Last-Click vs. Multi-Touch Attribution: Which Model Is Telling You the Truth?

Ask two analysts which campaign drove last month's revenue and you can get two confident, contradictory answers — from the same data. The difference is not the data. It is the attribution model: the rule that decides which click gets credit for each order. Change the rule, change the winner.

This post walks through what last-click and multi-touch actually credit, when the simple model is honest enough, when it quietly lies, and why the deciding factor is not which model is "correct" — it is whether you can inspect and reproduce whatever model you pick.

What Each Model Actually Credits

An attribution model answers one question: when an order had several ad clicks before it, who gets the revenue?

• Last-click gives 100% of the order value to the final click before purchase. Everything earlier is invisible.
• Multi-touch splits the value across every recorded click in the journey, using a weighting rule — linear, position-based, or time-decay.

Here is the difference on a single journey. The numbers are illustrative math, not benchmarks: a customer clicks a Meta prospecting ad on day 1, a TikTok retargeting ad on day 6, and a Google branded search ad on day 9, then buys for $100.

Same order, same clicks, four different "truths." Under last-click, the Meta campaign that started the journey looks like dead spend. Under position-based, it is your second-best performer. Neither table cell is a measurement — each is a policy decision about how to assign credit.

When Last-Click Is Honest Enough

Last-click gets a worse reputation than it deserves. There are real businesses where it is approximately the truth:

• Short purchase cycles. If most customers buy within hours of their first click, there usually is only one click. A model that splits credit across one touch and a model that gives it all to one touch agree by definition.
• A single dominant channel. If 90% of your paid traffic is one platform, cross-channel credit disputes barely exist. The interesting question is which campaign inside the channel — and last-click answers that cleanly.
• Direct-response offers. Impulse-priced products, flash sales, lead magnets: the click and the conversion are the same session. There is no journey to model.

Last-click also has two underrated virtues: it is deterministic (anyone can recompute it from the click log) and it is explainable in one sentence. For a small account, a simple model you fully understand beats a sophisticated model you have to take on faith.

When Last-Click Lies

The model breaks when journeys get longer than the rule:

• Long consideration cycles. High-ticket products, B2B, education — anything where a customer researches for weeks. The prospecting ads that created demand get $0 every time, because by purchase day the last click is almost always a branded search or a retargeting ad.
• Multi-channel mixes. Run Meta for discovery and Google for capture, and last-click will systematically tell you to cut Meta. Do it, and a few weeks later your "high-performing" branded search campaigns have fewer buyers to capture — because the channel that filled the pipeline is gone.
• Retargeting self-dealing. Retargeting clicks sit closest to the purchase by design. Under last-click, the campaign that merely reminded an already-convinced buyer outranks the campaign that convinced them.

The pattern: last-click rewards whoever shows up last, and the channels that show up last are the cheap, low-funnel ones. Optimize against that signal long enough and you starve the top of the funnel that made it look good.

Linear, Position-Based, Time-Decay: The Multi-Touch Menu

Multi-touch is not one model — it is a family, and each member encodes a different belief:

• Linear splits credit equally across all clicks. Belief: every touch mattered the same. Simple and neutral, but it flatters filler touches — five mediocre retargeting clicks dilute the one ad that did the work.
• Position-based (often 40/20/40) overweights the first and last clicks and shares the rest with the middle. Belief: starting the journey and closing it are the two hard jobs. A reasonable default for discovery-plus-capture channel mixes.
• Time-decay gives more weight the closer a click sits to the purchase, usually with a half-life (a click 7 days out is worth half a click on purchase day). Belief: recency signals influence. The gentlest step away from last-click — it still favors closers, just not exclusively.

None of these is "the accurate one." Each is a defensible assumption made explicit. That framing matters, because it leads to the real requirement.

The Model Must Be Inspectable — Never a Black Box

Platform dashboards increasingly report "data-driven attribution": a proprietary model whose weights you cannot see, applied by the same party selling you the ads (Google's attribution models, Meta's attribution settings). Whatever its statistical merits, you cannot audit it, you cannot recompute last quarter under it, and when its numbers move you cannot tell whether your marketing changed or the model did.

Your own attribution should clear three bars:

1. Inspectable. For any order, you can pull up the exact clicks that were matched to it and the weight each one received. Credit is evidence, not a score from an oracle.
2. Reproducible. Same clicks plus same model version equals same answer, every time you run it. If the model evolves, old results are stamped with the version that produced them, so history does not silently rewrite itself.
3. Comparable. You can run last-click and a multi-touch model over the same click log and diff them. The campaigns whose credit swings hardest between models are exactly the ones where the model choice is doing the deciding — look there first.

This is how the attribution layer in Decisa is built: every attributed conversion stores which click matched, under which model, at which model version — so a number on a dashboard can always be traced back to the clicks behind it.

How to Choose, Practically

1. Measure your journey length first. What share of orders had more than one recorded click? If it is small, last-click is honest enough — stop here and enjoy the simplicity.
2. If journeys are long and multi-channel, run two models in parallel — last-click and one multi-touch (position-based or time-decay) — over the same data.
3. Act on the deltas, not the absolutes. A campaign at $0 under last-click but meaningful credit under multi-touch is your undercredited demand creator. Cut it last, not first.
4. Write down the model and version you used next to any budget decision it informed. Future-you will need to know which rule produced the number.

The model you pick will shape where your budget goes. That is unavoidable. What is avoidable is not knowing which rule is doing the shaping — pick a model you can read, rerun, and argue with.