts badly" has none of these. "Visitors on phones who reach the delivery step leave at a much higher rate than desktop visitors; recordings show the address form clearing after a validation error; poll responses mention having to type the address twice" has all four. The gap between those two sentences is the research.

Research also has to find what is not worth testing. Broken functionality, outdated prices and error states are fixed, not tested. And it has to find what is working, because a redesign that removes the element quietly doing most of the persuading is a common way to lose a conversion rate (CVR) that took years to build. The research backlog in the conversion research hub is where findings of all three kinds are recorded with their location, size and sources. Each row in the research backlog also carries a decision, so the list shows what was concluded as well as what was seen.

Methods at a glance

MethodWhat it answersEffortMain bias
Analytics reportsWhich pages, sources and devices convert, and how that changes over timeLow once tracking is soundOnly records what was instrumented; silent on why
Funnel and path analysisWhich step loses the most people, and where they go insteadLow to mediumDepends on step events being complete and in order
Form analyticsWhich field people abandon on, correct repeatedly or leave blankMedium; needs field-level eventsMisses frustration that does not show as an interaction
Session recordingsWhat a visitor did on the page before leavingHigh to review, low to collectThe reviewer sees what they look for; a handful of sessions is not a pattern
Heatmaps and scroll mapsWhere attention and clicks go, and how far people readLowAggregates hide segments; clicks on non-links signal confusion, not interest
On-page pollsWhat visitors wanted or were missing at a given momentLowAnswered by a self-selected minority
Post-conversion and exit surveysWhy people bought, what nearly stopped them, why others leftLow to mediumBuyers rationalise; abandoners rarely reply
Customer interviewsThe decision in depth: alternatives, doubts, languageHighSmall numbers; the interviewer leads
Usability testingWhether people can complete the task, and where they struggleMedium to highParticipants behave differently when observed
Support and sales logsThe questions and objections that recurLowReflects people who bothered to contact you

The sections that follow take the groups in turn: quantitative, behavioural, attitudinal and usability, then how to combine them.

Quantitative: analytics and funnels

Quantitative research starts with the reports that already exist. The conversion rate by landing page, device, source and country, trended over a long enough period to see the normal week and the quiet one, shows where the rate is low and where it has changed. Changes are more informative than levels: a page that converted well until a release and badly since has a date, and a date narrows the search enormously.

The funnel report turns a site-wide rate into a sequence of step rates, and the step with the largest absolute loss is the place to start. Google Analytics (GA4) provides this through Explorations, which Google documents as a set of techniques including funnel exploration, path exploration, segment overlap and user explorer; the Explorations tutorial walks through building them, and the funnel analysis tutorial in the funnel section applies them to a conversion funnel step by step. Three practices matter more than the tool. Segment every funnel by device, because mobile and desktop funnels leak in different places and the blend hides both. Compare open and closed funnels, because a closed funnel shows the intended route and an open funnel shows what people actually do. And treat every step as a hypothesis about the tracking until it has been checked: a step with an implausibly high or low rate is more often a missing or doubled event than a visitor behaviour.

Form analytics, where field-level events exist, extends the funnel inside a single page: which field people reach, which they abandon on, which they correct repeatedly. It is the most direct quantitative method for lead-generation and checkout forms, and it needs instrumentation most sites do not have by default. Site search is one quantitative source that records intent in the visitor's own words, and the site search analysis tutorial shows how to find zero-result and abandoned searches in GA4.

Quantitative methods are precise about what happened and silent about why. Google Analytics does not record intent, confusion or doubt, and a report that shows people leaving the delivery step cannot say whether the cause is price, the form or something else. That is the hand-off to the behavioural and attitudinal methods. The site-wide measurement that quantitative research depends on, from event design to data quality, is the subject of the analytics hub and is not repeated here.

Behavioural: recordings and heatmaps

Behavioural methods show what people did without asking them. Session recordings replay a visitor's scrolling, clicking, typing and hesitation on a page. Heatmaps aggregate clicks and taps across many visitors; scroll maps show how far down the page each share of visitors reached; attention maps estimate where time was spent.

