Data-Driven Attribution: How It Works, With a Worked Example

Data-driven attribution is a model that splits the credit for each conversion across the touchpoints that led to it, based on how much each touchpoint actually raised the chance of converting in your own data. It compares the paths of people who converted with the paths of people who did not, instead of applying a fixed rule like “the last click gets everything.” It is the default model in Google Ads and in GA4.

This guide explains how the model works, how it differs from rule-based models, how much data it really needs (the top guides disagree, and Google’s docs settle it), and how to turn its fractional credit into revenue by channel and landing page, with the arithmetic shown.

What Is Data-Driven Attribution?

Attribution is the job of deciding which marketing touchpoints get credit for a conversion: a purchase, a lead form, a sign-up. Every model answers it with some logic. Rule-based models use a rule someone chose in advance. Data-driven attribution (often shortened to DDA, and also called algorithmic attribution) learns the split from the data.

The practical result is fractional credit. A single sale might be credited 0.6 to paid search, 0.3 to organic search and 0.1 to email, instead of 1.0 to whichever came last. Each model is built per account and per conversion action, so two businesses with the same channels can get very different splits.

How Data-Driven Attribution Works

Under the hood, every data-driven model does three things:

1. Collects paths. It records the sequence of touchpoints (ad clicks, video engagements, visits from each channel) for users who converted and for users who did not.
2. Estimates conversion probability. It models how likely a conversion is given which touchpoints were present, plus factors such as order and timing.
3. Assigns credit by lift. A touchpoint gets more credit when paths that include it convert noticeably more often than similar paths without it.

Google describes its version as counterfactual: the model compares the key event probability of users exposed to an ad with that of similar users in a holdback group. Its own illustration uses a path with four ad exposures that converts with a 3% probability; without the fourth exposure the probability drops to 2%, so the fourth exposure adds 50% to the probability of converting. GA4 also weighs time from the key event, device type, number of ad interactions, order of exposure and creative type, according to Google’s attribution documentation.

Is it the Shapley value?

Some guides say GA4 uses the Shapley value, a method from cooperative game theory that averages each player’s marginal contribution across every possible order of arrival. That is a fair way to understand the idea, and it is how many custom models are built. Google’s current help pages describe a counterfactual, holdback-based approach and do not name Shapley, so treat “GA4 is Shapley” as an approximation rather than a documented fact. The worked example below uses Shapley because you can do it by hand.

Data-Driven Attribution vs Rule-Based Models

Single-touch models give all the credit to one touchpoint. Multi-touch models split it. Data-driven attribution is a multi-touch model whose split is learned rather than fixed.

Several ranking guides still list six Google Ads models. Google has since removed first click, linear, time decay and position based; GA4 states they are no longer available as of November 2023. Today the choice in GA4 is data-driven, paid and organic last click, or Google paid channels last click. For a deeper look at the main alternative, see last click attribution.

Where You Find It: Google Ads and GA4

The two Google tools use the same idea but see different data.

Google Ads lists the supported networks on its data-driven attribution help page. Because Google Ads only sees its own ads, its conversion counts per campaign will not match GA4’s, which also credits organic search, email and other channels. Both are right about their own scope.

How to Turn On Data-Driven Attribution

In Google Ads

New conversion actions use data-driven attribution already. To switch an existing one:

1. Open Goals, then Conversions, then Summary (older interface: Tools and Settings, Measurement, Conversions).
2. Click the conversion action you want to change.
3. Click Edit settings.
4. Under Attribution model, choose Data-driven.
5. Save. Expect conversion columns to shift as credit moves between campaigns, and give automated bidding a couple of weeks to adjust.

Before switching, open the model comparison view in the Attribution section and compare last click with data-driven for the same conversion action. It shows how conversions and cost per conversion would move, without changing anything.

In GA4

1. Go to Admin, then Attribution settings (under Data display).
2. Set Reporting attribution model to Data-driven.
3. Set Channels that can receive credit to “Paid and organic channels” if you want organic search, email and referrals in the model. “Google paid channels” restricts credit to Google Ads.
4. Check the key event lookback windows.

Then use Advertising, then Attribution: the Model comparison report shows how credit shifts between models by channel, and Conversion paths shows which channels appear early, mid and late in journeys.

Why Use It: What Data-Driven Attribution Does Well

• It credits assists. Channels that start or advance journeys, such as content, generic search terms and video, get some credit instead of zero.
• It removes guesswork about weights. You do not pick 40/20/40 or a decay rate. The split comes from your paths.
• It improves the bidding signal. In Google Ads, the model feeds Smart Bidding (Target CPA, Target ROAS), so keywords and campaigns that assist conversions can win bids they would lose under last click.
• It adapts. As channels, offers and buyers change, the model is retrained on fresh data rather than keeping a stale rule.
• It makes partial credit visible. Seeing 0.4 of a conversion on an upper-funnel campaign is a better input for budget talks than a flat zero.

Limitations of Data-Driven Attribution

• It only sees what is tracked. Google Ads sees Google ads. GA4 sees visits that reach your site with a source. Ad impressions on other platforms, podcasts, events, word of mouth and offline sales calls are invisible, so channels that work there get undercredited.
• It is a black box. You see the credit split, not the weights behind it. That is hard to defend in a budget meeting.
• It measures association, not cause. Branded search shows up in many converting paths because people who already decided to buy search your name. Only holdout or incrementality tests prove what a channel causes.
• Privacy limits the paths. Consent choices, browser tracking protection and cross-device journeys break paths into pieces, and the model can only learn from the pieces it has.
• Lookback windows cap the journey. GA4 looks back 90 days at most. B2B deals that take six months or more lose their early touchpoints, and so do buying committees spread across several people.
• Low volume means noise. With few conversions, credit shares swing from month to month for no real reason.

