Why Your Marketing Attribution Breaks on Marketplaces

Lucas: There's a moment in nearly every marketing analyst's week where they stare at a dashboard and just know the numbers are lying to them. Luna: Usually right after a marketplace campaign report comes in. Lucas: Exactly. And the problem isn't that the data is faked - it's that the structure of how marketplaces like Amazon, Instacart, or Walmart.

com report conversions fundamentally breaks attribution models designed for direct to consumer sites. Luna: So not a bug, a feature of the walled garden? Lucas: Partly. But also a structural mismatch.

Let me give you a concrete example from something Instacart shared at their analyst day earlier this year. They disclosed that about 30 percent of orders on their platform are now placed through in-store pickup or delivery from a retailer's own site - not through the Instacart app. Luna: Wait - so if a shopper goes to a grocery store's website and chooses Instacart for fulfillment, who gets the attribution credit? Lucas: That's the problem.

The retailer's site usually runs its own analytics - Google Analytics, Adobe, whatever. Instacart's ad platform only tracks clicks that happen inside its own app or website. So a shopper sees an ad for a brand on Instacart, doesn't click, then later searches for that brand on the grocery site and buys. Instacart sees zero conversion.

The grocery site might credit a search or organic. The brand's internal MMM might credit a TV ad from two weeks ago. Everyone's right, and everyone's wrong. Luna: That's the attribution black hole Lucas was hinting at.

And it's not just Instacart - Amazon has similar issues with what they call 'subscribed and save' orders or bundle deals. Lucas: Right. Amazon attribution - the tool they offer brands to see click to conversion data - only works if the shopper clicks an ad within the Amazon ecosystem and buys within a certain attribution window, usually 14 days. But a huge share of Amazon sales happen through search, not ad clicks.

And Amazon's last-touch window resets every time the shopper clicks another product listing - even if they never saw another ad. Luna: So the window keeps extending, and eventually the conversion gets credited to some other click that might not even be an ad. The brand thinks their Amazon ad spend is more efficient than it actually is. Lucas: Bingo.

And the research firm Analytic Partners published a study last year showing that brands who rely solely on Amazon's built-in attribution see a 20 to 40 percent overstatement of return on ad spend compared to what you'd get from a proper geo holdout test. Luna: Forty percent? That's a whole campaign budget worth of misattribution. Lucas: Yeah.

And it's not malicious - it's just that the platform's incentives are to show that its ads work. The same dynamic plays out on Instacart, Walmart's marketplace, and even smaller players like Fanatics or Etsy. Any platform that both hosts the transaction and sells the ads is going to have a conflict of interest in attribution. Luna: So what's a marketer supposed to do?

You can't just stop advertising on marketplaces - that's where the customers are. Lucas: You can't. But you can stop relying on the platform's attribution dashboards as your source of truth. The most practical fix I've seen is to run clean-room match rate analyses.

You send your sales data - anonymized, with hashed identifiers - into a clean room like Amazon Marketing Cloud or Snowflake, and you match it against the platform's ad exposure data. Luna: And that gives you a real view of incremental lift? Lucas: Not perfectly, but it addresses the biggest blind spot: cross-platform exposure. The clean room lets you see if a shopper was exposed to an ad on Instacart but converted on the retailer's own site, for example.

It's not a full incrementality test, but it's way better than the platform's default last-touch model. Luna: Is that expensive to set up? I know for smaller brands the clean room infrastructure can be a barrier. Lucas: It's gotten cheaper.

Amazon Marketing Cloud is included with a minimum ad spend - I think it's around $50,000 a year? But for smaller brands, there's a lower-tech option: you can do a simple holdout test within the platform. Amazon lets you run a 'brand lift' study where a control group is not shown your ads. It's not perfect because the control group can still discover you through search, but it gives you a directional read on incrementality.

Luna: And for Instacart specifically, I recall they do offer vendor-supplied incrementality tests through their 'Carrot Insights' product. I think they use a geo-based approach. Lucas: They do. And that's actually the gold standard for marketplace attribution - a properly designed geo experiment where you compare sales in regions where you run ads against regions where you don't.

It bypasses all the click-level noise and gives you a clean read on the actual sales lift. But most brands don't run these because they're operationally heavy - you need a data scientist or a good analyst to design the test and interpret the results. Luna: So the takeaway is: treat any marketplace attribution dashboard as directional, not definitive. And invest the time in at least one incrementality study per year, per major platform.

Lucas: Exactly. And if you're a smaller team and can't run a full geo experiment, even something as simple as comparing your brand's share of search on the platform during ad-on and ad-off periods can give you a clue. It's not rigorous, but it's better than taking the platform's ROAS number at face value. Luna: That's a good low-effort sanity check.

I'll add one more: look at your cost per acquisition trend over time. If your CPA is flat but your attributed ROAS is going up, that's a red flag that attribution inflation is happening. Lucas: Really smart. That's the kind of signal you can catch without any special tools.

And on that note - if these marketing analytics conversations have sparked something you've actually used in your own work, it's worth mentioning that listener support is what keeps this show ad-free. People who find value here sometimes chip in at buy me a coffee dot com slash fexingo. No pressure at all, but it does help us keep the research time honest. Luna: And it means we can keep digging into these nuanced topics without worrying about sponsor constraints.

I think that shows in the depth we get into. Lucas: Alright, back to the mechanics. One more thing I want to flag about marketplace attribution that I think is underappreciated: the problem of bundled or 'subscribe and save' orders. Luna: Oh, that's a mess.

How does attribution even work when a shopper buys a bundle of five products and only one was advertised? Lucas: In most platform attribution models, the entire order value gets credited to the ad that drove the click, even if only one item in the basket was promoted. So a brand could advertise a low-margin staple like paper towels, and if the shopper adds a high-margin premium item to the same order, the ad gets full credit for that premium item's revenue. Luna: Which inflates the ROAS for that ad.

And the brand might then over-allocate budget to that product's ads, thinking it's driving way more value than it actually is. Lucas: Right. And the fix there is to request order-level attribution data from the platform - some, like Amazon, provide it through their 'attribution API', but it's not the default view. You have to ask for it or set it up through a third-party analytics partner.

Luna: So the real theme here is that marketplaces require an extra layer of scrutiny. The data is there, but it's not presented in a way that's useful for decision-making unless you know where to look. Lucas: Exactly. And to close out, I think the single most actionable thing a marketer can do this quarter is to schedule a 30-minute meeting with their analytics team to audit their marketplace attribution setup.

Ask one question: 'If a shopper sees our ad on this platform but buys through another channel, would we know?' If the answer is no, you've found your priority project for the next sprint. Luna: That's a great litmus test. And it might also reveal opportunities - like if you discover that a lot of marketplace-exposed shoppers convert on your direct site, you might want to adjust your bidding strategy on the platform to prioritize upper-funnel awareness over last-click conversion.

Lucas: Perfect example. That's the kind of insight you only get when you stop trusting the default dashboard and start asking better questions. I'm Lucas. Luna: I'm Luna.

And if you want to dig deeper, check out the show notes - we'll link to a few resources on clean room setups and geo experiment design.