In Brief
- Core Answer: A marketing attribution model is a rule for dividing credit for a conversion among the touchpoints that preceded it. Every model is a simplification, they disagree by design, and choosing one is a decision about which error you can live with.
- Why It Matters: Attribution models were already approximations. AI-mediated discovery breaks the touchpoint record itself, which changes which models remain defensible.
- Best For: Marketing and revenue leaders choosing an attribution approach, or defending the one they have.
Marketing attribution models assign credit for a conversion to the touchpoints that preceded it, according to a fixed rule. First-touch credits the first interaction, last-touch the final one, and multi-touch models spread credit across several. No model is correct; each encodes a different assumption about how influence works, and the disagreement between them is information rather than error.
The Marketing Attribution Models, and What Each Assumes
Model | Assumption | Systematic bias |
|---|---|---|
First-touch | Discovery is what matters | Overvalues awareness, ignores conversion work |
Last-touch | The final push closes it | Overvalues branded search and retargeting |
Linear | All touches count equally | Rewards volume of touches over quality |
Time-decay | Recent touches matter more | Penalises long consideration cycles |
Position-based (U-shaped) | Discovery and conversion dominate | Arbitrary weightings, usually 40/20/40 |
W-shaped | Adds opportunity creation as a third milestone | Requires reliable CRM stage data |
Data-driven | Weights learned from observed paths | Opaque; needs volume; inherits tracking gaps |
Every row shares one flaw: they all divide credit among recorded touchpoints. Influence that left no record receives zero credit under all seven models equally.
The Structural Problem: Unrecorded Influence
Attribution has always missed things — a conference conversation, a colleague's recommendation, a podcast. Historically these were a minority of influence in digital-led B2B, and the models were approximately useful.
Two developments changed the proportion. Privacy regulation and browser restrictions shortened cookie lifetimes and broke cross-device joins. Then AI-mediated discovery added a large category of influence that produces no touchpoint at all: a buyer asks an assistant, reads a summary that mentions you, and searches your brand name three days later. Every model records that as branded organic, first-touch and last-touch alike.
The consequence is not that attribution is useless. It is that the recorded path is now a biased sample of the real one, biased specifically against upper-funnel and unlinked influence — which is exactly what generative surfaces produce.
What to Do About It
1. Stop Treating One Model as Truth
Run at least two — typically first-touch and a position-based model — and compare. Where they agree, act with confidence. Where they diverge sharply, you have found a channel whose role is genuinely ambiguous, which is more useful than a single number that hides the ambiguity.
2. Add a Self-Reported Layer
One open question on the demo or contact form: "How did you first hear about us?" It is imprecise, subject to recall bias, and it catches an entire category of influence that no tracking can. In practice it is the only mechanism that surfaces assistant-mediated discovery at all, because the buyer is the only party that observed it.
Keep it open-text rather than a dropdown. Dropdowns constrain answers to the channels you already knew about, which defeats the purpose.
3. Classify What You Can
Some assistant traffic does identify itself through referrer or user agent. Segment it explicitly in analytics rather than letting it fall into direct. It is a floor on the real number, not the real number, but a measured floor is worth more than an unmeasured guess. The GA4 configuration is straightforward.
4. Baseline Branded Search
Branded search volume is the most reliable available proxy for uncredited exposure. Baseline it, then watch it against changes in AI visibility. A rise in branded search following a rise in mention rate, with no other campaign running, is real evidence — not proof, but the strongest signal available in this class.
5. Use Holdouts Where You Can
Geographic or segment holdouts measure incrementality directly and sidestep the attribution problem entirely. They are underused in B2B because sample sizes are small and patience is short, but where volume permits, a holdout answers the question the models are only approximating.
Choosing a Model, Practically
- Short cycle, few touchpoints, e-commerce-like. Last-touch is defensible and cheap. The error is small when the path is short.
- Long B2B cycle with a buying committee. W-shaped or position-based, paired with self-reported attribution. Last-touch will systematically credit branded search for demand created elsewhere.
- High volume, mature tracking. Data-driven, with the caveat that it inherits every gap in the underlying data and cannot be inspected when it produces something surprising.
- Heavy AI-mediated discovery. Any model, plus self-reported attribution and branded search baselines. The model choice matters less than acknowledging what it cannot see.
