Criteo vs Amazon DSP vs reMKTR: Which Demand Are You Buying?

Most Criteo comparisons put unlike things in the same column

Criteo sells to several different buyers under one brand: retail media for brands and agencies, self-service performance advertising, tooling for media owners, and monetisation for retailers running their own networks. A comparison that writes the company name without saying which of those is on the table is not a comparison.

Three matchups are worth separating.

• Criteo against Amazon DSP. A genuine choice about where your demand originates. This page is mostly about that one.
• Criteo against an open-internet demand-side platform. A choice between commerce-specific data with retailer integrations, and maximum reach and channel breadth across the wider web.
• Criteo against a campaign management platform. Usually a category error. Those platforms sit above the buying layer and manage spend across several networks, so the honest question is what each one replaces in your workflow rather than which is better.

Get the category right first. Half the disagreement in a shortlist meeting comes from two people comparing different things and both being right, and it costs a fortnight to discover that in a spreadsheet.

What Criteo is genuinely strong at

Criteo is a large, long-established commerce media business and the scale is not marketing gloss. Its own site states 200+ retailers and premium media owners, 17,000+ global clients, 740M daily active users and 5B product SKUs in its commerce data, with access from 60+ third-party demand-side platforms and 100k algorithm tests a year.

The 2026 record supports the direction of travel. Criteo published an expanded partnership with Loblaw Advance in August 2026, an onsite retail media partnership with Lidl in May 2026 and a South African expansion with Massmart in July 2026. It published two updates on its integration with OpenAI, in May and June 2026 — an early move into advertising inside AI assistants that nobody else in this set has described publicly. In June 2026 it was named a leader in an analyst matrix for retail media network and monetisation platforms, and in July 2026 it completed a redomiciliation from France to Luxembourg, with the board approving a further step toward a US domicile.

It also advertises incrementality testing on its own advertiser pages, so anyone telling you the platform cannot measure lift is describing an older product. That matters for how you should read the rest of this page: our argument is not that they cannot test, it is a narrower and fairer one about who designs the test and where it is written down.

Who Criteo suits better than us, plainly: brands whose sales are spread across many retailers, where the job is to buy shopper audiences at a long list of networks under one contract. No Amazon-focused service replaces that, and we would not try to sell you one.

What Amazon DSP is strong at

First the disambiguation, because the acronym is overloaded: Amazon DSP here is the demand-side platform for programmatic advertising, not Delivery Service Partner, the courier franchise business.

Its advantage is that the signal and the inventory sit on the same side of the wall. Amazon's own description covers audio, display, online video, streaming television and in-store advertising, with placements including Amazon Originals on Prime Video, live streams on Twitch, live sport, Amazon.com itself, and third-party premium sites and apps. Amazon states you can use the platform even if you do not sell on Amazon, that self-service customers keep full control of their campaigns, and that the managed-service option typically requires a minimum investment of USD 50,000.

The measurement side is the part that settles arguments. Amazon Marketing Cloud gives you event-level datasets from your own campaigns in a privacy-safe clean room, which makes it possible to ask a question no platform report answers on its own: did display cause purchases that sponsored ads would have captured anyway?

The weakness is the mirror image of the strength. If a meaningful share of your customers buy somewhere other than Amazon, an Amazon-only buy leaves them untouched, and no amount of clean-room sophistication changes that.

The question that decides it: where does the demand start

Not features. Where the journey to your product begins.

If a shopper searches, browses, reads reviews and buys entirely inside one marketplace, then the data that predicts their next purchase lives inside that marketplace and so does the inventory that reaches them at the right moment. Buying that audience through a network built on other retailers' data means paying for a longer, weaker path to the same person.

If your sales are genuinely distributed — a grocery multiple here, a big-box retailer there, a marketplace somewhere else — the opposite holds. A single-marketplace buy leaves most of your shoppers unreached, and a network with direct retailer integrations is doing something you cannot replicate from inside one storefront.

