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

Conversion Rate Optimisation › Conversion Research

Running a Conversion Research Sprint

How to run a conversion research sprint: scope one funnel and question, check the tracking, locate the leak, then watch, ask and test until a backlog is ready.
Tutorialintermediate10 min readUpdated

What a research sprint is for

Most conversion programmes do not lack ideas; they lack evidence about which ideas address a real problem. The conversion research hub describes research in three layers: continuous collection that runs all the time, a sprint of a few weeks that goes deeper on one funnel, and a measurement audit whenever the tracking changes. The sprint is the layer that produces most of a programme's hypotheses, and it is the one most often run badly, as a pile of recordings watched without a question or a survey sent to everyone about everything.

A sprint has three properties that distinguish it from research in general. It is scoped to one funnel and one question, so that every method looks at the same visitors at the same moment. It runs in a fixed order, quantitative first because it is cheap and it ranks, then behavioural and attitudinal methods at the place the numbers located. And it has a defined output: backlog rows, fixed bugs and a readout, not a report. The rows go into the programme's research backlog, where the next sprint can find them. The experimentation loop puts research before hypothesis, prioritisation and test; the sprint is that first step done deliberately.

The order matters because each method has a blind spot, as the methods guide sets out. The table shows which method answers which part of the question in a sprint, and where each step is covered in depth.

Sprint stageMethodWhat it answersCovered in depth by
Trust the numbersTracking checkWhether each step's events fire once, at the right momentTracking verification checklist
LocateFunnel exploration, segmentedWhich step, device and source loses the most peopleFunnel analysis tutorial
WatchRecordings, click and scroll mapsWhat lost visitors did at that stepReviewing recordings and heatmaps
AskOn-page poll, post-purchase questionWhat visitors say was missing or nearly stopped themOn-page polls and surveys
ObserveA small round of usability sessionsWhether people can complete the step, and where they stallUsability testing a journey
ReadSupport, sales and chat logsThe questions and objections that recur at the stepCustomer language for copy
DecideTriangulation and backlog rowsWhich findings are evidence and which are leadsResearch hub backlog

Before you start

  • The methods guide, read. This tutorial assumes you know what each method can and cannot show; the conversion research methods guide is the prerequisite.
  • A brief. One page that states the decision the sprint informs, the question, what is already known, the methods, the timing and how personal data will be handled. The research brief template has the fields, and the research design guide explains how to turn a decision into answerable questions; keep the brief to the one funnel.
  • Access. Rights to create explorations in the analytics property (GA4), to the recording and poll tools if the site uses them, and to the support, chat and sales notes for the period.
  • A backlog to write into. The research hub's backlog table, with its problem, location, evidence, reach, cost, status and decision fields, set up before the first finding arrives.
  • People and time. An owner who decides, an analyst for the funnel, someone to review recordings, a facilitator and note-taker for usability sessions, and a fixed end date for the readout.
  • The privacy position. What the site's consent banner allows each tool to collect, and the rules in the personal data guide for keeping identifiers out of every place research data lands. This page is educational, not legal advice.

The steps

  1. Agree the question, the funnel and the owner

    Write the question as a sentence about visitors, a place and a decision: "Why do mobile visitors who start checkout fail to reach payment, and what should we change first?" is a sprint question; "improve conversion" is not. Name the single funnel the sprint covers, from its first step to the conversion, and the person who will decide what happens to the findings. Fill in the research brief with the decision, the question, what is already known and the end date, and agree who will see recordings and poll answers. A sprint with two funnels or no owner produces two half-finished backlogs and nobody to act on them.

  2. Check the tracking before trusting any step rate

    Every later step rests on the funnel's numbers, so verify them first. Walk the journey with the tag manager's preview and the analytics debug view open, and confirm that each step's event fires once, at the right moment, with the parameters the funnel condition will use, as the tracking verification checklist sets out. Reconcile the conversion event against the business's own count for a closed day. Where the set-up has not been reviewed recently, the analytics implementation audit scores it more widely. A doubled or missing event is fixed before research continues; a step rate built on one measures the bug.

