15 attribution metrics and the question each one answers

What do all these attribution metrics actually mean, and which ones can I compare against each other?

Most attribution arguments are not disagreements about facts. They are two people comparing numbers that were never comparable, because nobody said out loud what each one counts.

So here is each metric with the question it answers and the trap that comes with it. The trap is the part worth reading. A definition without one is the version already on forty other sites, and it is also the version that lets someone put two of these in adjacent columns and draw a conclusion neither supports.

1. Conversions (the platform's column)

Answers: how many of the outcomes I told this platform to care about happened inside its attribution window?

The trap: it is a sum, not a count. It totals every conversion action the account has configured and marked primary, so add-to-cart, newsletter signups and phone calls sit in the same number as sales. Compared against orders it can read as several hundred percent on a perfectly healthy account. This is the most misread number in ecommerce marketing and the fix takes one screen: segment by conversion action.

2. Purchases (the purchase action specifically)

Answers: how many completed sales does this platform believe it drove?

The trap: none, and that is why it matters. This is the one platform metric that is genuinely comparable to your order count, which makes it the only one worth putting in a reconciliation. Getting the comparison right means using this rather than metric 1.

3. Attribution window

Answers: how long after seeing or clicking an ad does a sale still count?

The trap: it is two decisions, not one, and it changes which DAY a sale lands on rather than only whether it counts. Meta's has moved twice toward shorter, and on 12 January 2026 the 7-day and 28-day view windows were removed entirely while a newly created ad set now defaults to 7-day click alone. Existing ad sets keep their old setting, so one account routinely holds several windows at once.

4. View-through conversion

Answers: did a sale follow an ad that was shown but never clicked?

The trap: it produces no visit, so it can never appear anywhere in your own analytics, and its absence there is not a tracking failure. It is also the credit that is hardest to corroborate by any means available to you, which is why reading the click-only figure alongside the total is worth the extra click.

5. Engaged-view conversion

Answers: did a sale follow someone actually watching a meaningful portion of a video ad, without clicking?

The trap: it is easy to file mentally with view-through and it is a stronger claim, because a measured watch is more evidence than an impression. It is still a sale with no visit attached.

6. Modeled conversion

Answers: how many conversions does the platform estimate happened where it could not observe them?

The trap: an estimate can be accurate in aggregate and unauditable for any single sale, and both halves matter. You can reasonably budget against a modeled total. You cannot pull one order out of it and trace it, which means a modeled conversion can never be reconciled against a specific row in your store.

7. Last click

Answers: which touch closed the sale?

The trap: it is unambiguous and it systematically starves anything upstream. A brand search that closed a sale gets everything, and the ad three weeks earlier that caused the brand search gets nothing. It is not wrong, it is answering a narrow question precisely.

8. Last non-direct click

Answers: which non-direct touch closed the sale?

The trap: this is what web analytics acquisition reports default to, and it is winner-takes-all. One channel gets the whole order and every other channel gets zero, which is why it can never agree with an ad platform that credits itself for any touch inside its own window. The disagreement is definitional and no amount of tag fixing closes it.

9. Data-driven attribution

Answers: how much did each touch in the path contribute?

The trap: it produces fractional credit, which is why conversion counts show decimals, and it means per-campaign numbers cannot be checked against per-campaign order counts. Also worth knowing as of today: in Google Ads the first-click, linear, time-decay and position-based models were removed and the conversion actions using them were migrated here, so the model comparison report now compares data-driven against last click and nothing else.

10. ROAS (platform-reported)

Answers: for every unit of spend on this platform, how much revenue does this platform claim?

The trap: both halves of the fraction are the platform's own. It rises when you widen a window, when you add a conversion action, and when you install a server-side integration, none of which moves a sale. A reported ROAS improvement that coincides with a configuration change is a measurement of the change.

11. Blended ROAS and MER

Answers: for every unit of total marketing spend, how much total revenue did the business make?

The trap: it has no attribution in it at all, which is exactly its value and exactly its limit. It cannot be inflated, and it cannot tell you which channel to change. Its best use is as the reality check: if platform-reported ROAS is climbing while this stays flat, what improved was the reporting.

12. CAC and cost per new customer

Answers: what did it cost to acquire a customer who had not bought before?

The trap: it is entirely at the mercy of your definition of new, and the error runs one way. Any bounded lookback silently demotes a loyal buyer to new whenever their last purchase falls outside the window, which inflates the denominator with people you already had and makes the number look better than it is. Compute first-order status over the customer's whole history, and never apply a new-customer filter to your side of a platform comparison without applying it to theirs.

13. Event match quality

Answers: how well can the platform recognise the people in the events I send it?

The trap: read it per event, never per account. The account average is dominated by whichever event fires most, almost always PageView, and PageView matches easily. A healthy overall score sitting on top of a poor Purchase score is the normal shape, and Purchase is the only one that matters for revenue.

14. Deduplication key and event ID

Answers: did the platform recognise my browser event and my server event as the same sale?

The trap: this is a metric people never look at until their conversions inexplicably double. Sending purchases from both sides is correct practice and the shared event ID is the only thing that makes it safe. Meta publishes deduplication feedback per key, which tells you what share of events were actually collapsed and on which identifier, and that is stronger evidence than the absence of a complaint.

15. Incrementality or lift

Answers: would these sales have happened anyway?

The trap: it is the only metric on this list that answers causation, and it is the only one that costs real revenue to produce, because a holdout group is customers you deliberately did not advertise to. Every other metric here describes correlation with a credit rule attached. Any product claiming to prove causation without a holdout, including any attribution product, is describing something else.

The rule that makes all fifteen usable

Two metrics are comparable only when they share three things: a unit, a date basis, and a definition of what counts as a conversion. Sessions against visitors fails the first. A platform's click-date report against your order-date ledger fails the second. Metric 1 against your order count fails the third.

And there is exactly one number in this list with no credit rule inside it: your store's settled orders. Which is why the one test that survives every definitional argument above is the arithmetic one. A platform can legitimately resolve a buyer you never saw, but no platform can resolve an order that does not exist.