So do the arithmetic before the demos. What share of last year's revenue came through each channel, and what share of next year's growth plan does? If one marketplace is more than about two-thirds of the business, the honest shortlist is short. If no channel is above a third, an Amazon specialist is the wrong hire and we will say so on the call.

Write the split down before anyone presents to you. It is remarkably hard to hold that number in your head through two good demos.

The metric that decides it wrongly

Most head-to-heads are settled on return on ad spend, and for a comparison involving prospecting that is close to the worst available choice.

In our own accounts, display prospecting audiences run around two-thirds new-to-brand, and the first visible effect is not in the display report at all — it shows up as growth in branded search on Amazon, which then makes existing sponsored-ads campaigns cheaper because more of the traffic already knows you. That is what our team consistently observes across our managed accounts rather than a measured constant. The consequence is structural: a brand reading only its display dashboard will systematically undervalue prospecting, and a platform comparison judged on attributed return will pick whichever vendor was given the retargeting budget.

Streaming makes the same point more sharply. We set streaming television a return target of about 1.0 and judge it on video completion rate and branded-search lift, because attribution structurally undercounts it — shoppable features only fire for logged-in Prime members, so a meaningful share of streaming-driven demand arrives unattributed. Two operational details follow: premium streaming inventory and free ad-supported channel inventory behave differently and should not share a line item, and the completion-rate read tells you whether the creative works before the sales data is trustworthy. That is standing practice across our managed DSP accounts.

The fix for a comparison is to agree in advance, in writing, which segments are being compared and on which measure. Retargeting against retargeting on cost per acquisition. Prospecting against prospecting on cost per new customer. Streaming against nothing, on its own target. A single blended number across all three is a coin toss dressed as analysis.

The third option: buy the outcome and test it

Both of the platforms above are platforms. Somebody still has to build the audiences, rotate the creative, read the supply report and decide what to switch off. That is the option this page exists to argue for.

reMKTR does that work: Amazon DSP bought as a managed service, on seats we hold ourselves. We are part of Full Circle, a full-service Amazon management company with $500M+ in managed spend across 100+ brands, and we hold 109 live Amazon DSP advertiser seats.

Across 30 of those advertisers in July 2026 the book returned 6.04x on ad spend — the whole set for that month rather than the best campaign — from 78.4 million impressions at a $4.00 CPM, a blended $1.42 cost per click, and a $5.49 cost per acquisition across 57,137 attributed purchases, with 20.1% new to the brand. One window, one advertiser count, both stated so you can weigh it.

Our standard of proof is the differentiator, not the seat. Last-click attribution cannot prove incrementality and never could; holdouts and matched controls can. We reconcile in Amazon Marketing Cloud, where display and sponsored ads stop double-counting each other. A brand running both channels without that reconciliation is usually paying twice for some share of the same orders and calling the total a result.

Two redirects that are more honest than a pitch. Dr. DSP is the right call when you want this capability as a product your own team operates, with the first 30 days free — and note that Orbit, which comes with it, is not a DSP and should not be read as a substitute for the platform. Dr. PPC is the right call when the reconciliation shows the leak is in sponsored search rather than display.

Running a fair pilot between them

If the choice is genuinely open, run both properly rather than arguing about decks.

Split by geography or by product line, never by audience — two platforms chasing the same shoppers will each claim the same conversions and both look successful. Hold the creative concept constant across both, because a creative difference will swamp a platform difference and you will learn nothing about either. Agree the read-out metric in advance and in writing, and make it total sales in the tested area against a comparable untested area, not each platform's own attributed number.

Give it at least ninety days. Display works on a slower clock than search, and a thirty-day read on a new audience strategy mostly measures the learning period.

Watch the calendar as carefully as the design. A pilot that overlaps a tentpole event is being run in a fortnight when every advertiser on the platform is bidding into the same week, and whichever platform happens to hold the retargeting budget will win it. Either start after the event or agree upfront that those weeks are excluded from the read.

Then ask the awkward question at the end: if the untested area grew at the same rate, what was the spend for? A partner who has already designed the answer to that question — and who is willing to have it computed by someone not paid on the result — is a different kind of partner.