Attribution in paid media: models, biases and best practices

Shortening an attribution window lowers the conversions you see without a single sale disappearing. On Google and Meta alike, the same setting also changes which conversions bidding and delivery take into account.

On 12 January 2026, Meta removed the 7-day and 28-day view attribution windows from its reporting API (Meta for Developers). Dashboards built on those windows lost part of their conversions overnight, with no change to the campaigns. In March, Meta redefined the click itself for website conversions: only link clicks now count, and conversions that follow a share, a save or a like move to a category called "engage-through".

None of these decisions made a single sale disappear: they changed how Meta counts conversions, and therefore the figures you use to judge your campaigns.

Model, window and optimisation event

When a conversion count looks odd, three settings can explain it, and each answers a different question. The optimisation event is the conversion the campaign is optimised on. The window sets how long after a click or a view a conversion can still be credited to an ad. The attribution model splits the credit when several ads from the same platform preceded the conversion.

	Google Ads	Meta (Ads Manager)
Optimisation event	primary conversion actions, counted in the "Conversions" column; secondary actions stay observation-only in "All conversions"	performance goal and conversion event of the ad set
Window	"Click-through conversion window": 30 days by default, from 1 to 90 days; 3 days for engaged views, 1 day for view-through	ad set attribution setting: 7-day click and 1-day view by default
Model	"Data-driven" by default, or "Last click"	standard or incremental attribution, with no model choice comparable to Google's

Guides that still recommend a linear or position-based model are out of date. Google has retired those models, along with time decay and first click, and the conversion actions that used them were moved to data-driven attribution (Google Ads Help).

These settings also drive bidding

On Google Ads, the attribution model also affects bid strategies that use data from the "Conversions" column, Google's help centre states. The same documentation says that Target CPA and Target ROAS count and optimise conversions over whichever window you pick (conversion windows). At Meta, the attribution setting is used both for reporting and to optimise ad set delivery (Meta Business Help Centre).

Take a sofa maker whose customers buy, on average, three weeks after their first click. With a 7-day click-through window, a click older than seven days no longer receives any credit: the sale only enters reports and Smart Bidding data if another ad click took place in the week before it. Generic queries, typed early in the decision, then look less profitable than brand queries typed the day before purchase, and the bid strategy bids less on them. Conversely, a 90-day window on an impulse purchase credits the ads with sales that a two-month-old click hardly explains any more.

On Meta, an ad set set to 1-day click rather than 7-day click and 1-day view reports fewer conversions, and its delivery concentrates on people likely to convert within a day of clicking. We therefore recommend treating a window change as a campaign change, to be tested on a duplicated ad set rather than across the whole account at once.

The bias no model corrects

All these settings share credit among the ads that preceded a sale. None of them says whether the sale would have happened without those ads. The bias is strongest on brand search, remarketing and view-through conversions: someone who saw a Reels video yesterday and buys today may already have decided. Each platform also sees only its own touchpoints, which leads Google and Meta to count the same order, as we explain in our article on the limits of ROAS.

Meta implicitly acknowledges the problem. Since June 2025, an incremental attribution setting optimises delivery and counts conversions using models trained on Conversion Lift studies (PPC Land). It remains an estimate produced by Meta, and it does not replace a test with a control group, such as a Conversion Lift study or a geo test. Failing that, at the very least track total back-office revenue divided by total ad spend: no window can inflate it.

Before you change a setting

Before discussing any ROAS target, we would settle these four points:

• Measure the actual delay between first click and purchase, using your order data or the Google Ads days-to-conversion report, and choose a conversion window that covers the vast majority of your purchases, rather than the default.
• Keep what drives your revenue as the primary action. Micro-conversions (add to cart, sign-up) stay secondary, unless purchase volume is too low for bidding, in which case write that choice down.
• Log the date of every change to a window, model or attribution setting, including those imposed by the platforms. On Google, a new conversion window only applies to future conversions, so a comparison spanning that date mixes two rules.
• Only compare Google and Meta on equivalent windows. Thirty days after a click on one side and seven on the other do not measure the same thing.

If conversions are missing before they are even attributed, the window question comes second: start with data collection, which we cover in our guide to server-side tracking.

FAQ

Should you keep the Last click model in Google Ads?

Rarely. It gives all the credit to the last ad clicked, which usually favours brand queries typed just before purchase. We only recommend it to keep, during a transition, a series comparable with your history.

Why might your Meta click-through conversions have dropped in spring 2026?

One possible cause is the new click definition, rolled out gradually from March for campaigns optimised for website or in-store conversions: conversions preceded by a share, a save or a like no longer count as click-through conversions. Meta states that billing does not change. Look at the conversions classified as "engage-through" before concluding that performance has dropped.

Settings checked on 13 October 2026.