  3. Locate the leak in the funnel report

    Build the funnel in the analytics tool's explorations. Google documents Explorations, listed in this page's sources, as a set of techniques that includes funnel exploration, path exploration, segment overlap and the user explorer; the Explorations tutorial shows how to build them and the funnel analysis tutorial applies them step by step. Segment by device and then by source, read step rates rather than totals, and choose the step where a plausible improvement would recover the most conversions, not simply the lowest rate; the conversion rate calculator does the rate arithmetic. Build a segment of the people who left that step. It is the audience for every method that follows.

  4. Watch recordings and maps at the located step

    Point the behavioural methods at the segment the funnel produced: visitors on the located device who reached the step and left. Write the viewing question first, filter the recordings to that step and segment, set the rule for how many sessions must show a behaviour before it counts, and tally what you see rather than relying on memory. Read the click and scroll maps for the same segment, then compare with visitors who continued. The recordings and heatmaps tutorial sets out the coding sheet, the evidence rule and the consent and masking checks the tool needs before it collects anything.

  5. Ask visitors at the same spot

    Put one open question on the located step, triggered at the moment people hesitate or leave, and add a question after purchase about what nearly stopped the buyer. Ask about what happened rather than what people would do, keep personal data out of the answer box, and read the answers for themes and wording rather than proportions, because respondents choose themselves. The polls and post-purchase surveys tutorial covers the trigger, the wording, how often one visitor is asked and how to code the answers.

  6. Run a small round of usability sessions

    Give a few people who resemble real visitors a task that ends at the located step, written as a goal, and watch them attempt it on the device the funnel pointed to. Moderate without helping and note what they did and said. The methods guide's practical rule is to run a small round, fix what it finds and run another, rather than many sessions at once. Usability sessions find problems nobody knew to ask about, and they are the main method where the step has too little traffic for recordings or polls. The usability testing tutorial covers recruiting, consent, the script and severity ratings.

  7. Read support, sales and chat logs for the same step

    The questions customers already send are research nobody has to commission. Search the period's support tickets, chat transcripts and sales notes for the step, the page and the words visitors use for it, and list the questions and objections that recur. Remove names and contact details before the material is shared or analysed, as the personal data guide requires. Keep the exact phrases: they explain behaviour the recordings showed and they feed page copy later, which the tutorial on turning customer language into copy takes further.

  8. Triangulate the findings

    Lay the findings side by side by location. The methods guide states the rule: where the funnel has sized a loss, recordings have shown the behaviour and a poll or session has explained it, the finding is evidence a team can build on; a finding from one method alone is a lead to check with another. Disagreement is a finding in itself and usually means two segments are behaving differently, so split before concluding. Record which methods saw each finding and what each showed. Resist adding causes nobody observed; a location without a cause goes back for more research rather than forward as a guess.

  9. Write backlog rows and fix the bugs

    Write each finding as a row in the research backlog with the hub's fields: the problem in one sentence about what visitors fail to do, the location, the evidence from each method, the reach from the funnel, the cost as a range, the status and the decision. Then sort. Broken things, such as a field that rejects valid input, an event that fires twice or a price that differs between pages, are bugs: fix them now, without a test, and note the fix in the row. What remains are problems whose fix is uncertain, which become candidates for hypotheses, and problems too small or too costly to act on, which are parked with a reason.

  10. Present the readout and hand over hypotheses

    Close the sprint with a short readout in the what, so what, now what form: what the research found at the located step, why it matters in visitors and conversions, and what happens next. Take the strongest rows to the hypothesis tutorial, which turns each into a problem statement, a change, an expected effect and a metric, and to the prioritisation guide, which checks them against the traffic the step receives. The research KPI is not the number of findings; the hub defines it as the share of tests whose hypothesis came from evidence, so record which rows became tests.

