B2B marketing attribution: what I trust and what I don’t
Attribution assigns credit according to rules you choose. It does not prove that one activity caused a result. Most arguments I’ve seen between marketing and finance start when attribution is presented as proof of causation.
I rebuilt attribution at Penfold because the existing model could not answer the question that mattered: where should we spend the next pound? We were spending heavily on B2C paid acquisition in a category where trust takes time to build and customer value accrues over years. The reporting showed channel performance, but it could not reliably connect spend to downstream customer value.
My working threshold is simple: I want the model and the people closest to the revenue to tell the same story. If they do, I’ll take the number to the board, but label it as directional rather than causal. If nobody recognises the pattern, I won’t make an expensive decision from the model alone.
What attribution actually does
An attribution model distributes credit across the touchpoints it can track, according to rules you choose. First-touch attribution gives all the credit to the first interaction; last-touch gives it to the last. Linear models split credit evenly. Time-decay models weight recent touches more heavily. Position-based models favour the first and last touchpoints.
These are choices, not discoveries. Change the model and you change the answer, even when nothing in the business has changed. That is why two teams can look at the same quarter and reach opposite conclusions about channel performance.
GA4 is a useful example of why model labels should not be treated as permanent truth. Google removed first-click, linear, time-decay and position-based models from GA4 in November 2023. Its current reporting choices centre on data-driven and paid-and-organic last-click attribution. The menu changed. The customer journey did not.
This matters in B2B because buying decisions often involve several people, sales cycles are long and much of the influence is invisible to tracking. A prospect reads a post, hears your name from a peer, sees a competitor comparison, discusses it internally for six weeks and eventually arrives through branded search. Branded search gets the credit, despite having done little of the work.
The dark middle
Tracking captures some visible points in a B2B journey: an ad click, a form fill or a booked call. Much of the influence between them is dark: Slack communities, WhatsApp groups, podcasts, conversations at events and internal advocacy from a champion you have never spoken to.
Attribution models cannot see those interactions. They assign credit only to visible touchpoints, creating a confident-looking picture from partial evidence. The picture is not necessarily wrong. It is incomplete, and the model cannot tell you how incomplete.
That is why I use self-reported attribution alongside tracked data. Ask one open question on your contact form: “How did you hear about us?” The answers are messy, but they often reveal channels the dashboard misses.
Where the people come in
A cheap, useful check is to ask the people closest to the revenue whether they recognise what the model reports.
If the model attributes a third of pipeline to one channel, the sales team should be able to describe those conversations without prompting: what prospects ask, which objections recur and how well those opportunities convert. When the model and the sales team agree, I have more confidence in the direction of the finding, not the precision of the number. It is corroboration, not proof.
When they disagree, investigate the gap. The model may be over-crediting a touchpoint, or a real source of demand may be going unrecorded. You will not find either by reading the dashboard alone.
That is roughly how we challenged a targeting assumption at Penfold. The data pointed towards employers. Sales conversations showed a more specific problem: HR often could not move the process forward alone. By the time that became clear, I had already doubled down on HR as the buying centre, which cost us time.
The correction was to test founders and owners directly at smaller firms, where the person who cared was more likely to be able to sign.
The claim you cannot make
Attribution tells you which tracked touchpoints were present and how your model allocated credit. It does not tell you which were necessary to produce the result. Confusing those claims is a common source of overreach in marketing reporting.
To claim causation, you need a test designed to support it. Examples include geo holdouts, where you suppress a channel in matched regions; audience holdouts, where you withhold spend from a comparable group; and matched-market tests, where similar markets receive different treatment.
A clean before-and-after comparison can still be useful, but it is weaker evidence because other factors rarely stay still.
These tests are harder to run and slower to produce answers. They also give you evidence you can defend when a CFO asks whether the revenue would have arrived anyway. If you have not run one, be clear that attribution supports a directional judgement, not a causal claim.
A model you can defend
Agree the rules before you look at the output: which touchpoints count, the lookback window, how you handle self-reported data and what happens when the sources disagree.
Do this with sales and finance while the method is still abstract. Once a number is attached to somebody’s target, the conversation stops being only about method.
Use one model as the default rather than switching to whichever one flatters the quarter. Report on the same basis each month. A consistent, imperfect model is usually more useful for comparing trends than a model that changes each quarter. If you change it, record why and restate earlier periods if you can.
Keep a short list of what the model cannot see: partner introductions, community, word of mouth and the sales team’s own network. Name these gaps in reporting rather than letting that influence be absorbed by the last trackable channel.
What to report to a board
Boards do not need a technical walk-through of the attribution model. They need your best defensible view of marketing’s revenue contribution, whether efficiency is improving or deteriorating and what you plan to change.
Show the trend, not just the snapshot. Label what is measured, what is modelled and what should be read as directional.
Be direct about weak measurement. If a director identifies the limitation before you do, every other number becomes harder to trust. My most credible board reports have stated what we could not yet measure, why it mattered and what we were doing about it.
If attribution is producing numbers nobody in the business believes, start with the rules and the reality check, not a new tool. Confirm that sales and finance agreed the method, then compare the output with what sales actually sees. Do that before replacing the stack.
If your attribution cannot support a budget decision, let’s talk.
