What a poll can tell you that analytics cannot
A funnel report can say that visitors on phones leave the delivery step far more often than visitors on desktops. It cannot say whether they left because the delivery cost appeared, because the address form failed or because they went to check a measurement. The conversion research methods guide puts polls and surveys in the attitudinal group for that reason: they are the only route to why, and they are the methods most often done badly, because a bad question is easy to ask and its answers look like data.
The value of an on-page poll is specificity. An answer tied to a page and a moment is far more useful than a general survey answer, because it can be set beside the recordings and the funnel for the same step. The limit is who answers. The methods guide notes that respondents are self-selected, so a poll gives themes and language rather than proportions; it says what the reasons are, not how common each one is among all visitors. Its other product is wording: the phrases visitors use for a worry are often better copy than anything the team would write.
Polls in conversion research differ from the other surveys a business runs. A full questionnaire measured across a sample is the subject of the survey and questionnaire design guide; satisfaction programmes such as NPS and CSAT belong to the voice of the customer guide; and "how did you hear about us?" belongs to the self-reported attribution guide. The table shows the conversion research uses and what each is good for.
| Moment | Example question | What the answers are good for | Watch for |
|---|---|---|---|
| On a located step, after a pause | "Is anything missing from this page?" | Missing information and doubts at the step | Asking too early, before the visitor has read the page |
| On leaving a step | "Was there anything missing from this page?" | Reasons for leaving, in the visitor's words | Few answers; those who reply are not typical |
| On a pricing or plan page | "What is stopping you from choosing a plan today?" | Objections to price, terms and fit | Answers about price that are really about value |
| Confirmation page after purchase | "What nearly stopped you buying today?" | Friction the buyer overcame | Buyers rationalise after the event |
| Confirmation page after purchase | "What else did you consider before choosing us?" | Alternatives and comparisons | Mixing it with the attribution question |
Before you start
- A located step. A funnel analysis or research sprint that names the step, device and segment to ask about; a poll on every page asks everyone about nothing in particular.
- The methods guide, read. The conversion research methods guide is the prerequisite; this tutorial applies its attitudinal section.
- Control of the poll's settings. Who sees it, when it appears, how often a visitor can be asked and where answers are stored.
- A notice for respondents. A line that says who is asking, why and how answers are used, agreed with whoever handles data protection.
- A spreadsheet for coding. One row per answer, with the date, step, device and the answer text, and columns for codes added later.
The steps
Start from a located step and a question
Write down the step the funnel located, the device and segment that lose the most people there, and what you need to learn: "Why do phone visitors on the compatibility table leave without adding a part?" Write what the recordings or the funnel already suggest, so the poll can confirm or contradict it. A poll that starts from a step produces answers that can be set beside the other evidence for the same place; a poll that starts from curiosity produces a list of opinions about the whole site.
Choose the moment and the trigger
Decide when the question appears, which as general practice depends on what you want to hear. Time spent on the step catches visitors who are hesitating, after they have had time to read. An exit from the step catches visitors who are leaving, at the cost of fewer and less considered answers. The confirmation page after purchase catches buyers while the decision is fresh. Show the poll only on the located step and device, so that every answer can be read against that place. Avoid moments that interrupt the task itself, such as the middle of a form.
Write one open question about what happened
Ask about what happened, not what the visitor would do. The methods guide's rule is that hypothetical questions produce polite fiction, and the survey design guide sets out the wording traps: a leading questionA survey or interview question whose wording contains or suggests its expected answer, producing measurement bias rather than evidence. contains its expected answer, a double-barrelled one asks two things at once, and unfamiliar terms confuse. "What is stopping you from choosing a plan today?" is open and neutral; "Is our price too high?" leads. Keep it to one question with an open text box, because a list of options can only confirm what the team already believes. Read it aloud to someone outside the team before it goes live. A longer questionnaire needs a fuller check, which the tutorial on pretesting a questionnaire runs with think-aloud sessions and a soft launchReleasing a survey or test to a small share of its audience first, then pausing to read the data for problems before the full launch..
Add a question after purchase
On the confirmation page, ask buyers what nearly stopped them, or what else they considered before choosing you. The nearly-stopped answers are the most direct source of friction findings, as the methods guide says, and the alternatives answer shows what the page is competing with, which the jobs-to-be-done framework treats as the competition a business did not know it had. Keep the attribution question out of this breath: its wording, placement and coding belong to the self-reported attribution guide. Surveys sent later by email follow the consent rules of the email marketing hub, set out in the list building and consent guide; this page covers only the question on the site.