The Question Attribution Cannot Answer
Attribution divides credit for conversions that happened. It does not tell you what would have happened otherwise. A channel receiving forty percent of attributed revenue may be adding forty percent of incremental revenue, or it may be intercepting demand that would have converted anyway.
Branded search is the standard example: it attracts credit under last-touch and creates little incremental demand, because those buyers had already decided to find you. Cutting it would show a revenue drop under attribution and little change in reality — which is why holdout tests, not model refinements, are the correct tool for budget decisions of that kind.
The useful posture: use attribution for directional allocation and diagnosis, use incrementality testing for decisions with real money behind them, and use self-reported data to catch what neither can see. Precision claims beyond that should be treated with suspicion regardless of which vendor is making them. Measuring AI-attributed revenue honestly works the same way — triangulation rather than a single number.
Implementing a Model Without Rebuilding Everything
Most attribution projects fail on data plumbing rather than on model choice. Five prerequisites, in the order they usually break:
- Consistent UTM discipline. If campaign parameters are applied inconsistently, every model downstream is computing on noise. This is a governance problem with a documented convention as its solution, not a tooling problem.
- Identity resolution. Anonymous sessions must join to a known contact when a form is submitted, and that contact must join to an account and an opportunity. Where the chain breaks, credit is lost silently.
- An agreed conversion event. Attribution to a form fill and attribution to closed revenue produce different pictures and different budget decisions. Pick one as primary, and be explicit about which.
- A lookback window that matches the sales cycle. A ninety-day window on a two-hundred-day cycle discards the first two-thirds of the journey and then reports confidently on what remains.
- Offline touch capture. Events, calls and partner introductions do not self-record. Where they matter, someone has to log them, and where nobody does, they do not exist in any model.
Get these right and a simple position-based model produces useful output. Get them wrong and a sophisticated data-driven model produces confident output that is wrong in ways nobody can inspect.
Reading Disagreement Between Models
Running two models is only valuable if you know how to read the gap. Three patterns recur:
- Channel high on first-touch, low on last-touch. It creates demand and does not close it. Cutting it will look free for a quarter and expensive after two. This is the classic profile of content, organic and most upper-funnel activity — and increasingly of AI-surface visibility.
- Channel low on first-touch, high on last-touch. It harvests demand created elsewhere. Valuable, but its measured performance depends entirely on something else continuing to work. Branded search and retargeting live here.
- Channel high on both. Either genuinely excellent, or a tracking artefact — check whether it is the default bucket for unattributed traffic before celebrating. Direct traffic appearing strong on both is almost always the second.
That last check matters more each year. As unrecorded influence grows, the direct bucket inflates, and a model that credits direct traffic is crediting a residual category rather than a channel. Segmenting known assistant referrers out of it is the minimum hygiene, and comparing the remainder against branded search movement is the practical way to estimate what is hiding there.
Frequently Asked Questions
What is a marketing attribution model?
A marketing attribution model is a rule for dividing credit for a conversion among the touchpoints that preceded it. Common models include first-touch, last-touch, linear, time-decay, position-based, W-shaped and data-driven.
Which attribution model is most accurate?
None is accurate in an absolute sense; each encodes a different assumption and carries a systematic bias. Long B2B cycles are usually better served by position-based or W-shaped models paired with self-reported attribution.
What is the difference between first-touch and last-touch attribution?
First-touch credits the initial interaction and overvalues awareness activity. Last-touch credits the final interaction and overvalues branded search and retargeting. They disagree by design, and comparing them is more informative than choosing one.
Why does attribution miss AI search traffic?
Assistant referrals frequently arrive with no referrer header or campaign parameter, so analytics classifies them as direct. The buyer may also see a mention and search your brand days later, which every model records as branded organic.
What is the difference between attribution and incrementality?
Attribution divides credit for conversions that occurred. Incrementality measures what would have happened without the activity, usually through holdout tests. Budget decisions should rest on incrementality; attribution is better for diagnosis and directional allocation.
Should we use data-driven attribution?
It works where conversion volume is high and tracking is complete. It inherits every gap in the underlying data and cannot be inspected when it produces a surprising result, so pair it with a simple model as a sanity check.