Attribution answers the wrong question

Open your ad accounts and add up what each platform claims it earned you. Google says it drove so many orders. Meta claims its share. Your retargeting tool is proud of itself. Total it all, and you'll notice something odd: the sum is bigger than the revenue your business actually made. Everyone is taking credit for the same orders.

That's not a tracking bug you can fix with a better pixel. It's the nature of attribution itself — and once you see it, you can't unsee it. If you're setting budgets off platform-reported numbers, you're not optimizing your business. You're optimizing a story your tools tell about your business.

This isn't a rounding error. Nearly eight in ten marketing decision-makers say at least 10% of their spend is wasted because they can't measure well enough (eMarketer, 2026) — and in a thin-margin business like quick commerce, 10% is the gap between growth and a bad quarter.

Most writing on attribution tries to help you pick the “right” model — first-touch, last-touch, some clever weighted blend — and concludes, safely, that it depends. That entire conversation is a distraction. The problem isn't which model you choose. The problem is what attribution is.

Credit is not cause

Here's the distinction that changes everything, and almost nobody states plainly: attribution assigns credit. It does not measure cause.

Those sound like the same thing. They are completely different.

Picture a customer of your quick-commerce app. She sees a brand video one week, forgets about it, then a few days later gets a craving, opens the app out of habit, and orders. Last-touch attribution hands 100% of the credit to the last thing she tapped — maybe a retargeting ad, maybe nothing — and 0% to the video that put you in her head. Switch to first-touch and you flip the distortion the other way. Every model is just a different opinion about how to divide credit after the fact. None of them knows what caused the order.

And “data-driven” attribution — the AI-powered option everyone's moving to — isn't an exception. Google's own documentation describes it as a probabilistic model: it estimates credit by comparing the paths of people who converted against those who didn't. That's a statistical guess about correlation. It is useful. It is not a ledger of cause and effect, and treating an estimate as a fact is how smart teams talk themselves into confident, expensive mistakes.

What actually happened to tracking in 2026 (it's not what you were told)

For five years the industry repeated one line: the third-party cookie is dying, and attribution with it. Most founders filed it under “my agency's problem” and moved on.

Here's the real story. Google reversed course. It never finished removing cookies from Chrome, and in late 2025 it wound down the very “Privacy Sandbox” tools meant to replace them (industry reporting, 2026). Third-party cookies are still alive in Chrome, which is roughly two-thirds of the market.

So the crisis was cancelled? No — and this is the part worth understanding. Safari (about a third of browsing) and Firefox have blocked cross-site tracking for years. Apple's app-tracking prompt gutted the signal on iPhones. Consent banners, privacy law, ad blockers, and the simple fact that people hop between phone, laptop, and app all erode the data further. So a large, unstable slice of every customer journey is invisible to tracking regardless of what Chrome does. The ground didn't collapse in one dramatic event the way it was sold — it's been quietly subsiding the whole time, which is harder to notice and easier to ignore.

Put the two truths together: your tracking is partial and getting patchier, and even the parts it captures only ever measured credit, never cause. The reassuring number on your dashboard is built on sand, twice over.

The question that actually matters

If “which channel gets credit” is the wrong question, what's the right one? This: which of my spend produced results I wouldn't have gotten anyway?

That property has a name — incrementality. An incremental order is one that happened because of the marketing, not one the marketing merely stood next to. It's the only version of “performance” worth budgeting on, because it's the only one that tells you what would actually change if you spent more, or less.

And in quick commerce, incrementality and attribution disagree violently — because your business runs on habit. A huge share of your orders come from people who were going to reorder regardless. When you retarget those loyal users, or bid on your own brand name, the platform cheerfully reports every resulting order as a conversion it generated. Much of the time it simply stepped in front of a sale that was already walking through the door — and charged you for standing there.

This is why a retargeting campaign can show a dazzling 8× return and grow your business by almost nothing, while a campaign introducing your app to new neighborhoods shows a mediocre reported return and is, in fact, your only real growth. Optimize to the dashboard and you will pour money into the flattering channel that captures demand you already owned, and starve the one that actually creates it. You'll hit your ROAS target and wonder why the business isn't growing.

What we trust instead: three lenses, triangulated

You're not alone in this shift. A majority of marketers now put more trust in independent incrementality testing than in any attribution model, and most are moving toward “triangulated” measurement — combining incrementality, media-mix modeling and attribution rather than trusting one source (eMarketer, 2026). Here's how that works in practice. Since no single number is the truth, we stop hunting for one, and reason from three independent angles.

1. Incrementality tests — the closest thing to proof. For a hyperlocal business, the cleanest test is a geo holdout, and quick commerce is almost perfectly built for it. You take two sets of comparable cities or catchments — similar size, similar demand, similar maturity. You keep your campaign running in one set and deliberately switch it off in the other for a few weeks. Then you compare order volume between them. If the “on” cities genuinely outrun the “off” cities, that gap is real, caused lift — not a model's opinion. Divide your spend by those extra orders and you get your true incremental cost per order, which is often a world away from the CAC your platform reports.

It isn't free of pitfalls: the cities have to be genuinely comparable, you need enough order volume for the result to be meaningful, you must not run a city-specific promo that pollutes the test, and the window has to be long enough to catch delayed effects. But done carefully, it answers the only question that matters — and you don't test everything constantly, you test the big bets before you scale them.

2. Media-mix modeling — the long view. This steps back and uses your aggregate history — spend, orders, seasonality, pricing, weather, festivals — to estimate what's really moving the business over time. Crucially, it needs no user-level tracking, which makes it immune to the entire cookie-and-consent mess. It used to require a Fortune-500 budget; a disciplined small team can now run a lightweight version. It's directional, not surgical — but it sees the channels your tracking has gone blind to.

3. Platform data — a compass, not a map. The in-platform numbers are still useful for steering day to day: which creative is tiring, which city is heating up. Use them for that. Just don't mistake the compass for the territory. We shore up the signal that does exist with server-side tracking and consent-aware setups, recovering real data instead of losing it — while never treating a platform's self-graded report card as the verdict.

The discipline is in combining them. When the holdout, the model, and the platform roughly agree, you can move budget with confidence. When they disagree, that's not an annoyance to explain away — it's the most valuable signal you have. It's the yellow light telling you something is being mismeasured, before you've scaled a bad decision into a bad quarter.

What this means for you on Monday

You don't need to rebuild your whole measurement stack. Three habits separate founders who know their numbers from those who merely report them:

• Treat your best-looking channels with the most suspicion — especially retargeting and branded search, where non-incremental credit hides. A number that looks too good usually is.
• Separate demand capture from demand creation. Capturing people who already wanted you (retargeting loyal users, your own brand name) looks spectacular and grows nothing. Creating demand among people who'd never have found you looks worse on the dashboard and is your actual engine. Fund the engine.
• And before you scale any meaningful budget — a new city, a bigger paid-social line — validate it with a holdout first, not after. Judge it on incremental orders, not reported ROAS.

Attribution will always hand you a reassuring green number. Incrementality tells you whether that green is real — and warns you early when it isn't. We don't ask which ad to thank. We ask which spend actually changed the outcome.

We catch it at yellow.

Sources

1. Wasted spend (≈78% say ≥10% wasted on poor measurement) — eMarketer: Confidence in marketing measurement falters as budgets climb (2026)
2. Google's 2025 reversal / Privacy Sandbox wind-down — Marketing Attribution in 2026: What Still Works (2026)
3. Shift to triangulated measurement (MMM + incrementality + attribution) — eMarketer: Media mix modeling trends 2026