A garden-tools shop's research sprint (illustrative)

Illustrative example — not a real result

Thistlecog Garden Tools researches its mobile checkout

Thistlecog Garden Tools is an invented online shop selling hand tools, mowers and garden storage. Its owner sets a three-week sprint with one question: why do phone visitors who start checkout fail to reach payment? The tracking check comes first and finds that the purchase event fires again whenever a buyer reloads the confirmation page, so the reported mobile conversion rate (CVR) has been overstated; the tag is fixed on day two and the funnel is rebuilt on clean data. Segmented by device, the funnel puts the loss at the delivery step: in the previous four weeks 3,400 phone visitors reached it and 1,700 continued, while on desktop most visitors moved straight on. Recordings filtered to phone visitors who left the delivery step show repeated taps on the delivery-date picker and visitors scrolling back to the basket. A poll on the step collects answers about expecting free delivery and not finding it, and a handful of moderated usability sessions on phones agree: participants read the first date shown, a next-day slot with a charge, as the only option, because the free standard dates sit behind a link labelled "More dates". The sessions also find that choosing a date and then editing the address resets the date. Support logs add questions about whether mowers can be left with a neighbour. The sprint closes with two bugs fixed, the double-counted purchase event and the date reset, and three backlog rows with evidence from at least two methods each: the picker layout, the missing delivery cost on the basket page and the unanswered question about large items. The picker row goes to the hypothesis tutorial first, with a reach of about 3,400 phone visitors a month.

To practise the whole plan, the academy exercise on planning a conversion research sprint for a picture framer takes a made-to-measure configurator from the sprint question to the backlog rows.

Common mistakes

  • Starting with recordings. Without a located step, hours of viewing produce anecdotes. Locate first.
  • Trusting the funnel before checking it. A doubled or missing event creates a leak that is not there.
  • Two funnels in one sprint. Every method has to look at the same visitors at the same moment.
  • Reading poll answers as shares of all visitors. Respondents choose themselves; they give themes and wording, not proportions.
  • Testing bugs. A field that rejects valid input is fixed, not tested, and the fix is noted in the row.
  • One method, one finding, one test. A single source is a lead; check it with a second method first.
  • Collecting personal data by accident. Unmasked form fields in recordings and email addresses in poll answers make copies nobody needed.
  • Ending with a deck instead of rows. Findings outside the backlog are forgotten by the next sprint.
  • No owner. A sprint without someone who decides produces a list nobody acts on.

Frequently asked questions

How long should a conversion research sprint take?

There is no correct length. The conversion research hub describes a sprint of a few weeks, run once a quarter or when a funnel is being rebuilt, and the practical limit is the located step: recordings and polls need enough visitors to pass through it for patterns to repeat, which a busy checkout reaches quickly and a quiet enquiry form does not. Fix the end date before starting, size each method from the site's own traffic, and stop when the backlog rows are written rather than when the calendar says so.

Can a research sprint work on a low-traffic site?

Yes, with a different balance of methods. Where the funnel has too few visitors for recordings or polls to show a pattern, the sprint leans on usability sessions, customer interviews and a read of support and sales logs, which find serious problems without any quantitative data. What is lost is the sizing, so the backlog is ordered by severity and judgement rather than by measured cost, and more of the findings are shipped as considered changes than tested.

What should a sprint hand over at the end?

Three things. Bugs that were fixed during the sprint, each with a note of what was wrong and when it was fixed. Backlog rows for everything else, with the problem, location, evidence, reach, cost, status and decision fields the research hub uses. And a short readout that states what was found, why it matters and what happens next, with the strongest rows already passed to the hypothesis tutorial and the prioritisation guide. A slide deck of observations without rows is not a hand-over.

Terms used on this page

Tools to try

Frameworks

Where this page is used

Part of these learning paths

Sources

  1. [GA4] Get started with Explorations — Google Analytics Help

Related