Attribution models explained: who gets credit for a sale, and what they all miss

Take one made-up customer:

1. Day 1: clicks your Instagram ad, looks around, leaves.
2. Day 5: searches Google for your product type and clicks your search ad.
3. Day 6: clicks a link in your email and buys a $60 order.

Three things touched this sale, and the $60 has to be credited somehow. An attribution model is the rule that decides how. The rule changes which channel looks best, which changes where budget goes. That's why the choice of model matters more than it seems.

Single-touch models give all the credit to one interaction. Multi-touch models split it. Here's the $60 order under the classic rules:

• Time decay gives more credit to clicks closer to the purchase, so the email gets most and the Instagram ad least.
• Data-driven attribution learns the split from your own data, by comparing the paths of people who bought with the paths of people who didn't. Shao and Li described one of the first versions in 2011. Google Ads now uses it as the default for most conversion actions.
• View-through attribution adds credit for ads someone saw but didn't click, within a short window. Meta counts views by default; Amazon's Sponsored Products reports don't.

Google has retired the first-click, linear, time-decay and position-based models in Google Ads; only last click and data-driven remain. They still show up in other tools and in most explanations of attribution, so they're worth knowing.

Every model also has a window: how long after a click or view a purchase still counts. A purchase on day 8 counts in a 14-day window and disappears in a 7-day one. The defaults differ by platform:

You can change the window on Meta and Google. Longer windows credit more sales to ads; shorter ones credit fewer. Neither changes how many sales the ads actually caused.

Each ad platform runs its own attribution and gives its own ads the credit. A made-up brand sells $10,000 on its website in a month. Meta reports $6,000 of purchases from its ads and Google reports $5,500. Together that's $11,500, more than the brand sold. The customer above would be counted by both: Meta for the Instagram click, Google for the search click.

The opposite happens when sales land somewhere a platform can't see. Meta's pixel can't run on Amazon, so a shopper who sees an Instagram ad and buys on Amazon counts for nothing in Meta's report. If they click a Sponsored Products ad on the way, Amazon gives that ad the whole sale. The attribution gap calculator sets the reports side by side.

Every attribution model, data-driven ones included, answers "which ads did buyers touch?" The question that matters for budget is different: "which sales wouldn't have happened without the ads?" That's incrementality.

The two differ because ads tend to reach people who were already likely to buy. Retargeting shows ads to past visitors; search ads show up when someone types your brand. Those people buy often, so ads near them collect a lot of credit, ads or no ads. Comparing buyers with non-buyers doesn't fix this, because people who see more ads are different from people who see fewer.

Researchers have tested this directly. In large experiments at Facebook, Gordon and colleagues found that common non-experimental methods often failed to match the experiment's answer, sometimes by a lot. Lewis and Rao showed why it's hard: ad effects are small next to the normal swings in what people buy.

The reliable way to measure what ads cause is an experiment with a holdout group: some people or regions don't see the ads, and you compare. For ads that sell on Amazon, a geo test works because Amazon reports where orders ship. See how lift is calculated.

• For day-to-day decisions inside one platform, use its default and keep it fixed. Bidding learns from it, and changing models mid-flight makes trends unreadable.
• Don't compare ROAS across platforms. Different windows and rules make the numbers incompatible. Read what ROAS measures first.
• Check attribution against a test, then keep the ratio. A holdout test tells you how many sales a channel caused for each sale it was credited with. Use that ratio to read the reports between tests, and repeat the test when spend or creative changes a lot.
• Use a model for the big picture. Marketing mix models estimate each channel's effect from spend and sales over time, without tracking people. See MMM vs. testing.

This is where attribution breaks down most. Meta can't see Amazon purchases. Amazon Attribution tags only count shoppers who click and buy within 14 days. Shoppers who see the ad and search Amazon later are counted by neither, or credited to a Sponsored Products click on your brand name.

The result is that Meta ads often look weak and Amazon's own ads look strong, whatever is really happening. The halo effect calculator estimates the uncredited Amazon sales, and Facebook ads for Amazon covers setup and measurement.