The Outreach That Backfired and What It Taught

Why this matters

A reactivation push that annoys people does not just fail. It converts quiet lapsed customers, who were still winnable, into people who have now actively decided against you, and in one case it puts that decision in public where every future prospect can read it. The damage is permanent in a way that a campaign producing zero bookings is not.

What follows is one shop's push that went wrong, worked from the first signal through three confident wrong diagnoses to the actual cause. The useful part is not the fix. It is that every one of the wrong reads was about the message, and the cause had nothing to do with the message at all.

The signal, and what it looked like on day three

The shop sent a reactivation message to 412 records pulled from its customer table, in one batch, on a Monday.

By Wednesday the office had logged three angry phone calls. By Friday there were 27 replies asking to be removed. On the following Monday a one-star public review appeared, from a name that matched a record in the batch, saying the shop had been chasing them about a service they had already cancelled. And on the Tuesday, a call from a family member asking why the shop kept writing to someone who had died two years earlier.

The batch also produced 22 completed jobs, which is 5.3% of the 412 records sent. That number is why the owner's first instinct was to declare the campaign a modest success with an unpleasant edge, and it is worth naming as its own trap: a campaign can book work and still be net negative, and the bookings will always be the number somebody reaches for first.

Twenty-seven removal requests against 412 records sent is 6.6%. That rate is the actual signal, along with the composition of what came back: a bereavement call and a public review are not ordinary campaign noise.

First read: the copy was too salesy

The owner's first move was to re-read the message. It opened with a line about missing the customer's business, mentioned the shop's years in the trade, and closed with a discount. It reads as promotional, and that is a real weakness, so the working theory was that the tone had put people off.

What ruled it out. If the tone was the problem, the complaints should be spread roughly evenly across the list, because every recipient got the identical words. When the office matched the 27 removal requests back to the records, they were not spread evenly. They clustered hard, and the cluster had nothing to do with who read the message how. That is the test that kills a tone hypothesis: same message, uneven reaction, so the difference is in the recipients, not in the words.

The copy is still worth rewriting. It was not what happened here.

Second read: the discount cheapened the shop

The second theory was the offer. A discount to lapsed customers can read as an admission that the regular rate was padded, and long-tenured customers in particular can take it badly.

What ruled it out. Two of the three angry calls did not mention price at all. One was about being contacted after cancelling a service; the other was about being contacted twice in one day. The third mentioned the discount only in passing, as evidence the shop was desperate, and its main complaint was also about repeated contact.

The signal to notice here is that a theory has to explain the specific content of the complaints, not just be plausible. An offer theory that cannot account for two of the three loudest reactions is not the cause, however sensible it sounds.

Third read: the channel was wrong

The third theory was that the message had gone out on a channel people experience as intrusive, and that a mailed note or a phone call would have been received differently.

What ruled it out. The batch had gone out on two channels, split by what contact detail each record had. The removal requests came from both, at broadly similar rates. A channel that was itself the problem would have shown a lopsided split.

At this point three message-level theories had failed, which is the moment to stop looking at the message. The remaining variable was the list.

What the list audit turned up

The office lead pulled all 412 records and checked each one against what the shop already knew about it. The results:

The first four categories total 50 records, and adding the 91 not-actually-due gives 141 problematic records out of 412, which is 34% of everything that went out.

Then the decisive number. Of the 27 removal requests, 19 traced back to records inside that 141. That is 70% of the complaints coming from 34% of the list. If the message had been the problem, the complaints would have tracked the list's composition and the problematic subset would have produced around 9 of the 27. It produced 19.

That single comparison is what the diagnosis rests on. The batch did not annoy people because of what it said. It annoyed a specific, identifiable group of people who should never have received anything.

The 91 not-actually-due records deserve their own note, because they are the subtlest failure in the table. The batch was built on a flat rule of no job in 18 months. Some of the shop's service types run on much longer natural intervals, so those 91 customers were on schedule and were told the shop had noticed their absence. Being told you have been away when you have not is a specific kind of insult: it proves the sender does not know you, which is the opposite of what a reactivation message is trying to establish.

The cause behind the cause

The 141 bad records were not an accident of one query. Three structural facts made them inevitable.

The suppression list was a person, not a table. Prior do-not-contact requests lived in the head of the office lead, who filtered lists by hand from memory. She was on leave the week the batch was built. A suppression rule that depends on one person being present is not a rule.

The batch was built from a raw export, not a maintained set. The shop had a dormant-list procedure that removes bad records, but the person building this batch queried the customer table directly because it was faster. The maintained set existed and was bypassed.

Nobody read any of it. Four hundred and twelve records went out with no one having looked at a sample. Reading twenty records at random would have surfaced duplicates and at least one of the categories above in about ten minutes.

The repair, in two parts

Immediate, within 48 hours. The remaining scheduled wave was stopped before sending. Every one of the 27 removal requests was honored the same day and written to a permanent suppression table. The three angry callers and the bereaved family got a call from the owner, not a message, with a straight explanation and no attempt to save the sale. The public review got one short factual reply acknowledging the error and stating what had been changed, posted once and not argued.

Structural, before the next cycle. Five changes:

1. A permanent suppression table. Every do-not-contact request, ever, from any channel, written to a record field that survives exports and is applied to every list build automatically. Nobody may remove an entry.
2. The list may only be built from the maintained reachable set, never from a raw table export. That set already excludes open balances, disputes, returned mail, and sold properties.
3. Deduplicate by property and by contact detail before every send, not once a year. Two contacts in one day from the same shop reads as automated no matter how personal each message is.
4. Overdue is computed against each record's own interval, not a flat month count. This is what removes the 91.
5. A twenty-record read before any batch leaves. One person, ten minutes, actually reading the records the message will land on. This is the cheapest control in the list and it would have caught the problem alone.

Where it landed

The next cycle, run under the new rules, went to 260 records after cleaning - about 37% fewer messages than the 412 in the failed batch.

It produced 3 removal requests, which is 1.2% of the 260 sent, against 6.6% of the 412 sent last time. No angry calls and no review.

It produced 19 completed jobs, which is 7.3% of the 260 records sent, against 5.3% of the 412 sent in the failed batch.

Read those two job numbers carefully, because the flattering version is easy to write and wrong. The booking rate improved by 2 percentage points on the same base of records sent. The count went down, from 22 jobs to 19, because the shop deliberately sent to far fewer people. The shop traded 3 jobs for a clean list, no public damage, and a reusable procedure. That is a good trade, but it is a trade, and anyone reporting this as "the fix increased our bookings" has misread their own result and will be surprised when the next cycle is judged on volume.

What would have changed the conclusion

• If the complaints had been spread evenly across the list, the tone theory would have survived and the message would have been the right thing to fix. The uneven distribution is the entire evidentiary basis for the list diagnosis.
• If the shop had no duplicate records and no prior suppression requests, the remaining problem would have been the 91 not-actually-due records alone, which is an interval-definition failure and a much narrower fix.
• If the 27 removals had come mostly from the correctly-targeted 271 records, the cause would have been something in the message or the offer after all, and the list audit would have been a dead end that still needed doing.
• If the batch had been small, say 40 records, the same three structural failures would have been present but might have produced zero visible complaints, and the shop would have scaled a broken process to 400 records the following quarter with the same result at ten times the volume. A clean small campaign is not evidence of a clean process.

References

• See related: The Dormant List Audit SOP (the maintained set this batch bypassed)
• See related: How to Run a Win-Back Campaign That Does Not Feel Desperate
• See related: How to Decide Which Lapsed Customers to Call First
• See related: The Dormant Customer Definition Worth Setting (why a flat month count produced the 91)