Data-Driven Attribution: What It Needs to Work

What makes data-driven attribution different

Every model covered elsewhere on this site — last-touch, first-touch, linear, time-decay, position-based — applies a fixed, human-chosen rule to every path in the account. Data-driven attribution instead trains a model on your own historical conversion data: it compares paths that converted against paths that didn't, and learns which touch sequences and touch types actually correlate with a sale, then applies that learned weighting going forward. In principle, it removes the guesswork of picking a rule. In practice, it trades a transparent, checkable assumption for an opaque, data-hungry one — and the honesty of the trade depends entirely on whether you have enough data to make it.

The volume requirement, worked through

A useful rule of thumb from the broader ad-tech industry: reliable data-driven models need somewhere in the range of several hundred to a few thousand conversions a month, across enough path variation that the model can actually distinguish a common winning pattern from noise. Below that, the model either can't train at all, or trains on too few examples and produces weightings that shift dramatically from one reporting period to the next — a sign of overfitting to whatever happened to convert that particular month, not a genuine, stable pattern.

Take a mid-size DSP account running 400 attributed purchases a month. Split across five or six distinct path types — DSP-only, DSP-plus-sponsored, sponsored-only, streaming-assisted, and so on — that's often fewer than 80 conversions per path type, which is thin for a model to learn a reliable weighting from. Compare that to our own book: across 27 advertisers over 31 days in summer 2026, the portfolio logged 107,376 attributed purchases — enough volume, pooled, that a data-driven model would have something real to train on. Most single-brand accounts, especially newer or mid-size ones, simply aren't there yet.

Why the dispersion problem makes this worse

Even with enough raw volume, data-driven models are sensitive to how consistent the underlying performance is. In our own DSP book, individual advertiser ROAS ranged from 0.85x to 18.03x across 27 accounts in the same 31-day window, with a median of 4.30x — a wide spread pulled upward by a handful of strong retargeting-heavy accounts. A data-driven model trained across a pool with that much dispersion risks learning patterns specific to the outlier accounts rather than a genuinely representative one, unless the training set is large enough and varied enough to average that out. The honest question to ask a vendor offering data-driven attribution — including us, if we ever offered it as a packaged product — is what the median training-set size looked like, not just the total conversion count.

This is also why pooling data across many accounts, the way an agency's own book naturally does, can be a genuine advantage for building a data-driven model — but only if the pooled accounts are similar enough in category and funnel shape that the learned pattern actually transfers. Pooling a supplement brand's paths with a furniture brand's paths into one model produces a weighting that fits neither particularly well, which is worth checking before assuming scale alone solves the volume problem.

The common mistake, including ours

The mistake is treating "data-driven" as a synonym for "more accurate" regardless of the data feeding it, because the name itself implies rigor. We've seen — and, early on, built — attribution dashboards that offered a data-driven weighting option to a client running well under the volume needed to train it reliably, because the toggle existed and turning it on looked more sophisticated than a plain linear split. The output changed meaningfully month to month with no real change in the account, which in hindsight was the model overfitting on thin data rather than genuinely learning anything. We now check conversion volume against a reasonable threshold before recommending a data-driven approach, and default to a transparent rule-based model — usually position-based or time-decay — below it.

When to use it, and when not to

Data-driven attribution earns its complexity at real scale: high-volume accounts, or pooled analysis across a large enough set of comparable accounts, where the model has enough genuine variation to learn from and enough conversions to make that learning stable period over period. Below that volume, a rule-based model — linear, time-decay, or position-based, chosen deliberately rather than defaulted to — is the more honest choice, because its assumptions are visible and checkable rather than buried in a training process nobody outside the vendor can audit.

What to check before trusting a vendor's data-driven claim

Ask three questions of any tool or partner offering data-driven attribution: what conversion volume trained the model, how often it retrains, and whether the weighting has been stable or has swung meaningfully between reporting periods with no real change in the account. A model retraining on genuinely growing volume should stabilise over time; one that keeps swinging is a sign the underlying data still isn't dense enough to support the sophistication being sold, whatever platform is presenting it.

It's also worth asking what happens during a low-volume month — a slow season, a stockout, a paused campaign. A model that keeps producing confident-looking weightings even when conversion volume has temporarily collapsed is either falling back to a simpler rule under the hood without saying so, or continuing to report a stale weighting from a higher-volume period as if it still applies. Either behaviour is fine, as long as the vendor is willing to say which one is happening — the failure mode worth watching for is a dashboard that looks equally confident regardless of the data feeding it.