Marketing teams are comfortable measuring clicks, opens and conversions. They are far less comfortable measuring anything that arrives in a customer's hand. The usual result is that physical touches are either dismissed as unmeasurable or defended with stories rather than data.
That is a missed chance. A physical touch such as a handwritten card is, from an analytics point of view, one of the easiest interventions to test properly, because it is small, deliberate and deliverable to a chosen group. This article explains how to set up a clean experiment.
Most marketing channels are hard to isolate. A person sees an ad, an email and a post, and credit gets divided by a model. A card is different:
It goes to a specific, chosen list. You know exactly who received one.
It arrives at a known time. Delivery tracking gives you a date.
It is not mixed into a broad campaign. Little else in your marketing resembles it.
The volume is small enough to randomise. You can hold a control group back without losing meaningful revenue.
A useful test begins with one specific, falsifiable statement. For example:
"Customers who receive a handwritten thank-you card within a week of their first order are more likely to place a second order within 90 days than similar customers who do not."
Notice that it has a defined audience (first-time customers), a defined action (a card within a week), a defined outcome (a second order) and a defined window (90 days). Without all four, the result will be arguable.
Pick the metric that reflects what you want the card to do, and commit to it before you start. Reasonable options include:
repeat purchase within a set window
renewal or retention at a set date
reply or response rate to the next message
referral count
average order value on the next order
Secondary measures are fine, but only one should decide whether the test was a success. Choosing the best-looking metric afterwards is how teams fool themselves.
The most common flaw in a "card test" is a biased comparison. If the sales team picks which customers get a card, they will naturally choose the ones they like, and those customers were more likely to reorder anyway.
Use a random split:
List all eligible customers for the period.
Assign each a random number.
Send cards to those below the median, and none to those above.
Record the assignment before any cards go out.
If your volume is low, repeat the test across several weeks and combine results, rather than testing on too few people at once.
You cannot know the effect size in advance, and you should be wary of anyone who claims to. What you can do is estimate the sample you need to detect a difference of a size you would care about. If your baseline repeat rate is 20% and you would only change your practice for a gain of five percentage points, a standard two-proportion calculation will give you the group sizes needed. For small gains, that can run into the hundreds per group. If your volume cannot support that, say so in advance, and treat the result as indicative rather than conclusive.
Beyond the random split, a few design choices make the results cleaner:
A trackable element. A short URL, a code to quote, or a QR code on the card shows whether the person engaged directly. It captures only part of the effect, but it is a useful extra signal.
Delivery data. Use the posting date and tracking to line up the card's arrival with the customer's next action.
A consistent message. Do not change the wording mid-test.
No overlapping campaigns. Avoid launching a promotion to the same group during the window.
If you use a service that posts handwritten cards, such as MailMyCard, the cards are written with a real pen on a plotter and each is posted with tracking, which gives you the delivery dates this kind of analysis needs. It also supports a QR code that opens a page of photos, video or a voice note, which can double as an engagement signal.
Survivorship bias. Only counting customers who stayed long enough to receive a card.
Peeking. Stopping the test early because it looks good after three weeks.
Counting the wrong customers. Including returns, test orders or staff.
Ignoring cost. Compare any lift in revenue to the cost of cards and postage, not just conversion.
Over-generalising. A result in one customer segment does not necessarily hold in another.
If the card group outperforms the control by an amount larger than random variation would explain, and the lift in gross margin exceeds the cost of the cards, you have a case to extend the programme. If the difference is small or inconsistent, that is also a useful result: it tells you where not to spend.
Either way, record what you tested, how you randomised and what you found. A shared log of experiments, including the ones that failed, is what turns one-off tests into institutional knowledge.
Physical touches do not have to remain a leap of faith. With a clear hypothesis, a single primary metric, random assignment and an honest read of the numbers, a handwritten card can be tested as rigorously as any digital campaign. The discipline is the same; only the channel is different.