Data-driven attribution: how Google's model works in GA4, Google Ads and Search Ads 360

What data-driven attribution is

Data-driven attribution (DDA) is an algorithmic attribution model: instead of applying a fixed rule to every conversion path, it estimates from your own data how much each touchpoint contributed to the conversion, and assigns credit in proportion.

The idea underneath is counterfactual. For each touchpoint the model asks: how likely was this person to convert given that they saw this ad, compared with a similar person who did not? The difference is the touchpoint’s contribution. Add the contributions up along a path and you have the split. Game theory gives the formal version, the Shapley value, and the Markov chain’s removal effect asks the same question from the other side; both are worked through in the attribution models guide.

Two things follow that matter in practice. The model is specific to you: it is trained on your paths, so it cannot be copied from another account. And it needs non-converting paths as well as converting ones, because the whole method is a comparison between the two. A tool that only logs successes cannot run it.

How Google’s model works

Google publishes more about its method than most vendors, and the description on its attribution help page is worth reading in full. The short version, in Google’s own terms:

• The model “uses machine learning algorithms to evaluate both converting and non-converting paths”.
• It “incorporates factors such as time from key event, device type, number of ad interactions, the order of ad exposure, and the type of creative assets”.
• It uses “a counterfactual approach”, contrasting “what happened with what could have occurred to determine which touchpoints are most likely to drive key events”, and assigns credit “based on this likelihood”.
• “Each data-driven model is specific to each advertiser and each key event.”
• “Conversions can be reattributed for up to 7 days after the conversion”, so a path that is still open can change its credit for a week.

The method is sound, and it is the same one you would use yourself with enough data. The questions to keep in mind as you read the product sections are not about the method but about what the model can see (only Google-measured touches) and what you can see of the model (the output, not the weights).

DDA in Google Analytics 4

Google Analytics 4 made data-driven attribution free to every property, which in Universal Analytics was a 360-only feature with a volume gate. Today it is the recommended reporting model, and the one most teams should compare their platform reports against.

The three models GA4 offers

The Attribution reports offer exactly three models: data-driven, paid and organic last click, and Google paid channels last click. The rule-based models that Universal Analytics users remember, first click, linear, time decay and position-based, were removed in November 2023. If you want them, you build them yourself, which the last section covers.

Where the settings live

The reporting attribution model is set in Admin → Data display → Events → Attribution settings, and you need the Marketer role or above on the property. Changing it applies to historical and future data in every report that uses event-scoped traffic dimensions, so conversion and revenue figures move when you switch. Any viewer can still pick a model inside the Model comparison and Conversion paths reports without changing the property default.

Two more settings sit on the same page:

• Key event lookback window. For acquisition key events (first_open, first_visit) the default is 30 days, switchable to 7. For every other key event the default is 90 days, with 30 and 60 as options. The window applies to all models and all key event types, and changes apply going forward.
• Channels that can receive credit. Google paid channels only, or paid and organic channels. The second gives a cross-channel view in Analytics and affects the conversions shared with linked Google Ads accounts; it applies to web conversions only, since app conversions always use Google paid channels.

GA4 renamed conversions to “key events” in 2024, which is why the settings and reports use that word.

What it does well

• Free, and already attached to the data you collect.
• The model is trained on your property, not on a generic benchmark, and the comparison reports make it easy to see how credit shifts against last click.
• The lookback and channel settings are explicit, which is more control than most ad platforms give.

Where it falls short

• It is a black box. Google states no minimum data for DDA in GA4, but it also publishes no weights; you see the output and must take the method on trust.
• It sees what Google sees. Touches on Meta, LinkedIn, TikTok or email are only in the model if they reach GA4 as tagged sessions, and anything after the website, such as a CRM-qualified lead or a refund, is invisible.
• Only three models remain. A team that needs a position-based or time-decay baseline has to build it elsewhere.

DDA in Google Ads

In Google Ads, data-driven attribution is “the default attribution model for most conversion actions”. The only other model still supported is last click; conversion actions that used the deprecated rule-based models were upgraded to data-driven.

The model “looks at website, store visit, and Google Analytics conversions from Search (including Shopping), YouTube, Display, and Demand Gen ads”, compares the paths of customers who convert with those who do not, and credits the ad interactions that raise the probability of conversion. Because the same model feeds Smart Bidding, its practical job is less reporting than optimisation: it tells automated bidding which keywords, ads and campaigns to lean on.

Data requirements

The old gate is gone. Google’s page now says that “all conversion actions are eligible for data-driven attribution, regardless of conversion or interaction volume”, with a recommendation of “at least 200 conversions and 2,000 ad interactions in supported networks within a 30-day period” for the model to perform well. Below that it still runs; it just has less to learn from.

What it does well

• Default, free and tied directly to bidding, so the attribution output changes spend without anyone building a report.
• Covers YouTube and Display engagements as well as clicks, which last click ignores.

