Criteo Product Feed

Criteo’s setup documentation names CSV and TSV (comma- or tab-separated) and XML in Google feed format as the supported feed formats, and its troubleshooting guide additionally lists JSON among the structures the parser accepts. For delimited files the rules are strict and mechanical: the header must be declared in the first row, in lower case, without spaces. A header of Image_Link or image link is enough to break field mapping — the import may still run, but the column lands unmapped and every product loses its image.

FeedArc emits the CSV structure for Criteo (UTF-8, comma-separated, header row) because delimited files are the easiest to inspect and diff when something goes wrong. If you already run a Google Shopping XML feed and prefer to point Criteo at it unchanged, that is supported — the trade-off is that any Criteo-specific shaping (out-of-stock filtering, category grouping) then has to happen inside Criteo’s own feed rules rather than at generation time.

Reuse is safe when three conditions hold. First, the source feed already carries google_product_category — true for any working Google Shopping feed, frequently false for Meta-only catalogs, and the field is required on Criteo. Second, the identifiers in the reused feed match what your site’s pixel reports — the OneTag join (covered below) is by product id, and a reused feed with different ids than the pixel events retargets nothing. Third, the feed’s update cadence suits Criteo’s campaign model: a file regenerated weekly may be acceptable for a slow catalog on Google, but retargeting serves ads for products shoppers viewed yesterday, so stale availability directly buys clicks to sold-out pages.

When any condition fails, generate a dedicated Criteo feed from the same canonical catalog instead of sharing the file. Same data, channel-shaped output — the pattern the feed management pillar describes as one catalog, many views.

Most advertising channels define their own category tree. Criteo instead accepts Google’s product taxonomy directly — the same google_product_category values, by ID or full path, that a Merchant Center feed carries. For Criteo the taxonomy feeds the machinery that groups products into coherent dynamic-creative sets and gives the predictive model a category signal for lookalike expansion. For you, it means the categorisation work is done once and serves two channels: the mapping built for Google Shopping satisfies Criteo verbatim.

The failure mode is the store that never sold on Google. Its catalog has internal categories like apparel/mens/shirts and no Google taxonomy at all, so every product arrives at Criteo missing a required field. The fix is a mapping step, not a manual audit: FeedArc’s taxonomy auto-mapping infers the Google category from title, description, and existing category path — the same rule that would prepare the catalog for Merchant Center.

Alongside the Google taxonomy, Criteo supports a merchant-defined category_id used for product grouping — FeedArc’s channel preset lists it among its recommendations for exactly one reason: dynamic creative templates compose ads from groups of related products, and grouping by your own taxonomy ("summer-linen", "clearance-electronics") produces more coherent creatives than grouping by Google’s generic tree alone.

Treat the two fields as serving different masters. google_product_category is for Criteo’s models and is required; category_id is for your creative grouping and is optional. Populating category_id from the store’s native category path is a one-line transform rule and is worth doing at setup rather than retrofitting later — regrouping an established catalog reshuffles which products appear together in ads, which briefly disturbs performance while the model re-learns. If you skip it at launch, add it during a planned quiet period, not mid-campaign.

Criteo pulls rather than accepts pushes: you host the feed at a stable location, Criteo fetches it on the configured schedule. The troubleshooting documentation names HTTP, FTP, and SFTP as the supported protocols, and the setup flow configures scheduled imports — daily is the documented baseline for keeping the catalog current. The operational requirement hiding in that sentence is availability at import time: a feed URL that regenerates on demand and occasionally times out under load will produce intermittent import failures that look random until you correlate them with the fetch schedule.

A hosted feed URL — FeedArc’s model — sidesteps the availability problem by serving a pre-generated file rather than computing the export during the fetch. Whatever hosting you use, the invariant is the same: generation and serving should be decoupled, so a slow generation run delays freshness by one cycle instead of failing the import outright.

The right import frequency is a property of your catalog’s change rate, not a universal constant. Stable assortments with slow price movement are well served by the daily schedule. Fast-moving categories — fashion drops, flash pricing, thin stock positions — justify a tighter loop, because every hour between a sell-out and the next import is an hour Criteo may retarget shoppers toward a product they can no longer buy.

