Retargeting without stalking your customers

Remarketing produces the best numbers in almost any account. That is precisely the danger: because the metrics look good, budget rises, frequency climbs, and eventually an ad follows people for weeks with a product they long since bought or deliberately rejected.

Why the numbers look so good

Remarketing reaches people who already know your offer. Some of them would have bought anyway. Those purchases are credited to the channel although it did not cause them.

The term for this is incrementality. It can only be measured with a control group: a share of the audience is deliberately not served, then results are compared. In many such tests a substantial part of the attributed revenue turns out to have been going to happen regardless.

The two settings that decide everything

Frequency cap. Three to five impressions per person per week is the ceiling. Above that, rejection rises faster than revenue.

Membership duration. How long does someone stay on the list? The default is often 30 or 180 days and is almost never right. Take your actual decision window.

Exclusions that are compulsory

1. Buyers. Once someone has bought, they come off. Nothing reads as less professional than being advertised a product you already own.
2. Applicants and suppliers. Somebody who visited the careers page is not a customer.
3. Existing customers in acquisition campaigns.
4. Very short visits. Under ten seconds on site usually means a wrong turn.

Stagger the message over time

Good remarketing tells a sequence, not a repetition. Days 1 to 3: a reminder of the specific product or service. Days 4 to 10: objection handling, so lead time, guarantee, reviews. Days 11 to 21: social proof, references, numbers, experience. Thereafter: sharply reduced frequency, brand only, no product pressure.

Running one creative for 30 days instead produces exactly the feeling people dislike about online advertising.

The legal frame

Remarketing generally requires consent, because it rests on a recognition identifier. That consent must be obtained before the identifier is set, be freely given, and be as easy to withdraw as to grant.

The practical consequence: if your consent rate is 60 per cent, you work with 60 per cent of the data. That is not a flaw in the system, it is the frame. Anyone inflating the rate through confusing design risks the validity of every consent they hold.