Attribution answers a deceptively simple question: which marketing effort caused this customer? The honest answer is that no model tells you the truth — each is a different systematic distortion. The skill isn't picking the "right" model; it's knowing what each one hides, and building decisions that survive the uncertainty.

The problem in one illustrative journey

A future customer: reads your comparison blog post (organic search) → sees a retargeting ad, doesn't click → clicks a colleague's shared link → two weeks later searches your brand name, clicks the ad above the organic result → signs up → converts to paid after a lifecycle email.

Who gets credit? Every model answers differently — and every answer reallocates real budget.

The models, and what each systematically distorts

First touch

100% of credit to the first recorded interaction. Tells you: what fills the top of the funnel. Distorts: ignores everything that converted the interest into money; also fragile, since the true first touch (a podcast mention, a conversation) is often invisible to your tools.

Last touch

100% to the final interaction before conversion. The default in most tools. Tells you: what closes. Distorts: systematically over-credits bottom-funnel harvesters — brand search, retargeting — that mostly collect demand created elsewhere. Budget optimized on last touch drifts toward buying your own brand name ever harder.

Linear

Equal credit to every touch in the journey. Tells you: the breadth of the path. Distorts: treats a 2-second ad impression and a 40-minute webinar as equals; "fair" in the way that splitting a dinner bill evenly among people who ordered very differently is fair.

Position-based (U-shaped and friends)

Typically 40% first, 40% last, 20% spread across the middle. A reasonable compromise codified as arbitrary percentages. Distorts: the weights are convention, not measurement.

Time decay

More credit the closer to conversion. Distorts: structurally biased against the discovery channels that start journeys — often the ones a startup most needs to evaluate.

Data-driven attribution

Algorithmic credit assignment, essentially by comparing conversion rates of journeys that include a touchpoint versus similar journeys that don't. Conceptually the best of the bunch. The catch for startups: it needs conversion volume — thousands per month — to produce stable estimates, plus it only sees trackable touches and remains correlational. Below that volume, its outputs are confident-looking noise.

The limits no model escapes

Dark funnel. Word of mouth, communities, podcasts, private chats — a large share of B2B journeys (often the decisive share) never appears in your tracking.
Tracking decay. Consent banners, ad blockers, cross-device journeys: your recorded paths are a biased sample of real ones.
Correlation ≠ incrementality. Attribution credits touches that appear in converting journeys. Brand search "converts" spectacularly — but many of those users would have found you anyway. The only ground truth for "did this spend cause customers" is incrementality testing: holdouts and geo splits.
Attribution windows are choices. A 30-day window makes long-consideration channels (content, SEO) look worthless by construction.

One more B2B wrinkle: accounts aren't individuals. The person who discovers you (a developer reading your docs), the person who evaluates you, and the person who signs are often three different people whose journeys never join in click-based tools. Account-level attribution — mapping all touches from the same company domain into one journey — is imperfect but far closer to how B2B buying actually works.

A pragmatic playbook for startup-sized data

You don't have the volume for data-driven attribution, and you don't need it. What works:

1. Run two simple models in parallel. First touch and last touch, side by side. Read them as different questions: first touch ranks demand creation, last touch ranks demand capture. When a channel looks great in one and absent in the other, you've learned what role it plays.

2. Add the question nobody can block: "How did you hear about us?" A free-text field at signup. Self-reported attribution is imprecise but sees the dark funnel that tracking can't. Illustratively, teams commonly find a channel like podcasts or communities at 25% of self-reported answers while showing near 0% in click-based models. Reconcile both views monthly.

3. Judge channels on cohort quality, not signup counts. Attribution's real payoff is per-channel unit economics: activation rate, retention, and CAC payback by first-touch channel. A channel delivering half the signups but 3× the 6-month retention wins the budget.

4. Test incrementality on your biggest line item. Before scaling any channel past ~30% of budget, run the crude but honest test: pause it (or a region) for 2–4 weeks and watch total signups, not channel-reported conversions. Channels have a way of claiming conversions that would have happened anyway; the pause reveals it.

5. Fix UTM hygiene first. No model survives garbage labels. One documented convention — lowercase, fixed vocabulary for source and medium, campaign naming pattern — and a monthly audit of "direct / (not set)" share.

Attribution by go-to-market motion

The right rigor level depends on how you sell:

Product-led, self-serve: journeys are short and mostly digital, so click-based models capture more of the truth. First/last touch plus the signup survey covers most of it. The subtlety is free-to-paid lag: attribute the paid conversion back to the signup's original channel, or channels that attract patient users will look worse than they are.
Sales-assisted: journeys stretch over months and run through humans. Click-based attribution collapses here; what matters is source-of-opportunity tracking in the CRM (with the same discipline you'd apply to UTMs) plus self-reporting. Weight channels by pipeline value created, not lead count — illustratively, a webinar producing 20 leads and two $15k opportunities beats a directory producing 200 leads and none.
Content and community-heavy motions: accept that your best channels will be structurally under-measured, and lean on directional signals: branded search volume trend, self-reported share, and periodic "pause tests" on paid spend to estimate the organic baseline.

In every motion, the same hierarchy holds: tracking data for day-to-day steering, self-reporting for the map, incrementality tests for the big-money verdicts.

What to report

A monthly channel review that respects the uncertainty:

Channel	Signups (first touch)	Signups (last touch)	Self-reported share	Activation rate	CAC payback
Organic search	210	140	18%	34%	6 mo
Brand search ads	30	180	2%	41%	2 mo
Communities	25	15	24%	52%	n/a

(Illustrative.) The brand-search row is the classic tell: huge on last touch, negligible self-reported — mostly harvested demand. The communities row is the opposite: invisible to tracking, loud in self-reporting, best cohort quality on the sheet.

The bottom line

Attribution models are lenses, not verdicts: first touch for creation, last touch for capture, self-reporting for the dark funnel, incrementality tests for truth on big bets, and cohort quality as the final judge. Hold them together, loosely, and spend accordingly.

Growth Pilot ties acquisition channels to downstream reality — activation, retention, and revenue per cohort from GA4 and Stripe — so channel debates end with quality data instead of last-click folklore.