When Direct Mail Belongs in a Restaurant Marketing Strategy
Richard Yu shows how to test a direct mail marketing strategy for a restaurant with a dated code, a holdout group and a worked cost-per-visit calculation.
A postcard mailed to the households near your restaurant, with one dated offer and a unique code your point of sale records, gives you a more honest count of extra visits than any social ad dashboard, as long as a random slice of those addresses gets nothing. That holds only if the point of sale can record the code and the send is large enough to compare the mailed and unmailed groups; below that, the result is noise.
You get a worked calculation labeled Hypothetical that turns postage, redemptions and holdout visits into cost per incremental visit and a break-even number, plus the offer setup, the random split and a send schedule with owners. It fits a restaurant, cafe or local service business whose repeat guests live nearby and whose slow season is coming.
When a Direct Mail Marketing Strategy Fits a Restaurant
Mail fits when the guests you want back live a short walk or drive away and already know you. A neighborhood bistro in San Francisco whose regulars come from the surrounding blocks fits. A cafe beside a convention center does not, and the section on where this breaks says why.
Of the local marketing channels a restaurant can buy, mail is one of the few where you control who does not get it. A social ad platform picks who sees the ad and reports results it computes itself. With a postcard, you pick the addresses, so you can also pick, at random, the addresses that get nothing, and compare the two groups at your own register.
That comparison only works on households you can match to visits. Build the list from your own records: reservation guest profiles, online ordering and delivery addresses, catering invoices, loyalty sign-ups. When a household on the list comes in and pays, you need to see that visit under its profile whether or not it brought the card. Mail sent to every door on a carrier route can reach strangers, but strangers cannot be matched, so that send needs a different scorecard.
The honest scope of the test is narrow. It tells you whether a dated offer brings known neighbors back more often in a slow window than they would have come anyway. For a restaurant that lives on repeat guests, that is the question that pays the rent. For more on why repeat visits are the number to watch, see Repeat Visits and Renewals: Retention Is the Real Hospitality Metric.
Promo Code Tracking With One Dated Offer
One offer, one end date, one code field. Each extra variable splits a sample that is already small.
The date does two jobs. It pushes the guest to act inside your slow window, and it fixes the measurement window: you count visits by mailed and holdout households over the same days, from the drop date to the expiry date. An offer with no end date has no window, so it has no result.
I'd print a unique code on each card if your point of sale accepts a list of single-use codes. A redemption then ties back to one address, and a code passed around a neighborhood group chat cannot inflate the count. If your system only takes one campaign code, use it; the visit counts come from guest profiles, and the code tells you how many of those visits carried the discount.
Build the code as a named discount before the card goes to print, have a server ring a test check with it, and confirm it appears on the discount report under the campaign name. The tripwire is staff: a server who applies the discount as a manual comp because the guest forgot the code makes that visit vanish from the count. Check the discount report daily for the first week and correct the drift at the next pre-shift lineup.
The card needs less copy than most owners give it:
[Restaurant name], [street address]
[One line on why now: the new dish, the season, the room]
[Offer, for example $[amount] off a check of $[minimum] or more], good for [days and meal]
Good through [expiry date]. One per table.
Your code: [unique code]. Show the card or say the code when you pay.
Book at [booking link] or walk in [hours].
Holdout Groups for Postcard Campaigns
Growth and lifecycle marketing teams in software use a holdout group: a randomly selected slice of the audience deliberately left out of a campaign, so lift is measured as the difference between the contacted users and the untouched ones. Without it, the campaign gets credit for everything the contacted users did, including what they would have done anyway. Raw postcard redemptions have exactly that flaw, which is why the method transfers.
For you it means three steps, done before anything goes to print.
- Add a column of random numbers to the address list in your spreadsheet, sort by it, and cut off the share you will hold out. I'd decide the share before sorting and write it at the top of the file.
- Save the holdout file and lock it. No swapping a favorite regular into the mailed group, no moving the nicer streets.
- Before mailing, count visits per household for both groups over several earlier windows the same length as your offer window.
The third step is the noise check, and it is the decision rule for the whole test. Because the split is random, the two groups should visit at nearly the same rate before the send. The largest gap you find across those earlier windows, multiplied by the number of mailed households, is your noise band in visits. If the campaign's incremental visits come in below that band, the result is noise: do not scale the send, and do not call it a failure either. Rerun it with a bigger list or drop mail for that window.
