Zurück zum Blog
    Operations AIROASattributione-commerceDTC

    Your ROAS Is Wrong. Operations AI Starts by Admitting It.

    Platform-reported ROAS systematically overstates reality. Here's the math, the gap, and what Operations AI infrastructure does about it.

    Von Nylo Team•

    Your ROAS Is Wrong. Operations AI Starts by Admitting It.

    You know the number that comes out of Meta Ads Manager isn't quite right. The number from Google says something different. Your own margin model, when you have time to actually run it, lands at a third number, usually the lowest.

    And yet, every Monday, somebody scales spend based on the prettiest of the three.

    This page is the long version of that uncomfortable feeling. It's also the most honest argument for why you need real infrastructure above your ad platforms, not just smarter dashboards on top of them.

    If you're a CMO, performance lead, or e-commerce operator at a brand doing 1M to 50M EUR in revenue, this is for you. We name names. We show the math. And we explain what Operations AI infrastructure like Nylo actually does about it.

    The gap nobody quotes you when you sign up

    Here's the cleanest version of the issue, said plainly:

    Platform-reported ROAS is systematically inflated by 30 to 60 percent versus database truth in most DTC accounts we've seen.

    That's not a swing. That's a structural overstatement. Three reasons it happens:

    1. Attribution overlap. Meta claims a conversion. Google claims the same conversion. TikTok claims it too. If you sum up platform-reported ROAS across channels, you're double-counting the customer.
    2. View-through inflation. Meta in particular counts conversions that happened within 24 hours of an ad being shown, whether the ad was clicked or not. For a brand running broad prospecting, that's many conversions Meta will claim that would have happened anyway.
    3. Post-iOS measurement gaps. Since Apple's App Tracking Transparency, web-to-app attribution is partially modeled, not observed. The model is plausible. It's not truth.

    None of this is news to a sharp performance lead. The news is what to do about it.

    What "correct ROAS" actually requires

    The answer isn't "another tool that does attribution." Triple Whale, Northbeam, Polar Analytics, Lifesight, all of them do attribution. They each pick a model (last-click, MTA, MMM-lite) and report a number with more decimals.

    More decimals isn't more truth.

    What correct ROAS requires is infrastructure that:

    1. Pulls the source data, not the pre-aggregated number. Meta's API gives you both raw events and pre-computed ROAS. The pre-computed one inherits all the assumptions Meta made. The raw events let you compute your own ROAS, with your own model, transparently.
    2. Reconciles against first-party truth. Your Shopify (or your CRM) knows the orders that actually happened. The actual revenue. The actual margin. That's the ground truth. Everything else is a claim about that ground truth.
    3. Holds the business model between providers and decisions. When you compute "ROAS", the formula should live in one place, not eleven. When you change it, every report changes. When you add Pinterest, the formula extends, the agents don't have to be retrained.

    This is what we mean when we say correctness is an architecture, not a feature. You can't bolt it on top of an existing analytics stack. You build the substrate.

    Triple Whale, Northbeam, Polar Analytics: why none of them solve it

    We like all of these products and we're going to be direct about what they are and aren't.

    Triple Whale.

    • Strength: fastest dashboard tooling on top of Shopify-centric DTC stacks. Chat interface is competent.
    • What it structurally doesn't do: still inherits platform-reported numbers as the base. The reconciliation against first-party margin happens if you wire it correctly. The default trusts the platforms.

    Northbeam.

    • Strength: MTA model is more transparent than most.
    • What it structurally doesn't do: still measurement tooling, not operations infrastructure. You see the gap. You don't act on it from inside Northbeam.

    Polar Analytics.

    • Strength: clean visual dashboards, good for sub-10M EUR brands.
    • What it structurally doesn't do: same architectural family. Dashboards over fragmented data.

    Lifesight.

    • Strength: tries to do incrementality and MMM at a price small brands can afford.
    • What it structurally doesn't do: MMM is a model with confidence intervals. Without execution wired in, it produces a quarterly report, not a daily decision.

    None of these are wrong purchases. They're just measurement tools. Measurement is upstream of operations. Operations AI is the infrastructure that takes correct measurement and turns it into actions.

