First-touch vs last-touch attribution

First-touch attribution credits the channel that first brought a visitor to your site; last-touch credits the one active immediately before they paid. First-touch answers "which marketing finds people who eventually buy", which is the question a marketing budget is actually asking — last-touch systematically over-credits direct visits and branded search, because those are what people do once they have already decided.

What each model actually credits

An attribution model is a rule for deciding which marketing gets the credit when someone pays you. First-touch gives the whole sale to the source that brought the visitor to your site the very first time, however long ago. Last-touch gives it to whatever source was attached to the visit immediately before the payment. Both award 100% to a single touch; they disagree only about which end of the path counts.

The disagreement is not evenly spread. When a visitor clicks an ad and buys in the same session, both models agree. They diverge on the visitors who took time to decide — the ones you spent the most to acquire.

One customer, followed all the way through

A visitor searches for something your blog post ranks for, clicks the result, reads for four minutes and leaves. Eight days later they type your domain into the address bar, land on the pricing page, and pay $79.

Two touches. The first arrives with a search-engine referrer, so it classifies as organic search. The second has no referrer and no campaign parameters — a typed URL is indistinguishable from a bookmark or a referrer-stripped link — so it classifies as direct.

First-touch books $79 against organic search. Last-touch books $79 against direct. Nothing about the customer changed; only the rule did. Follow each into the decision it produces: first-touch says the blog post earned $79, so write more posts like it; last-touch says direct earned $79, and nobody can buy more direct. That $79 drops out of the marketing conversation, and the post shows up as having earned nothing.

Last-touch is not wrong about what happened. It just answers a question the decision in front of you — write more posts like that one, or not — does not depend on.

The two side by side

The bias each one carries

Last-touch over-credits direct and branded search structurally: those are what people do once they have already decided. Typing your domain, or searching your product by name, is the behaviour of someone the marketing already worked on — so the model puts a row you cannot buy at the top of the ranking.

The same bias flatters anything positioned late by design. A retargeting campaign sits between the visitor and the checkout on purpose, so it wins under last-touch whether or not it changed anyone’s mind — widen its audience to people who were buying anyway and its measured return goes up.

First-touch has the mirror-image problem: it cannot see what closed. If half your buyers need a nurture sequence to get over the line, first-touch reports nothing about it, and you could cut that sequence without the report objecting. It also hands full credit to whatever channel happened to be early, flattering broad-reach channels that catch people who would have found you anyway. Neither bias is fixable by care with the data; they are properties of the rule.

Where multi-touch and time-decay sit

Multi-touch models split the credit instead of awarding it whole. Linear splits it evenly across every touch; position-based gives most of it to the first and last; time-decay weights touches nearer the sale more heavily. All of them represent a path with several touches more faithfully than either single-touch rule. Faithful is not the same as useful.

The reason most small teams should not start there is that they need more data, and more trust in it, than a small site has. Take 40 sales a month across five channels averaging three touches each: that spreads over dozens of path shapes, most holding one or two sales. The split you read off it is noise, and it reorders itself next month.

The trust problem is worse. Every model divides credit only between the touches it can see, and touches go missing routinely — an ad blocker, a stripped referrer, a cleared cookie, a hop from mobile to desktop. Under a single-touch rule a missing touch is usually visible, as a direct or unattributed visitor. Under a split it quietly redistributes credit to the survivors, producing a plausible number with nothing flagged. Move to split credit once you have enough conversions that the split stops moving week to week.

Freeze the model at capture time

The choice has to be made before the data is recorded, not applied to it afterwards, because it determines what gets written down at all.

First touch is only knowable if something recorded it at the moment it happened and then refused to overwrite it. A system that keeps a “current source” per visitor and updates it every visit has not hidden your first-touch data — it has destroyed it, and the customer above would read as direct forever.

Freezing also buys reproducibility: if the credited channel is recomputed at read time, a report you ran in March returns a different answer in April — same customer, same $79, different row, because they came back through a newsletter link since.

Changing models mid-stream breaks the time series itself. Switch from last-touch to first-touch halfway through a quarter and direct collapses while organic and social jump on the changeover date — a step that reads exactly like a campaign working. If you must switch, recut the history under the new model, or pick a cut date and never compare across it.

Which to pick

Default to first-touch for deciding where money goes. The budget question is where to find more people like the ones who already paid you, and first touch is the only credit rule that answers it: it is the touch you chose, bought, and can buy more of. Ranking by first-touch revenue rather than traffic is what makes a small, high-intent channel visible — see attributing revenue to marketing channels for how that ranking is built, and revenue per visitor for the metric that keeps the comparison fair across channels of different sizes.

Use last-touch when the thing you are evaluating really is a closing mechanism and you already know where the demand came from: a cart-abandonment sequence, a checkout-page offer, a retargeting campaign whose only job is to convert people who arrived some other way. There, the last touch is precisely the question.

Two narrower cases. If your purchases are same-session — low price, little consideration — the models agree on most visitors anyway. And an ad network’s own dashboard credits the click it served, on its own model and its own attribution window, so a first-touch report will not tie out to it, and should not be expected to.

How Statlark does it

Statlark is first-touch, frozen at write time. The source, medium, campaign, ad-click id, referrer, landing path, country, device, browser and OS of a visitor’s first visit are written once and never overwritten; a later visit only moves the last-seen timestamp. Every revenue rollup keys on those frozen attributes, so a sale is credited to the channel, entry page, country and device of the visit that first found the buyer, not the session they checked out in.

Channels are a six-way classification of that frozen first touch, first match wins: Ads, Email, Social, Organic, Referral, and Direct — the case where there is no referrer and no source at all. Revenue in those rollups is net of refunds, and a payment in a currency other than the site currency is left out rather than converted at a guessed rate. Across devices, a person’s acquisition channel is the first touch of their earliest device active in the range you are looking at, which is the cut lifetime value by channel is built on — see Reading your data for what each number means on screen.

There is no last-touch or multi-touch report today, and no model switch — those are planned, and this page is the argument for why first touch is the one to start from. The per-visit data is kept, though: every event carries its own attribution snapshot, and a person’s timeline labels each event with the channel that visit arrived through. So you can read the closing touch on one customer, just not roll the site up that way.

One limitation: a frozen first touch is only as durable as the visitor id it hangs on. The first-party cookie lasts 365 days. In cookieless mode there is no cookie — the id comes from a salt that rotates every UTC day — so a visitor returning the next day counts as new, and the eight-day gap above becomes two unrelated people.