Then judge the campaign on visits above the holdout rate. Redemptions stay on the report because they set the discount cost, but they are not the result.
Cost per Incremental Visit, Worked Out
Hypothetical: a neighborhood restaurant mails a postcard to its own matched guest list for a slow stretch, offering $10 off with a 21-day window. Every figure is invented to show the arithmetic; none is a benchmark.
| Line | Item | Working | Result |
|---|---|---|---|
| 1 | Households on the matched list | From guest records | 4,000 |
| 2 | Holdout households, random, unmailed | Chosen before the send | 1,000 |
| 3 | Pieces mailed | 4,000 − 1,000 | 3,000 |
| 4 | Cost per piece | $0.25 print + $0.40 postage | $0.65 |
| 5 | Mail cost | 3,000 × $0.65 | $1,950 |
| 6 | Visits by mailed households in the window | From guest profiles, with or without code | 420 |
| 7 | Code redemptions among those visits | From the discount report | 280 |
| 8 | Discount cost | 280 × $10 | $2,800 |
| 9 | Total campaign cost | $1,950 + $2,800 | $4,750 |
| 10 | Visits by holdout households in the window | From guest profiles | 100 |
| 11 | Holdout visit rate | 100 ÷ 1,000 | 0.10 per household |
| 12 | Mailed-group visits expected without mail | 3,000 × 0.10 | 300 |
| 13 | Incremental visits | 420 − 300 | 120 |
| 14 | Noise band from the pre-send check | Largest earlier gap of 0.02 per household × 3,000 | 60 visits |
| 15 | Average check before discount | From the point of sale | $60 |
| 16 | Food and beverage margin | After food and beverage cost, before labor | 70% |
| 17 | Margin per visit | $60 × 70% | $42 |
| 18 | Cost per incremental visit | $4,750 ÷ 120 | $39.58 |
| 19 | Break-even incremental visits | $4,750 ÷ $42 = 113.1, rounded up | 114 |
| 20 | Margin from incremental visits | 120 × $42 | $5,040 |
| 21 | Net after campaign cost | $5,040 − $4,750 | $290 |
Hypothetical: read it from the bottom. Line 13 clears the noise band on line 14, since 120 is above 60, so the lift is real enough to act on. It also clears break-even on line 19, by 120 − 114 = 6 visits, and the campaign keeps $290 of margin after its costs.
Hypothetical: now read it the way a redemption report would. Line 9 divided by line 7 is $4,750 ÷ 280 = $16.96 per redeemed visit, a price that would justify mailing the whole zip code. The holdout puts the real price of an extra visit at $39.58, and it means at least 280 − 120 = 160 discounted visits would have happened without the card. That is at least 160 × $10 = $1,600 of line 8 spent on guests who were coming anyway.
Two decisions come out of the table. First, the offer looks too rich for the lift it buys, so the next test changes only the offer, a smaller discount or a minimum check, and checks whether it cuts line 8 more than it cuts line 13. Second, line 16 is margin before labor. If the slow window already has staff scheduled who would otherwise stand idle, that is the right number; if you add a shift to handle the visits, add its wages to line 9 before you work out line 19.
The stop rule: if line 13 lands under line 14, the result is noise. If it clears line 14 but lands under line 19, mail lost money on that list with that offer, and you stop mailing it.
Where Direct Mail Breaks for Restaurants
The strongest case against this: postage and printing cost far more per person reached than a social ad, and the people who redeem a postcard are mostly regulars who would have come in anyway. You pay to discount your best guests, and the redemption count flatters mail every time.
The second half is right, and the worked calculation shows it in the gap between line 7 and line 13. That gap is the reason for the holdout, and the holdout answers the objection only if addresses are split at random before the send. If you mail the blocks where your best guests live and hold out the rest, the mailed group was always going to visit more, and the lift you measure is your own selection.
The first half compares the wrong unit. Cost per person reached matters when you are buying reach; here you are buying extra visits, so line 18 is the number to beat. If your ad platform runs a lift study with a true holdout, set its result beside line 18. If it reports only clicks and attributed conversions, you are comparing a measured number with a modeled one.
