Adding self reported attribution to a registration form without wrecking conversion

Somebody proposes adding a dropdown to the registration form. How did you hear about us, twelve options, required. The audience acquisition lead says it will cost registrations. The data person says the answers will be useless. Both are right about the version being proposed.

Self reported attribution is worth having, and the version that works looks nothing like that dropdown. One optional open text question, placed after the point of commitment, coded weekly into the same channels as the rest of your reporting. It recovers the routes no tracking can see, and it does it without touching the conversion rate at all.

Where the question goes

Placement decides which of two costs you pay.

Put the question before the submit button and you get a high response rate, because the person is still in the flow and wants to finish. You also add a field to a form that is already losing people, and every field costs something.

Put it after the commitment step, on the confirmation page or in the confirmation email, and the completion rate cannot move, because the registration is already saved. What you lose is response rate: some people close the tab.

I would take the second trade on any show where registration volume carries a target. A response rate of 62 per cent on a question that is answering something nothing else can answer is a good deal. A response rate of 88 per cent bought with a point of form completion is not, and the arithmetic in a later section makes the size of that point concrete.

One more placement note. The question should be visible on the confirmation page rather than only in the email, because the confirmation page is the moment of highest attention, and email response will be depressed by everything that depresses email response.

Why does a fixed list of options bias the answer?

Because a list is a prompt, and prompts move answers in directions that have been measured for decades.

Krosnick and Alwin published the theory and the test in Public Opinion Quarterly in 1987, using a split ballot experiment on the 1984 General Social Survey. They predicted a primacy effect for visually presented lists, meaning respondents lean towards options appearing early, and predicted it would be strongest among respondents with lower cognitive sophistication. Both predictions held. Applied to a registration form, the third option down gets an advantage that has nothing to do with your marketing, and it gets a larger advantage among people rushing through the form.

The second problem is that a list constrains what can be said. Pew Research Center's methods guidance on writing survey questions, current in 2026, reports an experiment where 58 per cent of respondents chose the economy as the most important problem when it was offered in a list, against 35 per cent who volunteered it unprompted in the open ended version of the same question. A gap of 23 points, produced entirely by whether the answer was on the page.

Your list would contain email, search, social, and a colleague. It would not contain the thing you most want to learn, which is the specific reason somebody in a firm you have never marketed to decided to attend.

So the question is one open text box, optional, with a short label. Something like: how did you first hear about the show? The word first is doing work there, because without it a third of answers will describe the confirmation email.

Coding free text into the same channels

Free text costs somebody twenty minutes a week, and that is the entire operational burden.

Export the week's answers. Read them against a coding sheet that maps phrasings to the same channels your tracked reporting uses, so the two columns can sit next to each other. My colleague sent it, someone in my team shared it, my boss told me to go: peer share. Your email, newsletter, an email from you: email. Saw it on linkedin: platform mention, which needs a rule of its own. Google, searched for it: search. Read it in the trade press or a specific title name: partner and media. An exhibitor invited me, my supplier gave me a code: exhibitor.

Platform mentions need the rule written down in advance, because linkedin could be paid or organic and the respondent has no idea which. The honest treatment is a bucket named for the platform rather than for the mechanism, reported as such, and never silently folded into paid social to make the paid numbers look better.

Anything genuinely uncodeable goes to unclassified, and if unclassified runs above 10 per cent the coding sheet needs a new category rather than more judgement.

Two people should code the same fifty answers in the first week and compare. Every disagreement marks a rule the coding sheet does not yet have, and that fifty answer exercise usually adds three or four rules which then hold for the rest of the edition. After that, one person can code the weekly batch in twenty minutes, because the hard cases were settled at the start.

Keep the raw text next to the code, permanently. The codes are what you report and the text is what you learn from, and the sentence that explains why a whole department registered together will never survive being compressed into a category.

Reading the two columns side by side

Take one edition with 4,800 registrations and a 62 per cent response rate, giving 2,976 answers. Coded, they come out as email 780, peer or colleague 560, search 470, a platform mention 430, exhibitor invite 320, trade press 210, and other 206.

As shares that is email 26.2 per cent, peer or colleague 18.8 per cent, search 15.8 per cent, platform mention 14.4 per cent, exhibitor invite 10.8 per cent, trade press 7.1 per cent, other 6.9 per cent.

Now set the tracked shares beside them: email 26 per cent, paid social 18, organic search 15, direct 24, partner 9, other 8.

Two of those pairs match almost exactly, which is reassuring and uninteresting. The interesting line is direct at 24 per cent tracked, against a self reported picture with no direct in it at all. Instead there is peer or colleague at 18.8 per cent and exhibitor invite at 10.8 per cent, and together those are 29.6 per cent of respondents describing a route that arrives at your form carrying no source value.

That comparison is the argument for instrumenting peer forwarding properly, which B9 builds out with share links and cluster counts. It is also a check on the decomposition of your unknown bucket: if B7's four way split says most of your untagged rows are genuinely direct while the survey says they are colleagues and exhibitors, one of the two is wrong and the survey is usually closer.

What does it cost in completions?

Measure it rather than arguing about it, and the measurement takes two weeks.

Show the question to half the traffic, before the submit button, and leave the other half alone. Over a fortnight, suppose each arm sees 2,400 form starts. The control completes 1,632, which is 68.0 per cent. The test completes 1,608, which is 67.0 per cent. That is one percentage point, and 24 registrations across a fortnight.

Now scale it honestly. If that rate held across a nine month campaign delivering 4,800 registrations, one point of completion is roughly 70 registrations across the edition. Whether that is worth an 88 per cent response rate instead of a 62 per cent one depends on how large your unknown bucket is and what decision is waiting on it.

The placement after commitment sidesteps the trade entirely, which is why it is the default I would ship. Run the split test anyway, once, because it converts a permanent argument into a number your team has measured on its own form.

Where self reported attribution stops

People misremember. Somebody who saw four emails, a retargeting ad and a colleague's message will name the colleague, because that is the touch with a face on it. Self reported answers systematically favour the last memorable human contact and under report paid media that worked without being noticed, which is most paid media doing its job.

The non response is not random either. The 38 per cent who skip the question are likely to be the people in the biggest hurry, and hurry is correlated with the channels that convert fastest. So the survey column has a bias with a known direction and an unknown size.

The question also cannot resolve which of four emails did the work, or which keyword, or which ad set. It gives you routes at the level of a channel, which is why it belongs next to tracked data rather than in place of it, and why the media you genuinely cannot code still needs a withheld control cell as B6 sets out to say anything causal.

Two answers and the whole thing is worth running anyway, because the tracked column will never contain a row for a conversation.

Add the question to your confirmation page this week, as one optional text box, and read the first two hundred answers yourself before you build any coding sheet at all. The categories will write themselves out of what people actually say, and about a fifth of the answers will name something that does not appear anywhere in your acquisition and attribution reporting today.