Why the choice of model matters
Give the same customer journeys to six attribution models and you will get six different league tables of channels. Nothing about the customers changed; only the rule for sharing credit did. That is the most important thing to understand about attribution models: they do not discover the truth, they apply a point of view.
The practical consequence is that a model choice is a budget choice. Switch from last click to a model that credits early touches and display, video and social suddenly look better while branded search looks worse. Teams should pick a model knowing whose work it flatters, and hold it steady long enough to compare like with like.
Changing the model changes the scoreboard, not the match.
The six models in brief
Consider a journey with four touches: a social video ad, an organic search visit, an email and a branded paid search click before purchase. Here is how each model would treat it.
| Model | How credit is split | Flatters | Penalises |
|---|---|---|---|
| Last click | All to the final touch (branded search) | Closing channels: branded search, retargeting, email | Awareness and discovery |
| First click | All to the first touch (social video) | Discovery channels | Nurture and closing |
| Linear | Equal shares to all four | Channels with many touches | Decisive single touches |
| Time decay | More to recent touches, less to early ones | Late-stage channels | Early discovery |
| Position-based | Large shares to first and last, remainder spread across the middle | Openers and closers | Middle-of-journey nurture |
| Data-driven | Weights learned from converting and non-converting paths | Depends on your data | Channels the tool cannot observe |
Note that availability varies by tool and changes over time. GA4, for instance, has removed several rule-based models from its standard options and now centres on data-driven attribution and last click variants. Check your platform's current documentation before planning around a particular model.
Last click: simple, stable and biased
Last click gives all credit to the final click before conversion (usually ignoring direct visits, depending on the tool). It is easy to explain and consistent over time, which is why finance teams often like it. Its bias is severe and predictable: it rewards whoever is standing at the till.
Branded search, retargeting and email to existing subscribers look excellent under last click because they intercept people who were already on their way. Last click is a reasonable lens for optimising within a closing channel and a poor one for deciding how much to invest in creating demand.
First click and linear: correcting one bias with another
First click reverses the bias, crediting whatever introduced the customer. It is useful as a diagnostic (which channels bring in new people?) but rarely as a primary model, because it ignores everything that persuaded them afterwards.
Linear treats every touch as equal. It feels fair, and it is a decent antidote to last click in a review meeting. But it rewards volume of touches rather than influence: a channel that fires many low-value reminders gains more credit than one decisive moment. Both models are best used alongside last click to see how much the picture moves.
Time decay and position-based: shaped compromises
Time decay gives more weight to touches closer to the conversion, which suits short buying cycles where recent interactions plausibly matter more. Position-based (sometimes called U-shaped) gives large shares to the first and last touches and spreads the rest across the middle, reflecting a belief that discovery and closing matter most.
Both are reasonable hypotheses about how buying works. Neither is evidence. The weights are chosen by convention, not measured from your customers, so their outputs should be read as 'what if discovery matters this much' rather than as findings.
Calculator
See how models split the same sale
Illustration only. Enter a conversion value and the number of touches in a journey. Position-based here uses a common convention of 40% to the first touch, 40% to the last and 20% shared by the middle touches.
Last click: credit to final touch
₹10,000
Every other touch receives nothing.
= value
Linear: credit per touch
₹2,500
Identical for every touch.
= value / touches
Position-based: first or last touch
₹4,000
Each of the two ends.
= value * 0.4
Position-based: each middle touch
₹1,000
Middle touches share what remains.
= value * 0.2 / (touches - 2)
Defaults are illustrations. Use your own numbers. Nothing you enter leaves this page.
Data-driven attribution: learned, not assumed
Data-driven models compare the paths of people who converted with those who did not and estimate how much each touchpoint changes the likelihood of conversion. Credit is then distributed according to those estimates. In principle this replaces arbitrary weights with patterns from your own data.
In practice there are limits. The model can learn only from touchpoints it can observe, so views, offline media and untracked channels are invisible to it. It needs enough conversion volume to learn from, and its internal logic is usually not fully disclosed. It remains correlational: it finds touchpoints associated with conversion, which is not the same as proving they caused it.