Recordings are most useful when reviewed against a question. "Watch a batch of mobile sessions that reached the delivery step and left" produces findings; "watch some recordings" produces an afternoon. Look for repeated actions on the same element, which signal that it did not respond as expected; for rapid back-and-forth between two pages, which signals a comparison the page did not support; for form fields cleared, re-typed or abandoned; and for rage clicks on things that are not links. Keep a tally so that the pattern is counted, not just remembered. The bias is the reviewer's: people find what they are looking for, and a handful of dramatic sessions can feel like a pattern when they are not one. Set a rule in advance for how many sessions must show a behaviour before it goes on the backlog.

Heatmaps are quicker and shallower. A click map that shows many clicks on an image or a heading that is not a link says people expected it to do something. A scroll map that shows most visitors never reach the call to action says the page is too long or the top of it gives no reason to continue. The main trap is aggregation: a heatmap of all visitors blends new and returning, mobile and desktop, paid and organic, and the pattern that matters may belong to one of them. Filter before reading. Recording and heatmap tools raise consent and privacy obligations, since they can capture personal data typed into forms; mask inputs by default and check the consent position in each market before deploying them. The tutorial on reviewing session recordings and heatmaps turns this section into a procedure, with a coding sheet, an evidence rule set before watching and a comparison between visitors who left and visitors who converted.

Attitudinal: surveys, polls and interviews

Attitudinal methods ask. They are the only route to why, and they are the methods most often done badly, because asking a bad question is easy and the answers look like data.

On-page polls put one short question in front of visitors at a chosen moment: on the pricing page, "What is stopping you from choosing a plan today?"; on exit from checkout, "Was there anything missing from this page?". Response rates are low and respondents are self-selected, so a poll gives themes and language rather than proportions. Its value is specificity: an answer tied to a page and a moment.

Post-conversion surveys reach people who just completed the action and are, briefly, willing to talk. Ask what nearly stopped them, what else they considered, and how they would describe what they bought. The nearly-stopped answers are the most direct source of friction findings. Exit and abandonment surveys, by email where there is an address, reach the people who left; the response rate is lower still and those who reply are not typical, which is a reason to read them for themes rather than counts.

Interviews go deepest. A short conversation with a recent customer, and separately with someone who evaluated and did not buy, surfaces the alternatives they weighed, the doubts they had and the words they use, which later feed copy as well as hypotheses. The constraints are numbers and leading: a few interviews cannot support a proportion, and an interviewer who suggests an answer will receive it. Write the questions in advance, ask about what people did rather than what they would do, and let silences run. The audience research guide covers the broader practice of learning who the customers are; conversion research narrows it to the moment of decision on the site.

Customer support tickets, sales call notes and chat transcripts are attitudinal research that already exists. A monthly read of the recurring questions is cheap and frequently finds the objection the page never answers.

Usability testing

Usability testing gives a person a task, such as finding and buying a replacement filter for a named appliance, and watches them attempt it, thinking aloud, on the real site or a prototype. It is the one method that reliably finds problems nobody knew to ask about: a label that means something different to customers, a step that looks finished when it is not, a button that is visible to the designer and invisible to everyone else.

A session needs a realistic participant, a task written as a goal rather than a set of instructions, a facilitator who does not help, and a note-taker. Moderated sessions, where a facilitator is present, allow follow-up questions and work best for complex flows; unmoderated sessions, where the participant records themselves following a script, are cheaper and scale to more people. A small number of sessions finds a large share of the serious problems in a flow, and the practical rule is to run a handful, fix what they found, and run a handful more, rather than running many at once. The bias is observation: people try harder, read more carefully and complain less when watched, so a problem that appears in usability testing is almost certainly worse in the wild. The tutorial on running a usability test on a conversion journey takes a round of sessions from writing the tasks and recruiting through moderation and severity ratings to the fixes and the second round.

