What to Measure Instead of Last Click on Meta

On January 12, 2026, Meta removed the 7-day view and 28-day view attribution windows from Ads Manager and its reporting API, as Bloom Analytics documented. Two months later it redefined a click: click-through conversions require a click on a link, and conversions that follow likes, comments, shares, saves and other non-link clicks moved into a renamed category called engage-through, explained in detail by Jon Loomer. Advertisers watched reported conversions fall in both months. The definitions changed. The purchases didn’t.

That is the most useful fact in any discussion of what should replace last-click attribution. Platform-reported conversions are a definition, and definitions change. If your measurement can move sharply because someone at Meta renamed a category, it was never measuring your business. It was measuring Meta’s reporting.

What last click gets wrong, and what replaced it badly

Last-click attribution gives all the credit for a sale to the final touch before it. That systematically rewards whatever sits at the end of the journey, usually brand search and retargeting, and makes anything earlier look useless. A video that introduced someone to your brand three weeks before they searched for you by name gets nothing.

The common escape from last click was platform attribution: letting Meta report conversions using its own windows, including people who saw an ad and bought later without clicking. That fixed one bias and introduced another. Each platform counts every conversion it can plausibly claim, so add up what Meta, Google and TikTok report and you often get more sales than you made. And as 2026 showed, the platform can change the rules of the count whenever it likes.

The honest question is not which attribution model is right. It is what the ads caused that would not have happened anyway. That is incrementality, and Meta itself has moved in that direction.

What Meta offers, and how far to trust it

Since April 2025, Meta has offered incremental attribution as a setting in Ads Manager. Instead of counting conversions within a time window, it uses a model trained on Meta’s Conversion Lift experiments to estimate which conversions the ads actually caused, and can optimize delivery toward them. Meta also runs Conversion Lift studies, which compare people shown your ads against a holdout group that wasn’t.

These are real improvements, and they share a limitation: Meta is measuring Meta. An independent test by Seer Interactive in April 2025, covering about $1.05 million of spend across six accounts, found Meta reporting 87 percent of conversions as incremental while a cross-check against Google Analytics put the figure at 67 percent, according to the same write-up. Practitioners have also noted the incremental setting suits high-budget accounts with plenty of conversion volume, and can hurt accounts already struggling to exit the learning phase.

Use Meta’s incremental tools, but don’t make them the only judge of Meta. Meta’s delivery increasingly decides who sees your ads on its own, a shift covered in Andromeda and the end of audience targeting as you know it, which makes independent checks on what that delivery produces more valuable, not less.

A measurement setup a mid-size advertiser can run

Large brands buy marketing mix models and run lift studies every month. A company spending a few tens of thousands a month needs something lighter that still answers the causal question. Four layers do it.

Start with one number from your own systems. Total new customers and total revenue, from your store or CRM, against total marketing spend. It can’t tell you which channel worked, but it can’t be inflated by any platform either, and every other measurement has to reconcile to it eventually.

Use platform attribution for decisions inside each platform. Which ad, which audience, which placement. That is what in-platform numbers are good for, as long as you compare like with like and remember the definitions shift.

Run a few holdout tests a year. The simplest is geographic: turn off or sharply reduce a channel in a set of regions that resemble the rest of your market, keep it running elsewhere, and compare total sales from your own systems over a few weeks. It needs no vendor and answers the question attribution can’t. Start with the channels you most suspect of taking credit for sales that would have happened anyway, often brand search and retargeting.

Once you have two or more years of weekly data, add a lightweight mix model. Meta’s Robyn and Google’s Meridian are both open source. They estimate each channel’s contribution from how sales move with spend over time, and they get much more reliable when you calibrate them with the results of your holdout tests.

Setting this up so the layers agree with each other, and keeping it running, is the core of performance marketing measured on what the spend caused.

Count the channels attribution can’t see

Some of what drives your sales never appears in any ad platform: word of mouth, a podcast mention, a trade show, a billboard, a sales rep. A post-purchase question asking how the customer first heard of you catches a surprising amount of it, and costs nothing. Offline activity can also be tested the same way as digital, with regional holdouts, and folded into the same model, as covered in integrating offline channels into digital marketing.

Before you build any of this, it helps to know what you are currently measuring and where the gaps are: which conversions fire, which platforms double-count, which channels have no measurement at all. That inventory is part of an online presence analysis of your channels and tracking.

Measure what changes when you stop

Every good measurement method comes back to one test: what happens to sales when you change the spend. Attribution models describe who touched what before a purchase. Holdouts and lift studies describe what the spend caused. The difference shows up in budget decisions, because the channels that look best on attribution are often the ones that would lose least if you turned them off. That is also why cutting ad costs starts with measurement, a thread through how to lower ad costs.

Last click was never the truth. Platform attribution isn’t either. The closest thing to truth a mid-size advertiser can afford is simple: switch something off, somewhere, and count what you lose.

Frequently Asked Questions

What should replace last-click attribution?

No single model. Use a combination: total revenue and new customers from your own systems as the anchor, platform attribution for decisions within each platform, regular holdout tests to measure what each channel actually causes, and a marketing mix model once you have enough history. Each layer checks the others, which matters because platform-reported numbers can shift when the platform changes its definitions.

What is Meta’s incremental attribution?

It is an attribution setting in Meta Ads Manager, available since April 2025, that estimates which conversions your ads actually caused rather than counting all conversions within a time window. It uses a model trained on Meta’s Conversion Lift experiments and can optimize delivery toward those incremental conversions. It is most useful for accounts with plenty of conversion volume, and its results are still Meta measuring its own ads.

Why did my Meta conversions drop in 2026?

Probably because of measurement changes rather than performance. In January 2026 Meta removed the 7-day view and 28-day view attribution windows. In March 2026 it limited click-through attribution to link clicks and moved conversions after other interactions into a new engage-through category with a one-day window. Check total sales in your own systems before concluding that performance changed, and compare periods using the same attribution settings.

How do I run a geo holdout test?

Choose a set of regions that behave like the rest of your market, then pause or sharply reduce spend on one channel there while keeping it running elsewhere. After a few weeks, compare total sales from your own systems between the two groups against how they normally compare. The gap estimates what that channel caused. Keep the test simple, change one channel at a time, and avoid major promotions during the test window.