Every Attribution Model, Compared — and Where They Break in the AI Era

Strip away the vendor language and an attribution model is one simple thing: a rule for dividing conversion credit among the touchpoints you managed to observe. A customer clicked a Meta ad on Monday, an email on Thursday, and a branded search ad on Saturday before buying. Which of those “gets” the sale? The model is the answer to that question — nothing more.

Two consequences follow, and they matter more than any individual model choice. First, the model can only divide credit among observed touches: anything invisible — a podcast mention, a friend's recommendation, an answer from an AI assistant — gets exactly zero, always, under every model. Second, no model measures causation. Reassigning credit between the Meta ad and the email tells you nothing about whether either one actually changed the buyer's behavior. (For methods that do measure causation, see our comparison of MMM, MTA, and incrementality testing.)

With that framing, here is every model you'll encounter, what it rewards, and what it hides.

Last-click attribution

100% of credit to the final touchpoint before conversion. Still the default lens in most ad platforms and reports, because it's simple and it flatters the platform showing you the number. Systematically rewards bottom-funnel “harvesting” channels — branded search, retargeting — that show up right before conversions they often didn't cause. Under last-click, demand capture always beats demand creation, which is precisely backwards for judging where growth comes from.

First-click attribution

100% of credit to the first observed touchpoint. The mirror image: useful as a lens on which channels introduce customers, and essentially never used as a primary model. Its deeper flaw is quiet but fatal in 2026: the “first click” is only the first tracked click. When the true first touch was an AI answer, a podcast, or a friend, first-click credits whichever tracked channel happened to come next — usually branded search — and calls it discovery.

Linear attribution

Equal credit to every observed touchpoint. Feels diplomatic, informs nothing: if every touch is equally valuable, no touch is decision-relevant. In practice linear mostly functioned as a political compromise between teams arguing over budget.

Time-decay attribution

Credit increases the closer a touch is to conversion (typically a 7-day half-life). A refined version of last-click's worldview — reasonable for short sales cycles, but it structurally assumes that later touches matter more, which is exactly wrong for consideration-heavy purchases where the early research shaped the decision.

Position-based (U-shaped) attribution

Typically 40% to the first touch, 40% to the last, 20% spread across the middle. An honest attempt to value both discovery and closing. Variants exist for lead-gen funnels (W-shaped adds a 30% block for the lead-conversion touch; Z-shaped/full-path adds opportunity stages). All inherit the same weakness: the weights are arbitrary — why 40/20/40 and not 30/40/30? — and the “first touch” is still only the first one tracked.

Data-driven attribution (DDA)

Instead of fixed rules, a machine-learning model estimates each touchpoint's contribution by comparing conversion rates across observed paths that do and don't include it (conceptually related to Shapley values from game theory). This is the default in GA4 and Google Ads, and the marquee feature of most paid attribution tools. Genuinely better than fixed rules — and deserving of its own section below, because “the algorithm figured it out” hides important caveats.

If several of these models sound like history: they are. In 2023 Google removed first-click, linear, time-decay, and position-based attribution from GA4 and Google Ads, leaving only last-click (in paid-and-organic and ads-preferred flavors) and data-driven attribution, with DDA as the default. Google's stated reasoning was blunt: the rule-based models were barely used and, in its words, didn't reflect how people actually convert.

The subtext matters for how you read your own reports. Rule-based multi-touch models were built for a world of durable cookies and long observable paths. As privacy changes shortened observable journeys — often to a single tracked touch — elaborately weighting a path of one stopped making sense. Google consolidating on DDA was as much an admission about data quality as a statement about model quality. Note also that GA4's DDA quietly leans on Google's modeled and consent-gap-filled data; the model and the data it runs on are a package deal.

DDA earns its status as the best default. Comparing paths with and without a given touchpoint is a real methodological improvement over arbitrary weights, and at high conversion volumes it produces credit assignments that at least reflect observed correlations rather than a committee's opinion.

Its honest limitations:

• It's a black box you cannot audit. When DDA shifts 15% of credit from search to display this month, no report will tell you why. You are trusting the vendor's methodology wholesale.
• Platform DDA sees only the platform's touches. Google's DDA weighs Google touchpoints; it cannot devalue a Google click in favor of a Meta view it never saw. Every walled garden's “data-driven” number quietly overstates that garden.
• It needs volume. Below roughly 300–600 conversions per month (thresholds vary by tool), estimates get noisy and models silently fall back toward last-click-like behavior.
• It is still correlational. DDA answers “which observed touches predict conversion” — not “which touches caused it.” Retargeting predicts conversion beautifully; whether it causes any is a question only an incrementality test can answer.
• And, like every model, it divides only what it sees. An AI assistant's recommendation that never produced a tracked click receives zero credit — with full algorithmic confidence.

Walk through a journey that is rapidly becoming typical:

1. 1The owner of a marketing agency asks ChatGPT: “best project management tools for a 15-person agency — compare pricing.” ChatGPT names four products with a summary of each. No website visit occurs.
2. 2Over the next week she asks two follow-ups, reads a comparison the assistant cites (one click — from an AI referrer, if the analytics setup can even recognize it), and mentions two shortlisted tools to a colleague.
3. 3Ten days later she Googles the winning brand by name, clicks the (branded) search ad, and starts a trial.

Now score this journey under each model:

• Last-click: 100% to branded paid search — the one channel that demonstrably did not create this demand.
• First-click: 100% to whatever was tracked first — probably that single mid-journey content click, or the branded ad if the citation click was lost. The actual first touch (the AI answer) isn't in the graph.
• Linear / time-decay / position-based: credit elaborately divided between the branded ad and one content visit — a weighted average of two nearly irrelevant data points.
• Data-driven: the same misallocation, performed with statistical confidence and at scale, across every journey like this one.

Multiply this by the volume of research now happening in AI assistants and AI Overviews — where a large share of informational queries end without any click at all — and the aggregate effect on reporting is predictable and already visible in many accounts: direct and branded-search conversions drift upward, prospecting channels look weaker, and last-click logic recommends feeding the harvest while starving the planting. The content, digital PR, reviews, and community presence that determine what AI assistants say about a brand earn zero credit under every model on this page.

1. 1Use DDA as your default lens, last-click as a cross-check. Comparing the two in GA4's model comparison report shows you where credit is genuinely contested — that gap is where your interesting questions live.
2. 2Make the invisible visible where you can. Properly segment AI-assistant referrals so the citation clicks you do get are recognized rather than lumped into referral noise — step-by-step setup here.
3. 3Add self-reported attribution. A mandatory “How did you hear about us?” field on signup or checkout is the cheapest measurement upgrade available. When 12% of respondents say “ChatGPT” and your analytics says 1% of sessions are AI referrals, you've measured the iceberg below the waterline.
4. 4Track leading indicators of invisible influence: branded search volume, direct-traffic trends, and share of voice in AI answers (tools compared here).
5. 5Settle big-money arguments with experiments, not models. Whether to cut a six-figure channel is a question for a geo holdout, not for switching from last-click to DDA.