What is the attribution delta?

The attribution delta — Δ last vs linear — is one subtraction per channel: attributed margin under last-touch minus attributed margin under linear. It compresses the whole model-comparison table into a single signed number that says what role a channel plays. Positive: the channel closes journeys other channels start. Negative: it starts journeys other channels close. Near zero: the models agree, either because journeys are single-touch or the channel sits evenly across them.

Formula

Δ = last-touch attributed margin − linear attributed margin

Worked example

Retargeting shows €42,000 of margin under last-touch but €18,000 under linear: Δ = +€24,000. Most of what last-touch credits to retargeting was earned across journeys that other channels fed. Meanwhile a social prospecting channel shows €6,000 under last-touch and €21,000 under linear: Δ = −€15,000 — it rarely closes, but it is present early in a lot of journeys that end elsewhere.

How Saldo Metrics computes it

v_attribution_model_comparison pivots fact_attribution into one row per channel with the summed attributed contribution margin under each model — last-touch, first-touch, linear, time-decay, and Markov where populated. The delta subtracts the linear column from the last-touch column per channel, after both are converted to your display currency, so the subtraction happens on the numbers you actually see. Both inputs come from the same journeys, the same 30-day window and the same order margins; the only thing that differs is the credit rule, which is what makes the difference meaningful.

Why it matters

The delta is the cheapest funnel-role diagnostic in the app. Before moving budget on any single-model number, the delta tells you whether that number flatters or punishes the channel: a big positive delta warns that platform-style (last-touch) reporting is inflating the channel; a big negative one warns that cutting the channel on last-touch evidence would quietly damage the channels downstream of it.

Common mistakes

• Reading a positive delta as "bad channel". Closers are necessary; the delta says the channel's last-touch number overstates its solo contribution, not that the contribution is fake.
• Comparing deltas across orgs or periods with different journey lengths. More multi-touch journeys mechanically widen deltas everywhere; compare channels within the same period.
• Acting on a delta computed from thin data. A channel with a handful of attributed orders can swing sign month to month; check attributed order counts in the by-model table first.

Where you see this in the app

Marketing → Attribution, as the Δ last vs linear column of the model comparison.