Meta's Incremental Attribution has two settings and teams use the wrong one

Meta's Incremental Attribution exists in two forms and they do completely different things. One lives in reporting under Compare Attribution Settings and changes nothing about delivery. The other is an optimisation setting under the Sales objective and changes who Meta serves your ads to. Teams routinely enable one while expecting the behaviour of the other.

The confusion is understandable, because they share a name and a concept. The consequences are not symmetric.

Measurement mode

This is the reporting view. You add it as a comparison column and Meta re-reports your conversions using its incrementality model, showing what it believes would not have happened without the ad.

Nothing changes in the auction. Nothing changes in delivery. Your ads run exactly as before. The only thing that changes is the number on your screen, and it will be lower, often much lower, particularly for retargeting and for brand campaigns.

That drop is the feature. It is the platform telling you which of the conversions it previously claimed were genuinely caused rather than merely observed. Treat it as a second opinion, run it in parallel for at least one full purchase cycle, and do not let it move budget until you have seen how it behaves across a complete season.

The failure mode here is emotional rather than technical. Somebody adds the column, sees reported conversions halve, and escalates. Brief the team before you enable it, not after.

Optimisation mode

This is a different thing entirely. Under the Sales objective, you can tell Meta to optimise toward incremental conversions rather than all attributed conversions. Now the model is not just reporting differently, it is bidding differently, because the outcome it is trying to maximise has changed.

The practical effect is that Meta stops chasing people who were going to buy anyway. Retargeting pressure drops. Spend shifts toward audiences with genuine uncertainty attached. Reported conversion volume falls and the conversions you do get are more likely to be caused by the ad.

That is a real strategic change and it should be treated as one. Do not switch it on across the account mid-quarter because it sounded sensible in a webinar.

How to test optimisation mode properly

Four conditions.

Run a control. A matched campaign on standard optimisation, same budget shape, same creative, running concurrently. Without it you are comparing to a memory.

Give it a full purchase cycle plus the learning phase. Anything shorter and you are measuring the reset, not the mechanism.

Judge it on business outcomes, not platform CPA. In-platform cost per acquisition will look worse, by design. The metrics that matter are net new customers, blended CAC across all channels, and total revenue. If those move the right way while platform CPA moves the wrong way, the setting is working.

Validate outside the platform. A geo holdout is the only clean read available. Hold a set of comparable regions out entirely, run the test in the rest, and compare total business outcomes between them. It is more work than a dashboard column and it is the only version of this that survives scrutiny.

Why incrementality became the argument

Two structural changes pushed this forward.

Attribution model choice in the Google stack is effectively over. First-click, linear, time-decay and position-based were removed from Google Ads and GA4 in 2023, leaving last-click and data-driven. The long-running debate about which model to pick has no options left to argue over.

And the privacy story everybody planned for did not happen. Google announced in October 2025 that it was retiring the remaining Privacy Sandbox APIs, with deprecation in Chrome M144 in January 2026 and removal in M150 in July 2026. Third-party cookies remain enabled by default in Chrome under a user-choice model. Any 2026 measurement strategy still framed around cookie deprecation or Topics API readiness is two years stale.

What survived is the underlying problem: platform-reported conversions overstate causation, and no amount of tag configuration fixes that. Incrementality is the response, and both platforms are now shipping their version of it.

The limitation worth stating out loud

Meta's incrementality model is Meta grading its own homework. It is a better-informed estimate than last-click, and it is still an estimate produced by the party being evaluated.

That does not make it useless. It makes it one input. Run it alongside a geo test at least once a year, reconcile both against your own order data, and hold the platform number as a directional signal rather than a verdict.

Decide which mode you are actually turning on before the next planning cycle, and write the answer somewhere the whole team can see it.