Where it falls short

• It is a Google Ads model. Every other channel, and every conversion Google Ads cannot see, is outside it.
• The credit split optimises Google’s spend; it is not a view of your marketing mix.

DDA in Search Ads 360

Search Ads 360 is the Google Marketing Platform’s search management tool, built for advertisers running paid search across several engines. Its data-driven attribution works on Floodlight conversions, the platform’s own conversion tracking, and feeds its bid strategies.

The requirements are the strictest of the three products: “your sample must include at least 15,000 clicks and 600 Floodlight conversions during the last 30 days” to create a model. In return you get up to five DDA models, each covering as many as 50 Floodlight conversion actions, a model that “learns and updates each week”, and reporting applied “multiple times throughout the day” so the numbers are near real time. A separate option lets a Google Analytics 4 key event use data-driven attribution inside Search Ads 360, under the same thresholds.

What it does well

• Several models side by side, so channel groupings can be compared rather than argued about.
• Near-real-time reporting and direct connection to bid strategies.
• Offline conversions can be uploaded into Floodlight, so the model can see more than the website.

Where it falls short

• The volume gate rules out most mid-sized advertisers.
• It is paid-search attribution. Organic, email, social and impressions outside the Floodlight scope are not in the path.
• It is an enterprise product with its own setup: Campaign Manager 360, Floodlight configuration, channel-grouping labels.

What Google’s DDA cannot see

All three products run the same honest method on a partial view. Three gaps recur, and they are the reason a data-driven number from Google and a data-driven number from your warehouse can disagree by a wide margin.

• Other platforms’ touches. Each product models the interactions it measured. A journey that ran Meta → email → Google search → direct looks, to Google Ads, like a one-touch search conversion.
• Everything after the click. Qualified leads, closed revenue, refunds and returns live in your CRM and order system. A model trained on “key events” is trained on form fills and checkouts, not on money.
• The weights. You can see that search got 38% of a conversion; you cannot see why, reproduce it, or run it on last year’s data under a changed assumption. For a dashboard that is acceptable. For a budget decision someone has to defend, it is the weak link.

Build a model you can inspect

The method behind data-driven attribution is not Google’s. Shapley values and Markov removal effects are published mathematics, and the inputs they need, every session with its channel and every conversion with its value, are in the GA4 BigQuery Export and your ad and CRM data. Put those in one warehouse table and the model is yours: you choose the lookback window, you include the CRM revenue and the refunds, and you can print the weights.

The practical path has three steps. Marketing attribution models walks through the first, with a runnable BigQuery query that computes four rule-based models from the GA4 export in one pass, and explains where the Markov and Shapley computations live. Build a marketing attribution model in BigQuery covers the step before that: sessionising GA4 events and attaching ad cost to sessions, so the output is return on spend rather than a share of conversions.

The third step is the one that makes the model useful to anyone but its author. Published as a governed Data Mart in OWOX Data Marts, the attribution query has a name, an owner and a definition of what counts as a conversion; reports in Google Sheets and Looker Studio read from it, and a business user can ask for the channel breakdown in plain language and get the number with the query behind it. The analyst still owns the SQL; the product publishes and schedules it. That is the difference between a data-driven number you can show the board and one you have to take on faith.

Free guide: The Ultimate Guide to Marketing Attribution Modeling: Everything You Need to Know →

Which to use

• GA4’s data-driven model for the cross-channel view of what Google can measure, and as the baseline every platform report is compared against. Turn on paid and organic channels unless you have a reason not to.
• Google Ads’ model for what it was built for: feeding Smart Bidding. Read its credit split as a bidding signal, not as a marketing-mix report.
• Search Ads 360 only if you clear the volume gate and run paid search across engines at a scale that justifies the platform.
• Your own model in the warehouse when the number has to include non-Google channels, CRM revenue or refunds, or has to be explained to someone who will ask how it was calculated.

Key takeaways

• Data-driven attribution learns the credit split from converting and non-converting paths instead of applying a rule. The method is sound and the mathematics is public.
• Google runs it in three products with three scopes: GA4 sees tagged sessions, Google Ads sees its own ad interactions, Search Ads 360 sees Floodlight conversions at enterprise volume.
• GA4 offers three models, data-driven and two last-click variants, with settings under Admin → Data display → Events → Attribution settings, a 30- or 90-day default lookback, and a choice of which channels may receive credit.
• Google Ads has no minimum volume for DDA any more and recommends 200 conversions and 2,000 interactions a month; Search Ads 360 still requires 15,000 clicks and 600 Floodlight conversions.
• None of them can see other platforms, CRM revenue or refunds, and none shows its weights. When the number has to be defended, build the model in your warehouse and publish it as a governed Data Mart.

Marketing AttributionMarketing Attribution Models: How Each One Works and How to Build Your OwnOctober 1, 2026→AI in AnalyticsMachine Learning Use Cases in Digital Marketing in 2025January 18, 2025→Marketing AttributionHow to Build a Marketing Attribution Model in BigQueryOctober 14, 2026→

See all articles →