Marketing Attribution Models Explained (Honestly)

Marketing attribution models are rules for assigning conversion credit to the touchpoints that preceded it. A buyer clicked a LinkedIn ad in March, read three blog posts in April, and clicked a Google search result in May before signing up — attribution decides which of those touches “gets” the signup, and therefore which budget line looks like it’s working.

The first honest thing to say: every model is a simplification. Attribution doesn’t measure causation; it distributes credit among the touches you happened to observe, according to a rule you chose. Change the rule and the same history produces different winners — which is why attribution debates are really debates about what question you’re asking.

The second honest thing: models can only credit touches your analytics saw. Ad blockers, Safari’s tracking prevention, cross-device journeys and consent declines all delete touches before any model runs — one reason we run first-party, cookieless analytics as the substrate. And some touches are structurally invisible: the podcast mention, the Slack recommendation. That’s the dark funnel, and no model rule fixes it.

Here is the standard lineup, with what each model is actually good for — and the blind spot its fans don’t mention:

To make the differences concrete: take one journey with four touches — LinkedIn ad, blog visit, webinar, branded search — ending in a signup. Here’s the credit the LinkedIn ad (the first touch) receives under each rule:

If you learned attribution in Universal Analytics, GA4 will surprise you: Google has retired most of the classic models. Per Google’s own attribution documentation, GA4’s attribution reports now offer exactly three options: data-driven attribution, paid-and-organic last click, and Google-paid-channels last click. First-click, linear, time-decay and position-based are gone.

Two more GA4 details worth knowing. Attribution models there exclude direct visits from receiving credit unless the entire path is direct — a reasonable choice that also quietly hides how much of your traffic is unattributable. And data-driven attribution is a per-account machine-learning model: genuinely clever, but not inspectable, and like every model it can only weigh the touches that were tracked in the first place.

The practical takeaway isn’t “GA4 bad.” It’s that even Google concluded the rule-based multi-touch models were creating more debate than insight. What’s left — a last-click view, an algorithmic view, and (elsewhere in this guide) a first-touch view plus asking humans — is honestly about the real menu.

B2B is where attribution models go to be humbled, for four structural reasons:

• Buying committees. The person who researched you, the person who signed, and the person who found you on a podcast are often three different people. Click-level attribution sees three unrelated visitors; the deal belongs to an account.
• Long cycles. A six-to-eighteen-month journey outlives cookies (Safari caps script-set storage aggressively — see WebKit’s tracking prevention policy), outlives attribution windows, and often outlives the campaign that started it.
• Cross-device and cross-context journeys. The LinkedIn scroll happened on a phone; the signup happened on a work laptop. Without an identity graph that stitches on email, those are two strangers.
• The dark funnel. Slack communities, DMs, podcasts, word of mouth — heavy influence, zero referrer data. We wrote a full guide on the dark funnel because it deserves one.

This is why the roughly 98% of visitors who never fill out a form matter so much in B2B: most of the journey happens anonymously, and the model only gets to run on the fraction that surfaces. Two partial fixes help: company-level visitor identification reveals which accounts are in the anonymous traffic, and self-reported attribution lets the buyer tell you about the touches software can’t see.

After all the honesty, here’s what we actually recommend — and run ourselves. It’s less a model than a pipeline that preserves evidence:

The parts are covered in their own guides: UTM discipline upstream, revenue attribution downstream, and the self-reported layer alongside. The reason this beats a fancier model: every step preserves raw evidence (this person, this first touch, this invoice) rather than a modeled abstraction, so you can always re-ask the question later.

Our unexciting answer: use first-touch and last-touch side by side, on revenue rather than conversions, and treat both as lenses. If a channel looks strong on first-touch and weak on last-touch, it creates demand others harvest — fund it. If the reverse, it harvests demand — keep it, but don’t credit it with creation. If a channel looks weak on both and self-reported answers never mention it, cut it.

Save data-driven models for when you have the volume they need, and even then, sanity-check them against the simple views and against what buyers literally tell you. The goal was never a perfect credit split — it’s better resource allocation than last quarter’s, decided with fewer illusions.