Attribution concepts and their limits

0:00 Attribution

A shopper sees your TikTok video, searches your brand on Google a week later, opens an email, and finally buys after clicking a retargeting ad. Who gets the credit? TikTok says TikTok. Google says Google. Your email tool says email. And your finance team says, why do these add up to three sales when we only made one? In this lecture you'll learn how attribution models work, what G A four actually offers, why platforms over-claim, the limits of click-based attribution, and how incrementality tests and marketing mix models give you better answers.

0:40 Why it matters

Why does this matter? Because attribution decides budgets. If you trust last-click alone, you'll pour money into channels that close sales, like branded search and retargeting, and starve the channels that create demand, like video and social prospecting. If you trust every platform's self-reported numbers, you'll believe you made three times the sales you did. Getting attribution roughly right, and knowing its limits, is one of the most valuable things a marketing analyst can do.

1:13 Models answer different questions

Here's the core idea. Every attribution model is a simplification, and each answers a slightly different question. Think of a football goal. Who deserves credit? The striker who scored? The midfielder who made the pass? The defender who won the ball? Last click credits the striker. First click credits the defender. Linear shares credit equally. Position-based favors the first and last touches. Time decay favors recent touches. And data-driven attribution estimates credit algorithmically by comparing paths that did and didn't convert.

1:48 What GA4 offers

What does G A four offer? Data-driven attribution is the default reporting model for key events, with paid and organic last click, and a Google paid channels last-click variant, as alternatives. Google retired first click, linear, time decay and position-based from G A four in twenty twenty-three, but you should still understand them, because other platforms and analysts use them. Acquisition reports use session and first-user dimensions, not multi-touch credit. The attribution area shows model comparison and conversion paths, and lookback windows are configurable in property settings.

2:26 Platform self-attribution

Now platform self-attribution. Every ad platform reports conversions using its own rules, typically including view-through conversions and its own click windows. Meta, TikTok, Google Ads and others will often each claim the same sale. So platform totals usually add up to more than your actual orders. Use platform numbers for in-platform optimization, like which ad set or creative performs best within that platform. Use a neutral source, like G A four, your backend, or a modeling approach, for cross-channel budget decisions.

3:01 Limits of click-based attribution

And the limits of click-based attribution. It can't see impressions that influenced people without a click: a billboard, a podcast, an unclicked video. It loses users who deny consent, switch devices or browsers. And it confuses correlation with causation. Retargeting may get credit for sales that would have happened anyway. Data-driven attribution is smarter about sharing credit, but it's still built on observed clicks. It isn't proof of causality.

3:31 Better answers

So what gives better answers? Incrementality testing asks: what would have happened without this activity? Methods include geo holdouts, where you turn a channel off in comparable regions, conversion lift studies offered by platforms, and on and off tests. They measure causal impact. Marketing mix modeling uses aggregated historical data, like spend, sales, seasonality and prices, to estimate each channel's contribution. Open-source tools such as Google's Meridian and Meta's Robyn have made it more accessible, but it needs enough history, variation in spend, and statistical expertise.

4:08 Paths and model comparison

Two practical tools inside G A four help you have better attribution conversations. The conversion paths report shows the sequences of channels people touched before a key event, so you can see whether social or video usually appears early and search late. And the model comparison report shows how much credit each channel gets under data-driven versus last click. When a channel gains a lot of credit under data-driven, treat that as a question, not a verdict. Is that influence incremental? That's exactly the kind of question a holdout test can answer.

4:48 Example 1: is retargeting worth it?

First example, with illustrative results. A fashion brand sees retargeting with a strong return on ad spend in its ad platform and a solid share of last-click revenue in G A four. The team runs a holdout: a random portion of the retargeting audience is excluded for four weeks. Purchase rates in the holdout are almost identical to the exposed group. The true incremental effect is small. Most of those buyers would have returned anyway. Budget shifts toward prospecting that creates new demand.

5:24 Example 2: branded search

Second example, a business case, also illustrative. A UK coffee subscription brand sees branded search with an outstanding return. It runs a two-week, geo-split pause of branded search in comparable regions, at a time when competitors weren't bidding on the brand name. Most clicks shift to the organic listing, and purchases barely change. The brand reduces branded bids, keeps a small defensive budget, and rechecks quarterly, because competitor behavior can change the answer.

5:56 Watch me do it, part 1

Watch me design a geo holdout on one page. Question: does paid social prospecting create incremental purchases in Saudi Arabia? Units: twelve comparable cities or regions. Assignment: randomly split six and six, test with ads on, control with ads paused. Baseline: eight weeks before, check that test and control trends move together. Duration: four to six weeks, enough purchases per group. Metric: purchases per thousand residents from the backend, not platform-reported.

6:27 Watch me do it, part 2

Then the guardrails and decision rules. No other big changes in either group, and log promotions and holidays, like Ramadan or White Friday, which fall differently each year. The readout compares the change in test against the change in control versus the pre-period, with an uncertainty range. And the decision, scale, hold or cut, is agreed before the test starts, so nobody reinterprets the result afterward. That last line is the one most teams skip, and it's the one that makes tests trustworthy.

7:03 Common mistakes

Common mistakes. Summing platform-reported conversions and presenting them as total sales. Judging upper-funnel channels by last click alone. Switching attribution models frequently and comparing numbers across the switch. Treating data-driven attribution as proof of causality. And running tests without agreeing the decision rule first, then arguing about what the result means.

7:25 Triangulate

A mature approach triangulates. Platform data for tactics within each platform. G A four attribution for directional cross-channel views. And experiments or marketing mix models for strategic budget decisions. When they disagree, that's not failure. It's information about what each method can and can't see. Use this template for every attribution conversation: which decision, which source, which model and window, what it can't see, and how we'll validate with a test.

7:56 Recap and try this now

Recap. All attribution models are simplifications that answer different questions. G A four defaults to data-driven, with last-click alternatives. Platforms self-attribute, so their totals exceed real orders. Click-based attribution can't prove causality. Use incrementality tests and marketing mix modeling to estimate causal impact, and triangulate. Try this now: for one channel you use, write the decision you need to make, the source and model you'd rely on, and a one-page holdout design to validate it.