Marketing attribution models explained: last-click, first-click, time-decay, position-based and data-driven, plus the bias each one builds in.
Attribution is the most argued-about subject in marketing measurement and the least productive, because most of the argument is a search for the model that is correct. There isn't one. Every attribution model is a rule for dividing credit among touchpoints that all happened, and every rule is wrong in a direction you can name in advance.
That reframing is what makes attribution useful. Once a team stops looking for truth and starts choosing a known bias deliberately, the conversation shifts from which model is right to which distortion is acceptable for the decision in front of us.
An attribution model assigns credit for a conversion across the interactions that preceded it. A customer might see a paid social ad, read an article, search the brand name, click an email, and buy. Five touchpoints, one sale, and a question with no factual answer: which one caused it?
There is no observable answer because the counterfactual is unobservable. You cannot see what that customer would have done had the ad not existed. Attribution does not solve this. It applies a rule so that reporting can proceed. Understanding that attribution is a convention rather than a measurement is the single most useful thing a marketing leader can know about it.
All credit to the final interaction before conversion. It remains the most widely used model because it is simple, unambiguous, and available everywhere.
Its bias is well understood: it over-credits whatever sits closest to the purchase and undervalues everything that created the demand. Branded search is the standard illustration. Someone who discovers a company through a campaign, considers it for two weeks, then searches the company name and clicks — that sale is attributed entirely to branded search, a channel that mostly intercepted an intention already formed.
The predictable consequence is a budget that drifts toward capture and away from creation, until growth stalls because nothing is generating new demand. The stall usually arrives quarters after the drift begins, which makes the cause hard to see from inside the reports.
Last-click is defensible for short, low-consideration purchases where the journey genuinely is short. It is a poor basis for a considered B2B or high-value purchase, which is most of what the Saudi and wider GCC market actually sells.
All credit to the first recorded interaction. The mirror image, with the mirrored bias: it over-values discovery and ignores everything that converted the interest into a decision.
It is occasionally useful as a deliberate corrective — running first-click alongside last-click brackets the range and shows which channels are systematically over- or under-credited depending on where you stand. As a primary model it fails for the obvious reason that the first touch of a long journey is often nearly irrelevant to the outcome.
It also has a practical weakness: "first" is limited by measurement window and cookie lifetime, so the first recorded touch is frequently not the first actual touch. The model rewards whatever the tracking happened to catch.
These distribute credit across touchpoints rather than concentrating it.
Linear splits credit evenly. Its virtue is that it refuses to pretend it knows which touch mattered. Its weakness is the same: treating a display impression and a demo request as equal contributors is tidy and implausible.
Time-decay weights recent touches more heavily. It suits longer sales cycles where late interactions plausibly carry more decision weight, and it retains a last-click flavour in gentler form. The decay half-life is a judgement, and it is rarely justified from evidence — most teams accept a default without asking whether it matches their sales cycle.
Position-based, often forty per cent to the first touch, forty to the last, twenty spread across the middle, encodes the intuition that discovery and closing matter most. It is a reasonable compromise and entirely arbitrary. The weights are asserted, not derived.
All three are improvements on single-touch models for considered purchases, and none is more true than the others. They are different guesses with different shapes.
Data-driven models use observed conversion paths to assign credit statistically rather than by rule, typically by comparing paths that converted with paths that did not.
This is genuinely more sophisticated, and it carries two costs that are often glossed over. It needs substantial conversion volume to be stable, so smaller advertisers frequently get a model fitted to noise. And it is difficult to inspect — when the model says a channel deserves more credit, explaining why to a finance director is not straightforward.
There is a further limitation that applies to every model in this article: data-driven attribution still only sees the touchpoints it can observe. Interactions outside the measurable set — a conversation, an offline impression, a recommendation — are invisible, and their credit is silently redistributed to whatever the system can see. The model appears rigorous while being blind in exactly the places that matter for brand-driven demand.
Three limitations survive any choice of model, which is why model selection deserves less time than it gets.
Attribution cannot establish causation. It divides credit among observed touchpoints; it cannot tell you whether the conversion would have occurred anyway. A channel can receive high attributed credit while adding nothing incremental. Only a deliberate absence — a holdout, a geo test, a scheduled pause — answers that, which is why structured experimentation is the necessary companion to any attribution model rather than an optional extra.
