When an account gets handed to us the temptation is to start adjusting bids on day one, because that looks like progress and it is easy to explain. We do not. The first week of any criteo ads audit is spent reading, not changing, because retargeting accounts fail for structural reasons far more often than they fail for bidding reasons, and a bid change on top of a broken feed only makes the broken part harder to see.
The pattern repeats across the accounts we inherit. Performance looks acceptable in the platform interface and disappointing in the client’s own analytics, and the gap between the two numbers is where the real problem lives. Somebody is counting a view as a contribution, somebody else is counting only last click, and neither party has asked whether the campaign is producing sales that would not have happened anyway. The audit is designed to answer that question before anyone argues about creative.
The catalogue comes first
Criteo is a catalogue product. If the feed is stale, incomplete or wrong, everything downstream is wasted effort, and no amount of clever segmentation will fix it. We check how often the feed updates against how often prices and stock actually change on the site, because a daily refresh on a store that reprices hourly means recommendations are advertising yesterday’s offer. We check the proportion of catalogue items that are actually eligible to serve, since rejected or incomplete items quietly shrink the pool the engine can choose from.
Then we check the boring fields. Image quality and consistent aspect ratio, because a recommendation grid mixing lifestyle shots and cut-outs reads as a mistake. Product titles that were written for internal use and never for a shopper. Missing brand and category attributes, which limits what the engine can generalise from. Out-of-stock items still being served, which spends budget on a click that ends in a dead end and teaches the shopper not to click again.
What a criteo ads audit checks in the tracking layer
Tag coverage is next, and it is rarely as complete as the account team believes. We walk the full journey on desktop and mobile and confirm that each event fires once with the right parameters, and that product identifiers in the events match the identifiers in the feed exactly. A mismatch between the two, often something as small as a prefix added by a new platform version, breaks the match between the visitor and the product, and the symptom is a slow decline in performance that nobody can trace to a cause.
The concrete checks we run in this part of the review:
- Every event fires once per action, with no duplication from a tag manager container that was migrated but never cleaned out
- Product IDs in events match feed IDs character for character, including variant-level identifiers
- Basket and transaction values exclude tax and shipping consistently, so platform revenue can be compared with the back office
- Consent handling behaves as intended, and the drop in matched traffic after a consent banner change has been quantified rather than assumed
- Exclusion audiences for recent purchasers are actually applied, not merely created
Segmentation and frequency
Most inherited accounts run one retargeting audience for everyone who visited. That treats a shopper who abandoned a full basket ten minutes ago the same as somebody who bounced off a blog post three weeks ago, and it makes the aggregate numbers meaningless because two very different populations are averaged together. We split by recency and depth of intent, then look at whether the split reveals that one segment has been subsidising the reported performance of the entire campaign. It usually has.
Frequency is the other half of this. Retargeting has a point past which extra impressions stop persuading and start irritating, and that point varies by category and price. We look at the distribution of impressions per user rather than the average, because a mean of four can hide a tail of users seeing the same product forty times. Capping that tail typically costs almost nothing in conversions and buys back a meaningful share of budget.
Attribution and the incrementality question
A criteo ads audit that stops at platform reporting has not finished. Retargeting reaches people who already intended to buy, so any last-click model will flatter it. We rebuild the numbers under a view-through window agreed with the client, compare the platform figure to what the client’s own analytics and order data show, and document the size of the gap so future discussions start from an agreed baseline rather than two competing screenshots.
Where budget justifies it, we then recommend a holdout test: a randomly withheld share of the addressable audience, held for long enough to cover the category’s purchase cycle, measured on total orders rather than on attributed ones. It is the only method that answers the question everybody actually has, which is what would have happened with no retargeting at all.
What comes out of it
The deliverable of a criteo ads audit is a prioritised list, not a document. Items are ordered by expected impact against effort, each one names who has to do it, and anything that needs a developer is flagged separately so it can enter a sprint rather than sit in an email. Feed and tracking fixes are almost always at the top, because they cost the least and unlock everything else.
Only after those are cleared do we touch bids and budgets. Optimising on top of unreliable data produces confident decisions built on noise, and the account ends up worse than when it was simply left alone.
Keep reading: Criteo Ads · Advertising