Every marketing report that splits a sale across channels runs on an attribution model, whether anyone chose it or not. The model decides which team looks like it drove the sale. For a loyalty program, it also decides whether the points email, the app push or the paid ad gets the credit for a member’s next purchase.
What is an attribution model?
An attribution model is the rule that gives credit for a sale to the touches that came before it. Google calls it “a rule, a set of rules, or a data-driven algorithm” for handing out that credit (Google Analytics Help). The goal being credited can be a purchase, a sign-up or a new member joining.
The model works inside a lookback window, the stretch of time in which a touch can earn credit. In Google Analytics 4, the default window is 30 days for first visits and app opens, and 90 days for all other key events (Google Analytics Help). A touch before the window gets nothing.
What are the main types of attribution models?
Single-touch models give all the credit to one touch. Multi-touch models split it. Rule-based models split it by a fixed formula, and data-driven models split it by what the data shows. Google described the rule-based options in its Ads help pages before retiring them (Google Ads Help, 2022).
| Model | How it splits credit | Search ad, then email 7 days later, then push 7 days later, then a $120 sale |
|---|---|---|
| Last click | All credit to the last click | Push $120 |
| First click | All credit to the first click | Search $120 |
| Linear | Equal credit to every touch | $40 each |
| Time decay | More credit to recent touches, on a 7-day half-life | Search $17.14, email $34.29, push $68.57 |
| Position-based | 40% to the first and last touch, 20% shared by the rest | Search $48, email $24, push $48 |
| Data-driven | Credit by how much each touch raised the odds of converting | Depends on the account’s own path data |
The example path is an illustration. Each model reads the same three touches and tells a different story. Under last click, the push team owns the sale. Under first click, the search team does.
Why did Google retire most rule-based attribution models?
Google took first click, linear, time decay and position-based models out of Google Analytics in November 2023 (Google Analytics Help). In Google Ads, goals set to those models were moved to data-driven. Last click is still on offer. Data-driven is now the default for most goals (Google Ads Help).
Data-driven attribution learns from paths that converted and paths that did not. Google says its models compare users who saw an ad with similar users in a holdback group. They are trained on data from randomized controlled trials. Credit then follows how much each touch changed the chance of a key event (Google Analytics Help). By default, direct visits get no credit unless the whole path was direct.
Google gives a simple case. Four ads in a row lead to a 3% chance of a sale. Without the fourth ad, the chance drops to 2%, so the fourth ad adds 50% to the odds. The model repeats this for each ad and uses the results as credit weights (Google Analytics Help). That is why data-driven reports show fractions of a sale.
Does an attribution model measure incremental sales?
No. An attribution model splits credit for sales that happened. It does not ask which of those sales would have happened anyway. A model can give a channel full credit for sales it did not cause.
eBay showed how big the gap can be. In a series of large field tests, its researchers found that brand keyword ads had no measurable short-term benefit. Because search clicks and buying go together, paid search returns were a fraction of standard estimates (Blake, Nosko and Tadelis). A last-click report would have credited those brand ads with every sale that followed a click. A study of 15 US ad tests at Facebook found the same pattern. Methods that read effects from tracking data often failed to match the results of the randomized tests (Gordon et al.). Only a test with a holdout shows the true lift. Our incrementality entry covers that measure.
How do you choose an attribution model?
Start with what the model will decide. If it sets bids inside one ad platform, use the data-driven model that platform offers. If it splits a budget across channels, a single platform’s model sees only its own touches.
Then match the lookback window to the buying cycle. A grocery shopper buys within days. A hotel guest books weeks ahead. Check which touches the model can see, since email, app and store sales often sit outside the ad platform. Last, test the model’s claims. When a model says a channel drives sales, pause that channel for a random group and see if sales fall.
Data volume matters as well. A data-driven model needs enough sales to learn from, and Google’s model comparison report drops any campaign without enough data (Google Ads Help). Once you pick a model, keep it. In Google Analytics, a change of model is applied across all key event reports that use event-level traffic data (Google Analytics Help). Trends read across a model change can mislead.
How do loyalty programs use attribution models?
A loyalty ID ties touches to sales that ad platforms never see. Most store purchases leave no click behind. When the member scans a card or app at the till, the sale joins the member’s record. Ulta Beauty reported more than 46 million members at the end of fiscal 2025, with about 95% of total sales coming from members (Ulta Beauty 10-K). At that coverage, nearly every sale can be traced to a person.
Retail media runs on the same link. Kroger’s ad arm builds audiences from the purchase data of Kroger’s loyalty ecosystem, which covers 62+ million households. It matches each ad seen to real household purchases in stores and online. To find the true effect, it compares groups that saw the ad with groups that did not (Kroger Precision Marketing). Our entry on in-store attribution covers the matching step.
The program’s own messages need the same care. A member who opens the weekly points email and then shops gives the email a last click. That holds whether or not the email changed anything. Routine shoppers make program email look stronger than it is.
For illustration, take the path in the table and add a holdout. Members who got the push bought at a 22% rate within 14 days. A random group with no push bought at 20%. Only 2 of every 22 push-group sales are new, about 9%. Last click gave the push all $120 of each sale. On a fair count, it earned about $10.91 per sale, before the cost of any points.
How GRAVTY handles attribution
Attribution for a loyalty program starts with a complete member record. GRAVTY ties every earn, burn, tier change and offer response to one member through Member 360. On lji.io, AgenticTest attributes incremental lift at the segment level while a campaign is still running. Agentic AI Compass lets a team ask why a metric changed and compare any segment in plain language.
Frequently asked questions
What is the default attribution model in Google Ads?
In Google Ads, the default for most conversion actions is data-driven attribution. Last click is the only other model still on offer. Actions that used the four retired models were moved to data-driven (Google Ads Help).
What is a lookback window in attribution?
A lookback window is how far back a touch can be and still earn credit. In Google Analytics 4, the default is 30 days for first visits and first app opens. For all other key events, it is 90 days (Google Analytics Help). A 30-day window credits a January 30 purchase only to touches from January 1 to 30.
What is the difference between an attribution model and marketing mix modeling?
An attribution model splits credit for each conversion across the touches a user had. Marketing mix modeling works from totals instead. It relates weekly sales to spend by channel, price and season, with no user-level paths. Attribution is fast and detailed. Mix modeling covers channels that leave no click, such as TV and stores.
Why do attribution reports from different tools disagree?
Each tool sees different touches and counts them by different rules. Google Analytics, for example, gives no credit to direct visits unless the whole path was direct (Google Analytics Help). An ad platform sees only its own ads. Add different lookback windows, and the same sale can be claimed by two tools at once.
Which attribution model is best for a loyalty program?
The best model is the one checked against a holdout. Data-driven models suit bidding inside one platform. For program decisions, such as which offers to keep, attribution should rest on the member record and a random control group. That shows how much of the credited spend the touch really caused.
