Every attribution model shares out credit by its own rules. Here is where each one misleads you, and how to cross-check them before you move budget.
Picture a month in which Google Ads reports 80 sales, Meta reports 60 and your shop recorded 100. The figures cannot all be right, yet the channel that tells the best story tends to win the next budget. Attribution, the rules that decide which ad or channel gets credit for a sale, sits behind each of them.
Your attribution model is lying because every model is a set of assumptions about how to share out credit, and it can only share it among touchpoints it can see. Last click hands everything to the final step, ad platforms credit their own ads, and data-driven models miss offline conversations and visitors who declined cookies. You get closer to the truth by cross-checking three sources, not by hunting for a perfect model: platform reports, your own records and customers' answers, and tests that switch spend off to see what it really adds.
This article shows how each common model bends the picture, where Google's data-driven model helps and where it cannot, and how to run a simple test that measures what your ads actually add. It ends with a short routine you can run without a data team.
Every sale has a path behind it: an ad seen on a phone, a colleague's recommendation, a comparison article, a search for your name a week later. An attribution model looks at the parts of that path it can observe and divides the credit by a fixed rule or a statistical estimate. The parts it cannot observe get nothing, however much they mattered.
Much of what moves people leaves no trail. Word of mouth, podcasts, trade fairs, a link shared in a messaging app, a video watched without a click, a visitor who declined cookies or switched from phone to laptop: none of these show up cleanly in the data.
The bigger blind spot is that attribution describes rather than tests. It can tell you which ads came before a sale, not whether the sale would have happened without them. Sales an ad actually caused, the ones that would not have happened otherwise, are called incremental sales, and no attribution model measures them directly.
Last-click attribution gives all the credit to the final touchpoint before the sale, which is often a click on a brand search ad or a direct visit. Take a customer who sees three of your Meta ads, reads a review, and a week later types your company name into Google and clicks your ad: last click gives the entire sale to that brand search ad.
The result is predictable. Brand search campaigns look extremely profitable because they collect people who had already decided to look for you, while the video and social ads that created the interest look worthless. Some of those brand clicks would have reached you through your free listing anyway, especially when no competitor bids on your name.
Last-click logic is still how many businesses read their numbers, because it is simple and matches the everyday question of where a sale came from. In business-to-business (B2B) sales, where the first contact may come months before the final click, it is especially misleading. First click, its mirror image, has the opposite fault: it credits whatever started the journey and ignores everything that closed it. Rule-based alternatives such as linear, which splits credit equally across every touch, or time decay, which gives more to recent ones, only swap one fixed assumption for another.
Two practical rules follow. Report brand search separately from non-brand search, and never use brand search returns to justify spending anywhere else. If competitors do not bid on your name, pausing brand ads for a few weeks in part of your market will show what they really add.
Each ad platform counts the conversions it can connect to its own ads, within its own attribution window. That window is the number of days after a click, or after someone merely saw an ad, during which the platform will claim a sale. A customer who clicked a Google ad and also saw a Meta ad can be counted in full by both. Add the platform reports together and the total can easily exceed your actual sales, as in the month described above.
Since Apple's App Tracking Transparency arrived with iOS 14.5 in 2021, Meta has seen less of what happens after an ad and fills part of the gap with modelled estimates. Its Conversions API, which sends purchase and lead events from your server rather than from the visitor's browser, recovers some of that signal and is worth setting up. But a better-informed platform is still a platform reporting on its own ads.
Use platform figures for decisions inside the platform: which ad, audience or creative works best compared with the others. For decisions between platforms, compare the sum of what they claim with the orders or enquiries you actually recorded, and treat the gap as a measure of double counting.
Data-driven attribution, the standard model in Google Analytics 4 (GA4) and Google Ads, compares the paths of people who converted with those of people who did not, and estimates how much each touchpoint added. It is a genuine improvement on fixed rules, because it learns from your own data instead of assuming that the first or last click matters most.
It has three limits. It needs volume: with only a few dozen conversions a month, as in many B2B accounts, its estimates rest on thin evidence, so treat its splits as rough. It sees mainly what Google can see, so a Meta video someone watched without clicking does not exist in its paths. And it still observes rather than experiments, so a touchpoint that often appears before a sale gets credit whether or not it caused anything.
Use data-driven attribution for what it does well: comparing campaigns, keywords and ads within Google's own channels. Do not let it settle how much of your budget goes to Google rather than everything else.
The most honest approach is not to choose the right model but to triangulate: read three imperfect sources side by side and act where they agree. Each one covers a blind spot of the others:
Customers' own answers are messy, and some people simply pick the first option. They are still the only source that captures a podcast mention or a recommendation from a friend, so read them for patterns over a quarter rather than record by record.
The simplest incrementality test is a regional holdout. Choose regions that usually deliver a stable share of your sales and switch one channel off there for four to six weeks while it keeps running elsewhere. Then compare what the paused regions sold with what their usual share predicts. In a business selling across several countries the regions can be countries; within one country, groups of counties or postcode areas work if the platform lets you target them.
Here is an illustrative result. Before the test, Meta's reports credited the paused regions with about €8,000 of sales every four weeks; with the ads off, sales there fall by only €4,000 compared with what the control regions predict. Meta's real contribution is then about half of what it reports: possibly still profitable, depending on your margins, but worth half the credit in your planning.
Tests have limits too. Weekly sales in smaller businesses are noisy, so run the test long enough, avoid holidays and big promotions, and repeat it before making a large, permanent change. The large ad platforms also offer their own lift studies for bigger budgets, and the principle is the same.
When the three sources disagree, let the decision choose the source. Trust tests for moving large sums between channels, your own records for B2B deal sources, and platform data for creative and targeting choices.
Attribution does not tell you the truth; it tells you which version of the truth is most useful for the decision in front of you.