Keep personal data out and say how answers are used
Ask for nothing that identifies the visitor, and say so beside the box: "Please do not include your name, email or order number." People type them anyway, so remove them before answers are shared, as the personal data guide requires for anything that reaches a reporting tool. A post-purchase answer stored against an order is first-party dataData a business collects directly from its own customers, visitors and users. about a known customer and needs the same care. The survey design guide's section on consent, anonymity and data handling states the principles: tell people who is asking and why, collect only what the analysis needs, and note that data protection law, including the GDPR where it applies, adds legal duties such as a lawful basis and a clear notice. This page is educational, not legal advice. When the poll grows into a longer survey sent to a sample, the survey launch checklist gathers this line with the pretest, sampling and consent checks to tick before it goes out.
Limit who is asked and how often
As general practice, ask each visitor at most once in a period the team sets, and never show the poll to someone who has just answered another one. The voice of the customer guide describes contact rules for the same reason: customers asked too often stop answering, and those who keep answering become less representative. A contact ruleA business-wide rule that a customer surveyed or interviewed recently is not asked to take part again for a set period, whichever team wants to ask. sets that limit across every team for surveys and interviews; the poll's own limit applies the same idea to visitors. Exclude visitors whose task the question would interrupt, such as people already typing in a form, and keep the poll off pages where another research tool is asking something else.
Collect until the themes stop changing
There is no response count to aim for; the useful stopping point is when new answers stop adding new themes. Read the answers as they arrive, note each new theme, and stop when a batch of answers adds nothing that is not already on the list, or when the end date of the sprint arrives. Record the start and end dates and any change to the page during the period, because a release halfway through changes what visitors are reacting to.
Code the answers with a codebook
Build a codebook from a sample of answers: a short name for each theme, a one-line definition, what it excludes and an example answer. Then apply it to every answer, allowing more than one code per answer and an "uncoded" category for answers that fit nowhere. The voice of the customer guide treats the comment box as the most valuable part of most surveys and asks for themes checked by reading samples. Where a language model does the coding, the AI research synthesis guide sets the checks: only codes from the frame, excerpts copied exactly, counts taken from the coded table and a blind check of a sample by a person.
Keep the visitors' own words
Copy the phrases that say a worry most clearly into a separate list, exactly as written, with the date and the step. They are evidence for the backlog, and they are the raw material for copy: the anatomy of a landing page guide asks for objections to be answered in the visitor's own language, and the tutorial on turning customer language into page copy takes the list from here to a rewritten page. Never tidy a phrase; a paraphrase loses the reason it was worth keeping.
Record themes as evidence in the backlog
Add each theme to the research backlog in the conversion research hub as evidence for a problem at the step, beside the funnel's reach, never as a share of all visitors. Note which other methods agree: a theme that matches what the recordings showed is evidence; a theme on its own is a lead. Where the theme points at a fix that is uncertain, the hypothesis tutorial turns it into a hypothesisA testable statement of the change to make, the effect expected on a named metric and the evidence-based reason for expecting it. with the step's conversion rate (CVRShare of visits or clicks that convert.) as the metric; where it points at something plainly broken or missing, fix it.
A cycle-parts shop asks what is missing (illustrative)
Cogfennel Cycle Parts asks visitors on its compatibility tables
Cogfennel Cycle Parts is an invented online retailer of replacement parts for bicycles. Its funnel shows that visitors who open a part's compatibility table leave without adding to basket more often than visitors who do not, on pages that see around 6,000 visits a month. The team puts one question on product pages for visitors who have spent some time on the compatibility table, "Is anything missing from this page?", with a line asking people not to include their name or order number, and adds "What nearly stopped you ordering today?" to the confirmation page. Each visitor is asked at most once a month. Over three weeks the analyst reads answers as they arrive and stops when a week's answers add no new theme. A codebook of six themes is built from the first answers and applied to all of them, with a colleague coding a sample blind. One theme stands out in both places: fear of ordering a part that does not fit, phrased as "not sure this fits my bike", "worried about the wrong size" and, from buyers, "nearly didn't order because I couldn't tell if it was the right thread". The theme matches recordings of visitors scrolling up and down the table, so it enters the backlog as evidence with the step's reach beside it, not as a share of visitors. The exact phrases become the wording for a fit-check line above the add-to-basket button, "Not sure it fits your bike? Check your frame size here", which goes to the hypothesis tutorial as the change to test.
Common mistakes
- Polling every page. Answers that cannot be tied to a step cannot be set beside other evidence.
- Asking what visitors would do. Hypothetical questions produce polite guesses; ask what happened.
- Lists instead of an open box. Options confirm what the team already believes.
- Reporting respondents as visitors. People who answer choose to; report themes and wording, not shares.
- Two purposes in one poll. The attribution question has its own guide and its own record.
- Inviting personal data. An answer box with no warning collects names and order numbers.
- Asking the same visitor repeatedly. Fatigue lowers answers and makes the rest less typical.
- Tidying the quotes. The exact phrase is the evidence and the copy; a paraphrase is neither.
- Stopping at a word cloud. Themes become evidence only when they reach the backlog beside the step's reach.
