Platform, last-touch, promo code or multi-touch? Social commerce attribution compared

Attribution methods answer how observed outcomes receive credit. They do not by themselves answer what social activity caused. Compare methods on the same event, order maturity, population, source, timezone and privacy constraints.

What to take away

Choose platform attribution for platform optimisation, a reconciled cross-channel rule for descriptive allocation, and an appropriate counterfactual design for causal decisions. Publishing them side by side is useful; blending them into one conversion total is not.

Platform click or view attribution

A platform can claim a conversion after an eligible click or view within its configured window. TikTok's attribution overview defines click-through, view-through and engaged-view attribution and explains the window concept. This method supports campaign operation inside that platform.

Limit: it is self-attribution under platform rules. Record event, window, reporting date and match method. Do not add its claimed conversions to another source without deduplication.

TikTok Shop UK and the London-based resale marketplace Depop each report conversions on their own terms. A seller active on both cannot add the two attributed counts together without deduplication.

Cross-channel last-touch attribution

This rule gives credit to the final measured touch before conversion. It is straightforward to explain and prevents multiple channels receiving full last-touch credit within one reconciled system.

Limit: it ignores earlier measured influence and any unmeasured interaction. TikTok's reporting-discrepancy page notes that methodology, windows, timezones and interaction-versus-event dating can make its figures differ from measurement partners. Align settings before interpreting the gap.

Referral or promotion-code attribution

Stable campaign links, creator codes or order tags can connect a transaction to a route. Define eligibility, duplicate codes, returns, leakage, expiry and orders with more than one marker. Use settled orders, not merely submitted checkouts.

Limit: a code records use, not necessarily cause. It may be shared outside the intended content, while customers influenced by content may not use it.

Multi-touch allocation

This method divides credit among recorded interactions according to a declared rule or model. Publish all eligible touchpoints, lookback, weights, missing-data treatment and whether the total credit per order is capped.

Limit: a different allocation rule can produce a different answer from the same journeys. Keep modelled credit separate from the finance ledger.

Hold settings constant during comparison

Record the conversion event, click and view windows, account timezone, reporting date basis and identity rule before the period starts. Changing a window can change credited outcomes without changing any underlying order. If a setting changes, mark a series break and report both definitions only where the source data supports a reproducible recalculation.

Use one mature order population for all methods. Reconcile cancellations, returns, refunds and duplicate order identifiers before comparing credited contribution. A method that reports quickly should not receive an advantage by using placed orders while another uses settled orders.

Incrementality design

An experiment or suitable quasi-experimental comparison estimates change against a counterfactual. The Magenta Book says comparison data should be sufficient and genuinely comparable, with contamination and expected noise considered.

Limit: design, sample, spillover, seasonality and simultaneous activity can undermine inference. A statistical reviewer should approve the plan before results are visible.

Make the comparison like for like

For each method record purpose, eligible population, event, source, window, timezone, identity and deduplication rule, order maturity, exclusions, privacy basis and uncertainty. Report attributed order count and settled contribution under the same finance definition.

For example, under a 7-day click and 1-day view setting, a click six days before the order sits inside the click window, while a view two days before it falls outside the view window. A 30-day multi-touch lookback can credit three social touchpoints that the platform rule never counted.

Attribution can depend on cookies, pixels, server events, account information or promotion codes. The ICO's storage and access technology guidance covers cookies and comparable methods under PECR and, where relevant, data-protection law. Assess the configured method, purpose, collection, consent position, recipients and retention. Do not select a more invasive method merely because it assigns more credit.

Publish unmatched orders and customers who cannot lawfully or technically be measured as limitations. Missing observation does not mean no influence, and modelled recovery should not be presented as a directly observed transaction.

Before you act

Common questions

Why can platform attribution figures differ from measurement partners?

TikTok's reporting-discrepancy page notes that methodology, windows, timezones and interaction-versus-event dating can make its figures differ from measurement partners. Align settings before interpreting the gap.

What should be recorded for each attribution method in a like-for-like comparison?

For each method record purpose, eligible population, event, source, window, timezone, identity and deduplication rule, order maturity, exclusions, privacy basis and uncertainty. Report attributed order count and settled contribution under the same finance definition.

When is an incrementality design appropriate?

An experiment or suitable quasi-experimental comparison estimates change against a counterfactual. The Magenta Book says comparison data should be sufficient and genuinely comparable, with contamination and expected noise considered. A statistical reviewer should approve the plan before results are visible.