For some businesses the advice flips outright. If your guests are tourists or office workers who live elsewhere, the households near you are not your guests, and your guests are on no list you can mail. If your point of sale cannot record a code at checkout and tie a visit to a profile, neither group can be counted. If the send is so small that the gap between groups sits inside your normal week-to-week swing, the noise check from the holdout section catches it before you trust the result.
The software parallel stops short as well. In an app, a holdout user never sees the message and each session is logged to that user. Your holdout neighbor sees the card on the kitchen counter next door, and a guest who pays cash with no profile is invisible in both groups. That leakage pushes holdout visits up, so a passing result is conservative and a failing one is slightly unfair to mail. A software team can also rerun a holdout next week at almost no cost, while each postcard test costs real postage. You get few runs a year, so I'd spend them on the slow window that matters most.
Send a Test Mailing Before the Slow Season
The owner makes the first move this week: export the matched guest list and run the random split. Everything else waits on that file. The schedule counts back from the drop date; fill in your own dates. If you want to talk through where mail sits in your slow-season plan, you can book an intro call, and the GTM & Operations Plan is the ongoing version: a weekly 30-minute strategy and operations session at $999 per month with a one-year commitment paid up front.
| Task | Owner | By | Done when |
|---|---|---|---|
| Export households with an address and a guest profile ID | Owner | Drop date minus 6 weeks | One file, one row per household, no blank profile IDs |
| Random split, holdout file locked | Owner | Drop date minus 6 weeks | Holdout file saved read-only; the mailed file is everyone else |
| Noise check on earlier windows | General manager | Drop date minus 5 weeks | Largest gap written into line 14 |
| Discount built in the point of sale, test check rung | General manager | Drop date minus 4 weeks | Test redemption shows on the discount report under the campaign name |
| Postcard proof with expiry date and code field | Owner | Drop date minus 3 weeks | Printer proof signed; codes merge correctly on a sample card |
| Pre-shift lineup on keying the code, no manual comps | General manager | Drop date | Each server has rung the code once in training mode |
| Visits counted for both groups, calculation filled in | Owner | Expiry date plus 1 week | Lines 6 through 21 filled; decision written as scale, change the offer, or stop |
I would not mail until the noise check row is done. If the drop date is close and the check is not, move the drop date, because a test without line 14 tells you nothing the redemption report did not.
Frequently Asked Questions
Is direct mail for restaurants worth it compared with social media ads?
On cost per person reached, social ads win; on cost per incremental visit, a randomized mail test gives the number you can verify. Running ads during the mail window does not spoil the test, because a random split exposes both groups equally. It does break if you upload only the mailed addresses as an ad audience, since the mailed group then gets two campaigns and the holdout gets none.
What is the difference between go to market strategy vs marketing strategy?
Go-to-market strategy decides who you sell to, at what price and through which channels. Marketing strategy is the ongoing plan for creating demand within those choices. For a restaurant, the go-to-market decisions are the concept, the neighborhood, the price point and the booking channels. A postcard test belongs to marketing strategy: it works inside choices already made and tells you whether one channel brings those guests back.
Can I print one promo code on every postcard instead of unique codes?
Yes, if visits are matched to guest profiles, because the result comes from profile visits rather than the code. Watch for leakage: a single code shared online shows up on checks from guests who are on neither list. Add their discounts to line 9 so the total cost stays honest, but keep those visits out of line 6, since those guests belong to neither group.
What offer should a restaurant put on a direct mail postcard?
One that costs you less than the margin of the visit it is meant to create. A comped item is worth testing against a dollar discount, because the discount cost becomes what the item costs you to make, not its menu price. Pick an item the kitchen already preps in volume during the slow window, so the offer adds no station and no new 86 risk.

Richard Yu
Richard Yu is the Managing Partner of Rich Ventures, a San Francisco startup consulting and early-stage investment firm, and co-founder and Managing Partner of Equal Parts, a restaurant and cocktail bar in North Beach. He was Director of Strategic Account Management at Lob, founding Head of Solutions Engineering at ChartHop and at Naehas, and in customer success at LeadIQ. He has experience scaling startup post-sales organizations from $1M to $20M+ ARR and advises pre-seed and seed stage founders on go-to-market strategy.