Dynamic Retargeting Rebuilt Around Incremental Return

Across the retargeting rebuilds we have run since 2022, the share of platform-reported retargeting revenue that survived a holdout landed between eighteen and forty-six percent. The rebuild starts from that measurement, not from the return-on-ad-spend column, and in three of those engagements the recommendation was to spend materially less on this channel and put the difference into prospecting. The holdout comes first: A ten percent ghost-bid holdout where account size supports it, otherwise a regional split with matched pre-period revenue. Four to six weeks before we touch structure, because the answer to what should we build depends on how much of the current result is real. Clients occasionally ask to skip this and rebuild first. We do not, because then the rebuild takes credit for whatever the holdout would have revealed, and nobody learns anything they can use next year. Feed and catalogue: Most of the failure we find is upstream of the ad account. Availability fields that never update on digital products, a title field carrying SEO copy rather than a product name, images built for a shop grid at a size that is unreadable in a mobile feed. We rebuild the feed against the current specification, then use custom_label_0 through custom_label_4 for the things the account actually needs to segment on: margin band, licence tier, cohort start date, content rating. Not a second copy of the category tree, which is what those fields usually contain. Product sets by margin, not by category: A set that groups a high-margin annual plan with a discounted monthly one because they belong to the same product family bids identically for outcomes worth very different amounts. We build sets around what a conversion is worth to you. Cross-sell sets are mapped from what somebody already owns to what they buy next, taken from your own repeat-purchase data rather than from a plausible-sounding assumption about upgrade paths. Windows and exclusions: Purchaser exclusion windows come from your measured repeat-purchase interval, not from the thirty days everyone inherits. For a course business selling one enrolment a year that number is far longer; for a game selling seasonal content it is far shorter. Getting this wrong is how accounts end up paying to advertise a product to the people who bought it last Tuesday. What we cut, and usually do: The fifteen-minute cart reminder, which reaches people who are still in the tab. The warm-audience campaign whose audience definition is broad enough that it is prospecting under another name. Any campaign whose defence is that its ROAS looks good, which is the defence of every campaign that harvests demand somebody else created. Not included: Creative production, lifecycle email, on-site personalisation, prospecting campaigns. If the holdout indicates the budget belongs in prospecting, we will say so and scope that separately rather than quietly absorbing it here. Who this is not for: Catalogues under roughly fifty items, where a broad campaign with good creative does the same job without the feed work. Businesses with strong lifecycle email already reaching the same people, where retargeting is frequently buying the second touch on a conversion the newsletter had already earned. And any team that cannot act on a result reading spend less, because that is the finding often enough that commissioning the test without the appetite to follow it is money spent on reassurance.