    What Operations AI does with a wrong ROAS

    The difference is in what happens after the gap is detected.

    With a measurement tool: You see Meta says 4.1x. Your reconciled number says 1.9x. Slack message goes to the CMO. CMO decides. Maybe spend gets cut Thursday. Maybe.

    With Operations AI: Same gap detected. The system already knows your spend rules ("if reconciled ROAS on a channel drops below 2.0x for 5 days, reduce daily budget by 15 percent"). It drafts the action. A human signs off (or the rule auto-fires if you've configured it to). The action is in Google Ads Manager 30 seconds later.

    The gap moves from "insight" to "intervention." That's the infrastructure change.

    Which also means: when your reconciled ROAS is correct, the agent on top is acting on truth. When it's wrong, you're scaling spend on a lie at machine speed. The data infrastructure underneath matters more in the Operations AI era, not less, because the agents are now acting on it.

    What it looks like at a 15M EUR DTC brand

    Real pattern from a brand we know, pseudonymized.

    Before:

    • Meta-reported ROAS: 4.1x. CMO uncomfortable but using it.
    • Internal margin model (run quarterly): 1.9x.
    • Spend decisions: Monday standup, based on platform numbers.
    • One channel got over-scaled in Q2 last year because Meta's number was 30% inflated relative to truth. The brand realized it 6 weeks later.

    After Operations AI infrastructure in place:

    • Source events pulled from Meta, Google, TikTok, Shopify directly. Reconciled against database orders + margins daily.
    • Single reconciled ROAS per channel. The CMO knows which number to defend in the board meeting.
    • When a channel's reconciled ROAS drops below the threshold the team set, the system drafts a budget reduction and pings the performance lead. Action is one click away.
    • Net spend efficiency Q1: +18 percent. Not because the campaigns got better. Because the spend stopped going to channels that were lying.

    When this is worth the lift, and when it isn't

    Honestly:

    Worth it if:

    • You're doing 1M EUR/year+ in ad spend
    • You've ever scaled a channel because the platform said it was good, then regretted it 60 days later
    • You have a margin model that disagrees with platform-reported ROAS and you don't trust either
    • You're getting pushed on attribution by your board or investor and want a defensible answer

    Not yet, if:

    • Sub-500K EUR/year in spend. The lift cost is more than the ROAS-truth recovery.
    • You're pre-product-market-fit. Get to traction first, measurement gets meaningful at scale.
    • You haven't connected your Shopify to anything yet. Start with the connection, then the infrastructure.

    Frequently asked questions

    What's the difference between Operations AI infrastructure and Triple Whale? Triple Whale is measurement and dashboard tooling. Operations AI is infrastructure that owns correctness of the underlying data AND can act on it. Different stack position. You can use Triple Whale for visualization and Operations AI for decisions.

    Does Operations AI run MMM? It can. MMM is one model that lives on the business model Operations AI provides. The bigger value is having one substrate that runs MMM, last-click attribution, and incrementality tests with consistent assumptions. Today, most brands run them in different tools with different assumptions and get different answers.

    How fast do you see results? Usually within 4 to 6 weeks of onboarding. First week: the reconciled ROAS shows up and is uncomfortable. Weeks 2 to 4: spend rules get set up. Weeks 5 to 6: first cycle of automated budget adjustments. Brands that are honest about the gap see margin improvement immediately.

    Is the data secure? Yes. Source connections through provider APIs (read-only by default), data lives in your own data warehouse or in our managed instance. SOC2-aligned by design. You own the data, we run the infrastructure on top.

    What if my ROAS is actually right? If your platform-reported ROAS reconciles against your first-party margin truth within 10 percent, you've already done the hard work. Operations AI helps with execution speed and cross-channel decisions, not with measurement repair. Still useful, different value proposition.

    Get a real ROAS number

    The fastest way to find out how wrong your ROAS is: a 30-minute call. We'll look at one or two of your channels live and show you the reconciled view.

    Book 30 minutes with Jasmin

    More on the category frame: What is Operations AI?

    Operations AI is the category we're building at Nylo. Marketing today, operations in every data-driven business area tomorrow. If you're a DTC operator who's frustrated with platform attribution, we want to hear from you.