How Much Data Does Data-Driven Attribution Need?

The ranking guides give four different answers, all stated as requirements:

Google Ads now says all conversion actions are eligible, recommends at least 200 conversions and 2,000 ad interactions in supported networks within 30 days, and states that data-driven attribution will still function with less data, while getting more precise as volume grows, according to Google Ads Help. The higher numbers in older guides appear to reflect earlier eligibility rules. We could not confirm the claim that GA4 silently reverts to last click below a threshold in any current Google documentation.

Worked Example: From Path Data to Credit to Revenue

No ranking guide shows the arithmetic, so here it is with two channels. All numbers are illustrative, not from a real account, and this is a simplified Shapley calculation, not Google’s exact algorithm.

Step 1: Conversion rate by set of channels

An online furniture store groups its users by which channels they touched in the lookback window:

Step 2: Marginal contribution in each order

For a path with both channels, ask what each channel adds when it arrives first and when it arrives second, then average.

The two averages add up to the full 6 points, which is the point of the method: nothing is double counted. Organic search gets 2.5 / 6 = 41.7% of the credit and paid search gets 3.5 / 6 = 58.3%.

Step 3: Turn credit into revenue

Say 300 orders came from paths where organic search came first and paid search came last, with an average order value of $120. That is $36,000.

Step 4: Check the decisions it changes

• Paid search ROAS. With $9,000 of spend, ROAS is $36,000 / $9,000 = 4.0 under last click and $21,000 / $9,000 = 2.33 under data-driven. A bid strategy targeting a ROAS of 3.0 behaves very differently depending on which number it sees.
• Organic landing pages. If 200 of those 300 journeys started on a desk-chair buying guide and 100 on a standing-desk guide, the guides earn $10,000 and $5,000 of credit. Under last click, both look like they sell nothing.

The total never changes between models. Data-driven attribution only moves credit. If your totals change when you switch, you are comparing different data sources or scopes, not different models.

What GA4 Data-Driven Attribution Does to Organic Search, Direct and AI Traffic

Three GA4 rules decide how much SEO and AI assistant traffic shows up in data-driven reports.

• The channel setting. Organic search only receives credit when Channels that can receive credit is set to “Paid and organic channels.” With “Google paid channels,” every organic touchpoint is ignored. App conversions always use Google paid channels.
• The Direct rule. GA4 gives Direct visits credit only when the whole path is Direct. A buyer who found you through a blog post and later typed your URL gives all the credit to the blog visit’s channel, which is usually what you want.
• AI assistants with no referrer. When ChatGPT, Perplexity or another assistant passes a referrer, the visit is a referral and can earn credit. When it passes none, the visit lands as Direct, and the Direct rule then drops it from any mixed path. No model can credit a source it never saw. See AI conversion tracking for what can be identified.

Session-scoped reports, such as Traffic acquisition, ignore the reporting model and keep using last click, so organic search will look smaller there than in key event reports set to data-driven. For landing-page level reporting on organic traffic, see organic conversion attribution.

Failure Modes: Why Your DDA Numbers Look Wrong

Reattribution and history rules come from Google’s attribution settings page and its attribution overview.

Should You Trust Data-Driven Attribution? A Decision Rule

Using DDA for Budget Decisions: A Five-Step Routine

Data-driven marketing means letting measured results, not habit, set where the next dollar goes. Attribution is one input. A simple routine:

1. Ask one question. For example: “Should we move $5,000 a month from brand search to content?”
2. Pull the data. Data-driven and last click revenue for the channels involved, over a period that ended more than seven days ago.
3. Analyze the gap. Channels that gain a lot under data-driven are assisting. Channels that lose a lot were being overcredited by last click.
4. Decide with a test. Move part of the budget, or pause a campaign in some regions, rather than everything at once.
5. Review. After a full sales cycle, check whether total conversions and revenue moved, not just the credit split.

Example, continuing the furniture store (illustrative): data-driven credit puts paid search ROAS at 2.33 instead of 4.0, and the buying guides at $15,000. The team trims paid search spend by 10% in half of its regions for six weeks and keeps publishing guides. If total orders in the test regions hold steady, the cut was safe, and the data-driven view was closer to reality.

How to Build Your Own Data-Driven Attribution Model

If you need a model outside Google, for example to include email and CRM data, the Shapley approach above scales with tooling:

1. Export user or session paths, including non-converting ones (GA4’s BigQuery export works).
2. Group paths by the set of channels they contain and compute the conversion rate of each set.
3. For each channel, average its marginal lift over all orderings of the other channels.
4. Normalize so each path’s credit sums to one conversion, then multiply by conversion value.
5. Validate: rerun on a different month and check that the shares are stable before anyone uses them.

With more than a handful of channels the number of combinations grows fast, so teams usually move to Markov chain removal effects or a statistics library rather than a spreadsheet. For an overview of tools that do this for you, see marketing attribution tools.

Put a Value on Each Conversion First

Fractional credit is only useful when the conversions carry value. A newsletter sign-up and a demo request should not count the same, or the model will optimize toward cheap actions. Set a value for each conversion type (see how to calculate conversion value), then read every attribution report in revenue, not counts. For organic search and AI assistants, SEOConversion reports conversions and their value by landing page, and keeps visits with no referrer as Direct rather than guessing. The setup is covered in SEO conversion tracking.

FAQ