Print and direct mail attribution without a click to follow

The mailing went to 30,000 names on Tuesday. Three weeks later registrations are up, the show director is pleased, and the person who has to write the post campaign note cannot say whether the mailing did it or whether the early bird deadline did.

Print and direct mail attribution has a reputation for being impossible, and it is not. It is harder than digital in one specific way, which is that there is no click to follow, and easier in another, which is that you control exactly who receives it. That second property is worth more than the first, and most teams spend their effort on the half that matters less.

CEIR's Attendee Acquisition Study, presented in 2023, documented nearly twenty attendee marketing tactics in use across responding organisers and found that traditional approaches held up: telemarketing, direct mail to the house list, and trade magazine print advertising were all reported as effective ways to drive registration. If those channels are still delivering, they deserve better measurement than a shrug.

Three ways to make paper measurable

A vanity domain. Print an address that exists only on paper, such as tech26.show or yourshow.com/post, and have it redirect into a fully tagged registration url. Every visit is then a print visit by construction. Google's campaign parameter documentation, current in 2026, offers billboard as an example value for the source field, so the vocabulary already anticipates media with no click in it, and the redirect is what lets you use the same tagging convention as everything else, which B1 sets out.

A code per drop. A code entered at registration ties a record to a specific mailing rather than to the campaign as a whole. Make it short, avoid characters that look alike, and print it large enough to read without glasses on a train.

A withheld cell. Take the list you were going to mail and randomly remove a portion of it before the file goes to the print house. Those people get nothing. They are the only reason any of the other numbers mean anything.

One code per drop

The common mistake is a single code for the whole print programme. It saves nothing and it destroys the comparison you will want in six months.

A show that mails three times, at 120 days, 60 days and 21 days out, and prints one code across all three, learns that print produced some registrations. A show that prints three codes learns which distance from doors converts, and can decide next edition whether the 120 day drop earns its postage.

The same applies to publications. One code for the trade magazine and a different code for the association newsletter, even when the creative is identical, because the question you will be asked at renewal is which title delivered rather than whether print delivered.

The vanity domain deserves the same discipline and usually gets none. One hostname per drop is unnecessary, and one path per drop costs nothing: yourshow.com/mail1, /mail2, /mail3, each redirecting to the same registration page with a different campaign value attached. Use a temporary redirect rather than a permanent one, because a permanent redirect gets cached by the browser and by intermediaries, and a cached redirect will keep sending next year's traffic to this year's campaign value long after you have changed the destination.

Two details make the printed address usable. Keep it short enough to type on a phone with one hand, which in practice means no more than about twenty characters after the domain. And test it on the actual print proof rather than on screen, because a lowercase l and a digit 1 are distinguishable in your design tool and identical in the typeface the printer substituted.

The holdout that makes the number real

Here is the design, and it costs nothing except the registrations you deliberately forgo.

Start with 40,000 records on the house list that meet the selection criteria. Randomly withhold 10,000 of them, which is 25 per cent, and mail the remaining 30,000. Randomly is doing work in that sentence: not the 10,000 with the oldest records, not one region, not the ones with missing phone numbers. A random draw, seeded and written down.

After the campaign window, count registrations from each cell regardless of how they arrived. No codes, no vanity domain, no source column. Just the count of people in each cell who ended up registered. The match is on the identifier you mailed against, usually the email address or the record id, and it has to be exact, because a fuzzy match that pulls a few control records into the mailed cell will shrink the measured difference in a direction you cannot detect afterwards.

Suppose the mailed cell produces 960 registrations from 30,000, which is 3.2 per cent, and the withheld cell produces 260 from 10,000, which is 2.6 per cent. The difference is 0.6 percentage points. Applied to the 30,000 who were mailed, that is 180 registrations the mailing caused.

What does a code count actually measure?

Now compare that with what the codes reported. At a redemption rate of 0.8 per cent, the codes and the vanity domain between them accounted for 240 of those 960 registrations.

So the code count says 240 and the holdout says 180. The code count is not wrong, and it is answering a different question. It counts people who used the mailing to register. The holdout counts people who registered because of the mailing. The 60 record difference is people who had the postcard on their desk, typed the address because it was in front of them, and would have registered anyway through email or search.

That gap runs in the same direction almost every time, and it is why a print programme defended on redemption alone tends to look better than it is. It can also run the other way. A mailing that prompts somebody to search for the show and register through organic search leaves no code at all, and the holdout catches it while the code count misses it entirely.

I would report both, labelled, in the same table. Redemption tells the print team whether the creative and the code placement worked. Incremental registrations tell finance whether the postage was worth it. Reporting only one of them starts an argument that the other number would have settled.

How big does the control cell need to be?

This is where most holdouts fall over, because a cell that is too small produces a difference nobody can interpret and everybody argues about.

Work it backwards from the effect you expect to see. To detect a rise from 2.6 per cent to 3.2 per cent at conventional confidence and power, the sample size arithmetic for two proportions gives roughly 7.84 times the sum of the two variances, divided by the squared difference. The variances are 0.026 times 0.974, which is 0.025324, and 0.032 times 0.968, which is 0.030976. Their sum is 0.0563. The squared difference is 0.006 squared, which is 0.000036. So 7.84 times 0.0563 divided by 0.000036 gives about 12,300 per cell.

That looks fatal for a 10,000 record control. It is not, because that figure assumes equal cells. With unequal cells the quantity that matters is the harmonic mean of the two sizes, and for 30,000 mailed against 10,000 withheld that is two divided by the sum of one over 30,000 and one over 10,000, which comes to 15,000. Comfortably above 12,300, so the design holds.

Run that calculation before the file goes to print rather than after the results come back. If the arithmetic says your list is too small to detect the effect you are hoping for, you have three choices: accept a larger control, accept that you are measuring redemption only this edition, or pool two editions of the same drop and analyse them together.

Where this stops

The holdout only works for media where you choose the recipients. A house list mailing qualifies. A trade magazine advertisement does not, because you cannot withhold the magazine from a random quarter of its subscribers, and a regional holdout across every channel at once is B35's subject rather than this post's.

Contamination is the second limit. People forward mail, offices share magazines, and a withheld colleague sitting next to a mailed one sees the postcard on the desk. That leakage biases the measured effect downwards, so the honest framing is that the holdout gives you a floor on the incremental effect rather than a point estimate.

The third limit is patience. A mailing landing 120 days out produces registrations for months, and a holdout read at three weeks will understate it badly. Fix the measurement window before you start, tie it to a days out value rather than a calendar date, and resist reading the result early.

Everything above still leaves registrations that arrived with nothing attached, which is a larger bucket than print on most shows and needs its own decomposition in B7. And for the media you genuinely cannot code, the fallback is asking the registrant directly, which B10 handles with the survey design problems that come with it.

Start with the next drop. Before the file leaves your hands, pull a random 25 per cent into a suppression list, name it after the drop, and note the date. That single step costs you a few hundred registrations and buys you the only defensible number your acquisition and attribution reporting will ever have for print.