Forrester predicts AI agents will influence 75% of B2B purchase decisions by 2027. Whatever you think of the exact number, the direction is not in dispute: buyers delegate the boring first pass of vendor research to machines, then bring human judgement to the final few options. The click that finally lands on your pricing page is the last mile of a journey that largely happened outside your analytics.
That is a problem for attribution, because most marketing attribution models were built to watch a human clicking through a funnel they walked alone. When the research happens inside a chat window, a private community, a podcast or an assistant's summary, your tracking sees nothing until the very end — and then credits whatever captured the click, not what earned the decision.
This is the Wednesday companion to our ROI series. Tuesday's piece covered how to measure what AI-assisted marketing returns; this one covers the other half of the question — where the revenue actually came from. Six steps, no machine-learning theatre.
What AI-Era Buying Actually Breaks
Last-click attribution doesn't die in the AI era. It gets demoted — from truth to trivia. Three failure modes do the demoting:
- The research went dark. Assistants, communities, newsletters, podcasts, DMs — the dark funnel where B2B buying increasingly starts leaves no referrer, no UTM, no cookie. Your analytics record a direct visit or a branded search, and the model shrugs.
- The click is a conclusion, not a step. When an AI-assisted buying journey compresses three weeks of comparison into one summary, the link the buyer clicks was chosen for them. Last-click over-credits capture and under-credits whatever taught the machine to recommend you in the first place.
- Committees decay self-reported data. A B2B decision has many stakeholders; the person who fills in your form is rarely the person who first found you. Single-respondent attribution inherits their blind spots.
What still works: last-click still measures bottom-funnel capture honestly. The goal is not to throw it out — it's to build a marketing attribution model for the AI era that can also see influence.
Step 1 — Start From the Decisions, Not the Data
A model built without a decision to serve becomes dashboard furniture. Before touching a tool, write down the two or three decisions this model must inform in the next two quarters: channel-mix shifts, kill-or-scale calls on a program, how much to move from capture to demand creation. Then name who consumes the output — if the answer includes a CFO, the model must produce CFO-recognisable outcomes (pipeline, closed revenue), not platform metrics.
This is the same discipline as our guide to attribution ROI: proving which 30% of budget actually works — the model exists to find the 20–30% you can reallocate with confidence, not to decorate a monthly deck.
Step 2 — Pick a Model Class You Can Actually Feed
| Model class | The question it answers | Data it needs | The honest limitation |
|---|---|---|---|
| Last-click | Which channel captured the demand | Basic web analytics | Blind to everything upstream |
| First-click | Which channel introduced you | Analytics + strict UTM discipline | Ignores everything that nurtured |
| Position-based (U/W-shape) | Weighted credit for first, middle, last | CRM stage timestamps | The weights are assumptions — say so |
| Data-driven / algorithmic MTA | Marginal contribution of each touch | High-volume conversion paths + identity resolution | Starves at mid-market B2B volumes; opaque |
| Marketing mix modelling | Aggregate spend-to-outcome elasticity | 2–3 years of weekly data, clean spend ledger | Channel-level only; slow to react |
| Self-reported attribution | What buyers believe influenced them | One free-text form field + discipline | Recall bias — a signal, not a ledger |
Pragmatic call for multi-touch attribution in B2B 2026: a W-shape model (40/20/40 across first touch, pipeline creation, closed-won) plus a self-reported layer plus a quarterly incrementality test will beat a data-driven model fed on forty conversions a quarter. The expensive model isn't the sophisticated one — it's the honest one your data can actually support.
Step 3 — Instrument the Dark Funnel
Dark funnel attribution is not a product you buy; it's a set of habits that capture influence your analytics can't see:
- UTM everything you control — newsletters, community posts, guest appearances, partner links. Half of dark-funnel blindness is untagged links you own.
- Ask one honest question. A free-text "How did you hear about us?" on every form, reviewed monthly. Buyers name podcasts, communities and AI tools your tracking will never see. Where self-reported and tracked data diverge, treat the divergence as a map of your blind spots — the same validation-layer logic we use in the attribution ROI method above.
- Watch for AI-mediated arrivals. Referrers from assistant domains, untagged traffic spikes after AI-visible mentions, branded-search surges that follow them. None of it proves causation; all of it belongs in the model's evidence file.
- Trend, don't point. Dark-funnel signals are noisy week to week. Monthly direction is the unit of truth.
Step 4 — Fix the Foundation Before the Mathematics
An attribution model's ceiling is your data quality — identity resolution across form fills, de-duplicated accounts, trustworthy stage timestamps. If cost data lives in the ads platform, lead data in the CRM and outcome data nowhere, no model class will save you. Start where business intelligence for mid-market B2B starts: one source of truth, owned, with definitions people actually follow. Then run the data hygiene audit before you attribute anything: you cannot attribute what you never captured cleanly.
Step 5 — Validate With Experiments, Not Faith
Every attribution model is a hypothesis with good PR. The correction is incrementality: geo holdouts, channel pause tests, budget swings you can read two quarters later. Run one clean experiment per quarter, sized to matter, and let the results overrule the model — "model proposes, experiment disposes." A channel the model loves but the holdout can't defend is a channel you've been mismeasuring, and finding that once pays for the whole programme.
Step 6 — Govern It Like a Product
Give the model a named owner, a monthly refresh cadence, a change log (model tweaks get logged, or your trend lines lie), and a quarterly readout tied to real kill/scale decisions — the same governance rhythm as our honest marketing automation ROI method and the measurement framework in this series. Governance is what separates the teams that can prove return from the 59% that can't: Jasper's 2026 survey found only 41% of marketers confidently prove AI ROI, down from 49% the year before. The model that survives budget season is the boring one — documented, versioned, and argued with quarterly.
The Whole Build on One Page
- Write down the decisions the model serves and who consumes it.
- Pick a model class your data volume can honestly feed (W-shape is the mid-market default).
- Instrument the dark funnel: UTMs, one self-reported field, AI-referral watch.
- Fix the data foundation — identity, de-dup, timestamps — before the maths.
- Validate quarterly with incrementality experiments that can overrule the model.
- Govern it: owner, refresh cadence, change log, decision readout.
Where We Come In
Building attribution that survives AI-era buying is mostly plumbing: clean data, honest weights, one experiment a quarter, and a decision log finance recognises. That's the work we do — if you want your model pressure-tested or installed in one structured engagement, book a call. Better to find your 30% yourself than keep guessing which half is wasted.