Fig. 01 · Matrix
Tap to explore
Attribution models by transparency and adaptiveness
How to choose a model
Fig. 02 · Scorecard
Tap to explore
What to weigh when choosing
Bars show relative emphasis, not measured data
Our recommendation for most organisations is straightforward. Use the data-driven model your main analytics tool offers as the primary view if you have adequate volume, keep last click as a stable comparison, and never let the difference between them alone decide a large budget change. Use incrementality testing for that.
Self-diagnostic
0/4Are you using models responsibly?
A quick test for any team that reports attributed results.
01Is the model named on every report that shows attributed figures?
If yes: Readers can interpret the numbers correctly. If no: Add it. An unnamed model invites false certainty.02Have you kept the same primary model for at least two quarters?
If yes: Trends are comparable. If no: Settle on one; switching hides real change behind method change.03Do you compare at least two models when reviewing channel mix?
If yes: You can see which channels are model-sensitive. If no: Add last click beside your primary model for perspective.04Are large budget shifts validated with a test or model before committing?
If yes: Good governance. If no: Use a holdout or geo test before moving major spend.
Model choice by business type
Ecommerce and D2C brands with short cycles and high volume are the natural home of data-driven attribution: there are enough paths to learn from and the purchase sits close to the marketing. The main risk is over-crediting retargeting and branded search, which a periodic holdout test corrects.
B2B and high-consideration services rarely suit web attribution models at all. Journeys last months, involve several people and end in a sales conversation the website never sees. Here the CRM should be the system of record: capture first and most recent source on each lead, attribute pipeline and revenue by campaign, and send qualified stages back to ad platforms. A simple, explainable model on good CRM data beats a sophisticated model on web sessions.
Businesses with heavy offline media (television, outdoor, print, retail presence) should treat any digital attribution model as partial by design and lean on mix modelling and geo experiments for channel decisions.
Beyond models: the measurement stack
Attribution models are one layer. Above them sit experiments, which estimate cause directly, and marketing mix modelling, which estimates channel contribution from aggregate data including offline media. The teams that make the best decisions use all three and reconcile them, rather than searching for the perfect attribution model. Start with our explainer on marketing attribution for the wider context, and see ROAS versus CPA for how attributed figures feed bidding.
Key takeaways
- 01Every attribution model encodes a bias about how buying works; none reveals the truth on its own.
- 02Last click flatters closing channels; first click flatters discovery; linear rewards touch volume.
- 03Data-driven models learn weights from your paths but cannot see untracked or offline touchpoints.
- 04Choose one primary model, name it on every report and keep it stable for comparability.
- 05Validate large budget moves with experiments or mix modelling, not with a model switch.
Frequently asked
- What is the most accurate attribution model?
- No model is accurate in the sense of measuring cause. Data-driven models adapt to your own data and are usually a better default than fixed rules when volume allows. Accuracy about what marketing actually added comes from controlled experiments, which can then be used to calibrate whichever model you use.
- What attribution models does GA4 offer?
- GA4 has changed its options over time and currently centres on data-driven attribution alongside last click variants, having retired several rule-based models. Settings exist at property level and in comparison reports. Because the menu changes, check Google's current documentation before building a process around a specific model.
- Is last-click attribution still useful?
- Yes, as a stable, transparent comparison and for optimising within closing channels such as search. It is misleading for deciding investment in awareness and discovery, because it ignores everything before the final click. Keep it as a reference view rather than the sole basis for channel budgets.
- What is position-based attribution?
- Position-based, or U-shaped, attribution gives most credit to the first and last touches and divides the remainder among the middle ones. A common convention gives large equal shares to the two ends. It reflects a belief that discovery and closing matter most, but the weights are chosen, not measured.
- Should I build my own attribution model?
- Only if you have clean, joined first-party data, enough conversions and analysts who can maintain it. A custom model can be more transparent than a platform's. For most organisations, effort is better spent on data quality, CRM source capture and running incrementality tests than on building bespoke models.
Published by Fabulous.Media, a network of specialist marketing agencies. Updated 9 October 2026. Platform features change often; check current official documentation before acting on platform-specific detail.
