Retargeting was one of the first advertising disciplines to be handed over to machines, and for good reason. Deciding what to bid for a specific user looking at a specific product at a specific moment is a calculation problem with far too many variables for a person to do by hand, thousands of times a second. Criteo ads automation handles that class of decision well, and arguing with it usually makes results worse. The difficulty is that automation has kept expanding into decisions that are not calculation problems at all, and the boundary between the two is where most performance is won or lost.
The pattern we see repeatedly is an account that has automated everything it was offered and then cannot explain its own results. Bidding is automatic, creative assembly is automatic, product selection is automatic, and when performance drifts there is no lever anybody understands well enough to pull. The alternative is not manual control, which is worse. It is being deliberate about which decisions belong to the system and which belong to whoever is accountable for the budget.
What the machine does better than you
Bid setting is the clearest case. The system sees signals you do not, updates faster than any human process, and optimises against an objective continuously. Overriding it with manual bids or aggressive caps almost always reduces delivery without improving efficiency, because the cap is a blunt instrument applied to a decision that needed nuance.
Product selection within a feed is the second. Deciding which items to show a returning visitor, in what order, based on what they browsed and what typically sells alongside it, is exactly the kind of pattern matching that machines do well and people do badly. The same applies to creative assembly, where the system builds layouts from feed data and tests combinations at a rate no design team could match. Handing these over is not a compromise, it is the correct allocation of the work.
What criteo ads automation should not be deciding
The decisions worth keeping are the ones where the objective itself is in question, or where the cost of a mistake is not visible in the platform’s own metrics. In practice that list is short and fairly consistent across accounts.
- What counts as a conversion and what it is worth. Feed a system the wrong value and it will optimise faithfully towards the wrong outcome.
- Who to exclude. Recent purchasers, service customers, job applicants and anyone in a support queue should not be followed around the internet by product ads.
- Frequency limits, which the platform has little incentive to keep conservative and which affect brand perception in ways no dashboard reports.
- Which products may be advertised at all, including margin thresholds, stock reality and items the business does not want to be known for.
- Brand safety and placement boundaries, where the acceptable answer is a business judgement rather than a performance calculation.
The feed is the part nobody wants to own
Automated retargeting is only as good as the product data underneath it. If titles are inconsistent, images vary in quality, categories are wrong or availability is stale, the system will make confident decisions on bad inputs and the failure will look like a performance problem rather than a data problem.
This is the least glamorous work in the channel and reliably the highest return. Cleaning titles so they read as a shopper would search, making image treatment consistent across the catalogue, correcting category assignments and keeping stock status genuinely current usually improves results more than any bidding change available in the interface. We audit the feed before touching campaign settings on every account we take over, and in most cases there is material improvement sitting in it.
Measuring automated retargeting honestly
Retargeting reporting flatters itself by design. It is shown to people who have already visited the site and already demonstrated intent, so a share of the conversions it claims would have happened anyway. Judging the channel on last click return produces a number that is technically accurate and practically misleading.
The check worth running is an incrementality test. Hold out a segment of the audience, run the campaign to the rest, and compare conversion rates between the two. It costs some revenue during the test window and it is the only way to know what the spend is actually producing. We run it periodically rather than once, because the answer changes as the site, the audience and the rest of the media mix change around it. No amount of criteo ads automation settles that question for you.
A workable division of labour
The arrangement that works in the accounts we manage is straightforward to describe. The platform owns bids, product selection and creative combination. We own conversion definitions and values, exclusion lists, frequency caps, eligible product rules and the feed. Performance reviews focus on whether the constraints are still correct rather than on second guessing individual bids, because the constraints are where human judgement adds something.
This also changes what an agency should be spending its hours on. Very little of the value in a modern retargeting account comes from sitting inside the interface adjusting things. It comes from the data going in, the boundaries around what the system may do, and the honesty of the measurement coming out. Criteo ads automation is genuinely good at its half of the job. The other half does not automate, and pretending it does is how accounts end up efficient on paper and disappointing in the bank.
Keep reading: Criteo Ads · Advertising