Usability testing is also the right method when there is not enough traffic for behavioural or quantitative methods to say anything, which makes it the backbone of research for low-traffic sites and for new flows that have no visitors yet. A quick complement that reaches beyond the website is for the team to go through the whole journey themselves, by form, phone, email and after the sale, which the tutorial on walking the customer journey yourself sets out, always as a supplement to sessions with real participants.

Triangulation

Each method has a blind spot, and triangulation is the practice of pointing several of them at the same question so that the blind spots do not overlap. Analytics locates and sizes the problem but cannot explain it. Recordings show behaviour but not the thought behind it. Polls and interviews give the thought but cannot say how common it is. Usability testing finds the problem but with a handful of people under observation. A finding that analytics has sized, recordings have shown and a poll has explained is one a team can build a hypothesis on; a finding from any one of them alone is a lead to be checked by another.

In practice triangulation runs in a sequence. Start quantitative, because it is cheap and it ranks: find the step, device and segment with the biggest loss. Go behavioural on that location: review recordings and heatmaps filtered to the visitors who were lost there. Go attitudinal at the same spot: a poll on the step, interviews with people who abandoned it, a read of support tickets that mention it. Where the three agree, write the hypothesis. Where they disagree, the disagreement is itself a finding, and usually means the segments are different. The key performance indicator (KPI) that justifies the effort is a higher share of tests whose hypothesis came from evidence, and a higher share of those that moved the metric the way the evidence predicted.

Illustrative example — not a real result

Triangulating a checkout drop on phones

A homeware retailer sees a checkout completion rate of 38% on desktop and 21% on mobile. The funnel exploration, segmented by device, puts the mobile loss almost entirely at the delivery step: 6,100 mobile visitors reach it in a month and 2,800 continue, against a desktop step rate of 71%. Forty mobile recordings filtered to that step show 23 sessions in which the address form reloads after a postcode lookup and the previously typed fields are blank. An on-page poll on the step, shown after fifteen seconds of inactivity, collects 90 responses in two weeks; 31 mention entering the address twice and 18 mention delivery cost. Support logs for the month hold 14 tickets about the address form. The backlog entry records all four sources, a reach of 6,100 visitors a month, and an estimated 500 to 700 lost orders. The address bug is fixed without a test; the delivery-cost finding becomes a hypothesis for a test that shows the cost on the basket page.

Common mistakes

  • Starting with recordings. Hours of unfocused viewing, no location, no count. Start with the funnel, then watch the sessions at the step it points to.
  • Treating poll results as proportions. A self-selected handful of respondents gives themes, not shares of all visitors.
  • One method, one finding, one test. A single source is a lead, not evidence. Check it with a second method before it costs a test.
  • Asking what people would do. Hypothetical questions produce polite fiction. Ask what they did, and watch.
  • Ignoring the tracking. A step rate that looks wrong is usually a measurement problem. Audit before researching.
  • Researching without a backlog. Findings in a slide deck are forgotten by the next sprint. Record location, size, sources and decision in one place, and take the strongest to the hypothesis tutorial.

Frequently asked questions

Which research method should a team start with?

Start with the funnel report, because it is cheap, it already exists in most analytics setups, and it ranks: it tells you which step, device and segment loses the most people, which is where every other method should then be pointed. Recordings, polls and interviews are far more productive when they are filtered to the step the funnel identified than when they are run across the whole site.

How many session recordings do you need to watch?

Enough to see a behaviour repeat under a specific question, which is usually a modest batch filtered to one step and one segment rather than hours of unfiltered viewing. Decide in advance how many sessions must show a behaviour before it counts as a finding, keep a tally as you watch, and stop when the tally is clear. A handful of dramatic sessions is an anecdote; a counted pattern across a filtered batch is evidence.

Can conversion research be done without analytics?

Yes, and for new flows or very low-traffic sites it has to be. Usability testing, customer interviews and a read of support tickets and sales notes will locate most serious problems in a flow without any quantitative data at all. What is missing is the sizing: without analytics you cannot say how many visitors a problem affects, so the backlog is ordered by severity and judgement rather than by measured cost.

Terms used on this page

Where this page is used

Part of these learning paths

Also related

Sources

  1. Google Analytics Help — Google Analytics Help

Related