What is multi-touch attribution?

What MTA was built to fix

Single-touch models discard almost all of the evidence. A user who saw a video, read a review, clicked a search ad and then converted has produced four signals, and last click keeps one.

MTA keeps them all and applies weights — see attribution modeling for the common weighting schemes. Done well, it corrects the structural undervaluation of anything that is not the final click, which is most brand and upper-funnel activity.

What it requires, and why that became hard

MTA depends on stitching a single user's touchpoints together across sites, apps and devices. That stitching required a durable identifier.

Third-party cookie deprecation, ATT, and platform-level restrictions removed it. What remains is a partial view: touchpoints inside one platform or one device, with the rest of the path missing. An MTA model fed a systematically incomplete path does not degrade gracefully — it over-credits whatever is still visible, which is usually the channels closest to conversion. The model looks like it is working and is quietly reproducing last-click bias.

Where MTA still earns its place

• Inside a single ecosystem, where the identifier survives and the path is genuinely complete.
• On logged-in properties, where you have first-party identity across sessions.
• For directional channel comparison, treated as one input among several rather than as the answer.
• Not as a causal claim. It describes correlation in observed paths. Incrementality is the causal instrument.

MTA and AI surfaces

AI assistants add touchpoints that MTA cannot see at all. A user can be materially influenced by a recommendation inside a conversation and arrive at conversion with no recorded interaction, so the path MTA reconstructs simply omits the step that mattered.

Where the journey runs through agents rather than the user's own browsing, the shape changes further — the relevant chain is agent calls rather than human touchpoints, which is what A2A attribution is trying to model. Treating an AI-influenced conversion as an MTA gap to be closed with better tracking misreads the problem: the touchpoint was never observable.