Sales Attribution: Deciding Which Touches Get Credit for Revenue

Sales attribution is the set of rules a business uses to decide which marketing and sales interactions get credit for a closed deal. Buyers rarely arrive through a single door: they read an article, see an advertisement, attend a webinar, ask a colleague, search for the company by name and eventually speak to somebody. All of those contributed something, and attribution is an attempt to say how much. It is worth being clear from the outset that this is a modelling exercise rather than a measurement. The rules are chosen, not discovered, and the same journey produces very different answers depending on which rules you apply.

Single-touch models

First touch credits the interaction that created the lead. It answers where demand originates and systematically undervalues everything that happens afterwards. Last touch credits the final interaction before the deal. It answers what closed the sale and systematically overvalues channels that capture intent someone else created, most obviously branded search and direct visits.

Multi-touch models

Linear divides credit equally across every recorded touch, which is simple and treats a passing web visit as equal to an hour-long meeting. Time decay weights later touches more heavily, on the assumption that recent contact carried more influence. Position-based, often U-shaped, assigns large fixed shares to the first and last touches and splits the rest among the middle. Data-driven models derive the weights statistically from patterns across many completed journeys, which is the most defensible approach and requires more volume than most businesses have.

A worked example on one deal (illustrative)

A deal worth ₹1,00,000 has four recorded touches in order: an organic article, a paid advertisement, a webinar, and a direct visit to the pricing page. First touch credits the article with the whole ₹1,00,000. Last touch credits the direct visit with the whole ₹1,00,000. Linear gives ₹25,000 to each of the four. Position-based at a forty, twenty, forty weighting gives ₹40,000 to the article, ₹40,000 to the direct visit, and ₹10,000 each to the advertisement and the webinar. Time decay might give roughly ₹10,000, ₹15,000, ₹30,000 and ₹45,000 across the four in sequence. Nothing about the deal changed. The apparent value of the paid advertisement moved between zero and ₹25,000 purely on the basis of a modelling choice, which is why running more than one view is more honest than defending a single one.

The decision it drives is budget allocation. If a business is spending across six channels and can only grow three, it needs some basis beyond preference for choosing. Attribution supplies that basis, provided it is treated as evidence rather than accounting. The more useful version of the exercise is comparative across the funnel rather than a single revenue figure per channel: which sources produce enquiries that qualify, which produce deals that close, which produce larger deals, and which produce deals that close faster.

It also settles a recurring organisational argument. Marketing tends to argue from lead volume and sales from deal quality, and the two conversations never meet because they use different data. Attribution that follows a source all the way through to won revenue and loss reasons gives both sides the same picture, and it usually reveals that the highest-volume channel is not the most valuable one.

Mistaking it for causation

The most consequential error is reading attributed revenue as the value a channel created. Attribution describes the paths that were observed; it cannot say what would have happened had a channel not existed. Branded search is the standard illustration: it absorbs a great deal of credit while often converting demand that other activity created, and a business that cuts everything else on the strength of that report will find its branded search volume falling a quarter later.

Modelling only what is trackable

Every untracked influence, a recommendation from a peer, a conversation at an event, an internal advocate, a podcast, has its credit silently redistributed to the channels that do carry parameters. This does not make the model slightly wrong; it makes it systematically biased towards measurable channels. A self-reported source question on the enquiry form is crude, inconsistent and genuinely useful precisely because it captures what no tracking system can see.

Changing the model and comparing periods

Switching from last touch to position-based part way through a year and then comparing quarters produces conclusions that are entirely artefacts of the switch. If a model has to change, restate the prior periods on the new basis before drawing any comparison, and expect the restatement to be more work than the change itself.

Treat wide agreement between models as a strong signal and disagreement as a question rather than a problem. A channel that looks valuable under every model probably is. A channel whose apparent contribution swings dramatically between first touch and last touch is playing a specific role in the journey, and the right response is to understand that role rather than to pick the model that suits the argument.

Look past revenue to the shape of the business each channel produces. A source with modest attributed revenue but a high qualification rate, large deals and short cycles may deserve more investment than a source with high attributed revenue and a poor win rate. Pairing source data with loss reasons is usually the fastest route to finding a channel that generates volume from the wrong segment entirely.

Each model answers a slightly different question, and choosing one is choosing a question.

Sitting outside this table is incrementality testing, which is not a model at all but an experiment. Where a spending decision is large enough to matter, a holdout or a geographic split answers the causal question that no attribution model can.

Attribution is downstream of record keeping: if sources are not captured cleanly at the start, no model will rescue it. Start with what a CRM is, then look at lead management software for capture and source tracking, sales automation, AI CRM for scoring, free sales tools, and CRM versus Excel for why journeys cannot be reconstructed from a spreadsheet.