Attribution and Measurement glossary

Two versions, one variable, random split. The simplest reliable way to learn anything.

The general name for model-based credit assignment.

Systematic over-crediting of one channel by the model you chose.

How quickly credit fades as time passes since the touch.

The delay before a conversion is credited back to its source.

The rule set deciding which touchpoint gets credit. A choice you make, not a fact you discover.

The level of performance you would expect with no intervention.

A survey-based measure of change in awareness, recall or consideration.

Crediting only what actually caused the outcome, rather than what appeared near it.

The family of methods for separating cause from correlation. Underpins every honest incrementality claim.

A model built to answer what would happen if, rather than what happened alongside.

How long after a click a conversion still counts.

The range the true value probably sits in. Report it and half the arguments about tests disappear.

How sure you want to be before calling a result. Ninety-five percent is convention, not law.

Something affecting both the treatment and the outcome, quietly faking a relationship.

The people deliberately kept away from the campaign, so you have something to compare against.

The delay between the click and the conversion. Long lags make weekly reporting misleading.

A platform-run test measuring extra conversions against a control group.

A formal test comparing conversions between exposed and held-out groups.

What would have happened without the campaign. Never observed, always estimated.

The modelled no-advertising scenario an incrementality test measures against.

The delay between something happening and it appearing in your systems.

Information from the outcome sneaking into the model input, making results look better than they are.

Credit assigned by a model trained on your own converting and non-converting paths. Better than rules, and a black box.

Matching users by a known identifier such as a login. Accurate where it works, and it covers less every year.

Comparing the change in a test group against the change in a control group. Cancels out shared trends.

The shape of the data. Look at it before trusting any single summary number.

How big the difference is, as opposed to how confident you are it exists.

The window after someone watched a meaningful chunk of video without clicking.

All credit to the first interaction. Flatters awareness channels and ignores everything that closed the sale.

A structured test using geography as the split. Needs enough regions to survive normal variation.

Switching advertising off in some regions and comparing against the rest. The cleanest test most advertisers can run.

A control method where the platform records who would have seen your ad but shows something else.

A lift study built on ghost ads. Cleaner than a PSA control because the audience is matched by the auction itself.

The bidding version of the same idea, logging auctions you would have won.

Short for incremental CPA.

Conversions that would not have occurred without the ad.

Spend divided by incremental conversions. The real cost per customer, and usually a shock.

New customers the campaign genuinely created.

The extra outcome caused by the advertising, over what would have happened anyway.

Revenue that would not have arrived without the ad. Almost always lower than attributed revenue.

A model estimating the genuinely caused portion of results.

A pre-post comparison that models the underlying trend first. Better, still not a controlled test.

Heavier weight on the final touch than on the first. A last-click model with manners.

All credit to the final interaction. Flatters branded search and retargeting.

Credit split evenly across every touch. Simple, and treats a banner impression as equal to a demo call.

Fitting a straight-line relationship. Good for continuous outcomes like revenue.

Fitting the probability of a yes or no outcome. Good for conversion likelihood.

A general phrase for whichever credit rules a company has settled on.

Statistical modelling of channel contribution using aggregate spend and outcome data. Privacy-proof, slow, and needs years of history.

Comparing similar regions where one gets the campaign and one does not.

The average. Fragile whenever a few large values dominate.

Error introduced by how you collected the data rather than by the data itself.

The same technique under the older name.

The middle value. Almost always more honest than the mean for order values and session times.

The smallest change your test could reliably find. Set it before you start, or you will chase noise.

Short for marketing mix modelling. Back in fashion because tracking got worse, not because the method got better.

The most common value. Useful for spotting a dominant price point or bundle size.

Error baked into the assumptions of a model rather than the inputs.

Several variables at once, testing combinations. Needs far more traffic than most sites have.

Sales that arrive without paid support. Ignore it and every paid channel looks better than it is.

A value far from the rest. One wholesale order can move a monthly average on its own.

The probability of seeing a difference this large if there were no real difference.

The level of paid-driven performance before a change was made.

Weighted credit to first and last touch, the rest split between the middle.

Working out the sample size needed before running the test. Skipping it is why most tests are inconclusive.

Comparing before and after a change with no control group. Confounded by everything else that happened.

Matching users by inference such as device, timing and location. Scales without identifiers and gets some of it wrong.

Showing a public service ad to the control group. Better than nothing, and it still occupies attention.

The gold standard, where exposure is assigned at random. Rare in media because inventory does not work that way.

Fitting a relationship between variables. The workhorse behind most marketing models.

How long before numbers settle in a report. Judging yesterday today is usually judging incomplete data.

Any model where a human set the weights in advance.

Measured change in actual sales, usually via retail or panel data.

How many people the test needs. Low conversion rates and small effects demand enormous numbers.

Increase in branded search volume caused by upper-funnel activity. A useful proxy when sales data is slow.

Adjusting for predictable time-of-year effects before claiming a result.

When the group you measured is not representative. Retargeting suffers from this badly.

Any model giving all credit to one interaction.

Another name for an A/B test, sometimes meaning whole-page rather than element level.

The typical distance from the average in a dataset.

How much a sample estimate is likely to wobble from the true value.

The chance of detecting a real effect if one exists. Eighty percent is the usual target.

Drawing conclusions only from the cases that made it, ignoring the ones that dropped out.

Building a weighted combination of untreated regions to stand in as a control.

Another name for the treatment group.

A general term for the gap between exposure and outcome.

More credit to touches closer to the conversion. Sensible for short cycles, punishing for long ones.

The delay introduced by the tracking setup itself, such as batched server-side events.

The people who saw the campaign.

Adjusting for an underlying growth or decline so it is not mistaken for campaign effect.

Position-based with forty percent each to first and last touch.

How spread out the values are. High variance means you need more data to see anything.

How long after an impression a conversion still counts, with no click involved. Shorter is safer.

Three weighted points: first touch, lead creation and opportunity creation. Common in B2B.