Note what the cadence does not fix: an hourly import of a stale file achieves nothing. The full freshness chain is store → canonical catalog → generated feed → Criteo import, and the slowest link sets the real latency. If the feed regenerates nightly, an hourly Criteo fetch simply re-reads the same file 23 times. Align the generation schedule and the import schedule deliberately, and monitor the timestamp gap between them — the operating-cadence section of the feed management pillar covers this loop across channels.

Criteo surfaces ingestion results under Assets → Product Catalog → Imports, and its troubleshooting guide groups errors into four categories that map cleanly onto layers of the stack: Credentials (authentication to your feed source failed), Downloads (the file could not be retrieved — missing file, network fault, timeout), FTP services (transfer-level failures specific to FTP/SFTP), and Parsing (the file arrived but could not be read as valid CSV, XML, or JSON).

The category tells you which team owns the fix. Credentials and FTP errors belong to whoever manages the hosting; Downloads to infrastructure and availability; Parsing to the feed generator. Only parsing errors ever implicate the file’s content — and a parsing error on a previously-working feed almost always means the generator emitted something structurally new: an unescaped delimiter inside a product title, a BOM, a truncated write. Diff the current file against the last successfully imported one before touching anything else.

Two opposing pressures shape what belongs in a Criteo feed, and getting each right matters when you optimize a product feed for Criteo. Exclude unavailable products: retargeting is memory-based, so a shopper who viewed a product yesterday can be shown it today even if it sold out overnight — paying for clicks that land on a dead product page. FeedArc’s Criteo preset applies exactly one filter by default, excluding rows with out-of-stock availability, and the tighter the import cadence, the tighter that filter tracks reality.

Beyond availability, resist the urge to curate. Criteo’s own channel guidance in the FeedArc registry is to include as many products as possible: the predictive model builds audiences and creative sets from catalog breadth, and long-tail items contribute signal even when they rarely serve. The correct Criteo catalog is close to the full sellable assortment minus unavailable rows — narrower than that, and you are starving the model of its raw material.

Dynamic Ads are assembled from feed attributes, so attribute quality is creative quality. Images matter most: Criteo composites image_link into generated display creatives, and clean product shots on plain backgrounds composite better than lifestyle photography with busy contexts. Titles render on the creative itself, so front-load the identifying words — brand, product type, key attribute — the way you would for any channel where titles truncate. sale_price alongside price lets the creative render the strikethrough discount framing, which is the single highest-leverage attribute pair for promotion periods.

These are the same title, image, and pricing disciplines covered in depth in the product feed optimization guide — Criteo simply consumes them through a different renderer. The channel-specific addition is grouping: coherent category_id groups produce coherent multi-product creatives, as covered in the taxonomy section above.

FeedArc’s Criteo channel definition encodes everything this guide describes as an executable preset rather than a documentation page. On every export it emits the CSV structure with the lower-case header row, carries currency as a dedicated column (defaulting from your store region, EUR for the platform’s French home market), populates google_product_category through the same taxonomy auto-mapping that serves Google Shopping, applies the out-of-stock filter before a single row is written, enforces the 150-character title and 5,000-character description caps by truncation at generation time, and validates every row against the required-field list before the hosted feed URL updates.

The id column passes through from your canonical catalog untransformed and stays stable across regenerations, so the OneTag join survives every export. The result is that Criteo’s fetch always finds a pre-generated, pre-validated file — the generation-versus-serving decoupling the ingestion section argued for.

The Criteo feed is one output of the same pipeline that shapes your catalog for every other destination — the same canonical product data, transformed per channel, validated per channel’s rules, and served on per-channel URLs. Adding Criteo to an existing FeedArc setup is configuration, not integration work: select the channel, confirm region and currency, review the preset’s filter, and paste the generated URL into Criteo’s catalog setup. The taxonomy mapping, if it exists for Google Shopping, is reused; if it does not, the auto-mapping rule builds it.

That is the practical payoff of the feed-versus-catalog distinction this guide opened with: because the feed is just a shaped view of the canonical catalog, the marginal cost of one more channel approaches zero while correctness stays centralised. The commercial details — plans, sync frequency by tier, and the setup walkthrough — live on the Criteo channel page.