Attribution cannot see what it does not track. Privacy changes, cookie restrictions, cross-device journeys and offline influence all remove touchpoints from the record. Their credit does not vanish; it is reassigned to the survivors, quietly inflating whatever remains measurable.
Attribution says nothing about quality. A model credits conversions, not good conversions. A channel producing high volumes of poor-fit leads can look efficient in every model until the numbers are joined to what actually closed. That join is the fix, and it belongs in the tracking layer rather than the attribution layer.
The practical recommendation is unglamorous. Pick one default model appropriate to your sales cycle as part of a wider marketing measurement framework, write down the bias you are accepting, and keep it stable long enough for comparisons to mean something.
For short cycles, last-click is adequate and its distortion is small. For considered purchases, time-decay or position-based is a better default. Data-driven is worth adopting where volume genuinely supports it, with the inspectability cost understood.
Then supplement rather than switch. When a decision is large — a significant reallocation, a channel exit, a budget increase — bring a second lens. Run a holdout. Compare periods around a deliberate change. Look at blended efficiency against total spend, which is model-independent and harder to fool.
The instinct to change models when results are disappointing should be resisted specifically. Switching attribution mid-year makes the trend line unreadable, and it produces apparent improvements that are artefacts of the change. If a model is switched, the honest move is to restate history under the new model before drawing any conclusion.
Attribution disputes intensify when different agencies own different channels, because each partner reports through the lens that favours it. Every channel's own platform will happily claim credit for the same sale, and the sum of confident claims will exceed the revenue that exists.
This is not usually dishonesty. It is the predictable result of asking each participant to score their own contribution. The structural remedy is a single attribution standard owned by the business rather than by any supplier, reconciled to the system of record. Where performance strategy and measurement sit together, that reconciliation happens by default; where they are split across vendors, it happens only if someone insists on it.
Three checks, all cheap. If total attributed revenue exceeds actual recorded revenue, the models are double-counting and the reporting is fiction. If a channel's attributed performance is strong but pausing it produces no measurable revenue decline, the attribution is crediting interception rather than creation. If lead volume rises while closed revenue does not, the model is counting conversions that the business does not value.
Any of those is a signal to stop refining the model and start testing causation. Attribution is a reporting convention. Incrementality is the evidence. Confusing the two is how well-instrumented teams end up confidently wrong.
The useful question is not which attribution model is correct. It is which decisions you are making that would change if the model you chose were biased in exactly the way it is.
Search is quietly changing shape. For twenty years the job was to earn a position in a list of links and wait for the click. Increasingly, the answer arrives before the list does — assembled by a language model, delivered in a paragraph, with a handful of sources credited underneath. Generative Engine Optimization is the discipline of making sure your brand is one of those sources.
The shift matters commercially, not just technically. If a potential client asks an AI assistant which firms handle performance marketing in Riyadh and receives a confident three-sentence answer naming three companies, the competition for that query was decided before any website was visited. Ranking fourth on a page nobody scrolls to is not a consolation prize. It is invisibility with extra steps.
Generative Engine Optimization, usually shortened to GEO, is the practice of making a brand's content retrievable, quotable, and attributable by AI answer engines — Google's AI Overviews, ChatGPT's search mode, Perplexity, and Microsoft Copilot among them. The objective is citation and inclusion rather than a numbered position.
In classic search, the unit of competition is the page. In generative search, the unit of competition is closer to the passage. The system is looking for a piece of text that cleanly answers the question it is trying to resolve. A page can be excellent overall and still be passed over because no individual passage inside it states an answer plainly enough to lift.
The honest case for acting early is not that GEO is a solved discipline. It is that it is an unsolved one, and the cost of entry is currently low. Generative search has no incumbency yet — the brands being cited today are frequently the ones whose content happens to be structured in a way the model can use, not the ones with the largest domain authority.
That window will close. As more organisations publish specifically for retrieval, the same accumulation dynamics that made classic SEO expensive will apply here too. The advantage available in the next year is a timing advantage, and timing advantages expire.