Building a B2B Attribution Model Your CFO Will Trust

Your attribution report says marketing sourced $4.2M in pipeline last quarter. Finance’s report says marketing sourced $1.6M. Both numbers were pulled from the same CRM on the same day. Neither team can explain the gap in the meeting, so the meeting ends the way these meetings usually end, with the CFO deciding to treat the marketing number as directional and fund accordingly.

The instinct after that meeting is to go buy a better model. Multi-touch, data-driven, algorithmic, something with more mathematics in it than the last one. That instinct is the reason the next meeting goes the same way. A CFO was never objecting to the arithmetic. The objection was to a number that cannot be traced, cannot be reconciled, and changes without explanation.

Attribution is a finance-facing product. Building one that survives review is an engineering problem with a governance layer, and the mathematics is the smallest part of it.

What a CFO actually means by “I don’t trust that number”

A B2B attribution model is a documented set of rules for assigning credit for pipeline and revenue to marketing activity, together with the data definitions, reconciliation method and change controls that make the output auditable. The model is the rules plus the governance. A credit-allocation algorithm on its own is only half of one.

That definition is doing real work, because it names the half that most teams skip. When a CFO says the number is not trusted, the objection is almost never “your decay curve is wrong.” Decoded, it is usually one of four things:

• It does not tie out. Marketing’s revenue number and the general ledger disagree, and no one can produce the bridge between them.
• It moved and nobody can say why. The figure shifted 30% quarter over quarter. If a definition changed, that change was never logged.
• It counts things twice. Sourced and influenced pipeline are added together, or two channels each claim the same deal at 100%.
• Nobody owns it. When finance asks who signed off on the methodology, the answer is a tool vendor.

Every one of those is a process failure rather than a modeling failure. That is good news, because process failures are fixable without a data science team.

The trust gap is measurable, and it starts inside marketing

The uncomfortable finding is that marketing usually does not believe its own numbers either. In Forrester’s Marketing Survey, 2024, 64% of B2B marketing leaders said they do not trust their organization’s marketing measurement for decision-making. The credibility gap opens up long before the number reaches finance.

Finance notices. Gartner’s 2024 survey of 378 senior marketing leaders found that CMOs ranked CFOs at 40% and CEOs at 39% as the executives most skeptical of marketing’s value (Gartner, September 2024). The two offices holding the budget are the two least convinced by how it is being justified.

What makes this solvable is the shape of the gap. A 2026 Bain & Company and Google survey of nearly 1,400 senior marketing and finance executives found that only 41% of marketers feel appropriately equipped with the data, tools and measurement capability to tie performance to business outcomes, while companies with strong marketing and finance relationships were 2.5 times more likely to have credible data. Bain’s read is worth sitting with, because it cuts against the usual framing: marketing and finance largely agree on which outcomes matter. The divide is about proof, not priorities.

Proof is a build problem. The next four sections are the build.

There is a budget consequence to getting this wrong. Gartner’s 2026 CMO Spend Survey found marketing budgets effectively flat at 7.8% of company revenue, barely moved from 7.7% the year before. In a flat-budget environment, the team that can defend its contribution line by line keeps its allocation. The team that cannot gets trimmed at the margin every planning cycle.

Step 1: Match the model to your data, not your ambition

Most attribution advice is written for companies with enough deal volume to support algorithmic models. Most companies reading it are not those companies. Buying a Shapley-value model for a business closing forty deals a quarter produces confident-looking output built on samples too small to mean anything, which is a faster route to losing finance’s trust than using no model at all.

Improvado’s published practitioner guidance for building custom attribution models, which is vendor operational guidance rather than independently published research, puts the working thresholds at roughly 300 or more monthly conversions and 30 or more unique conversion paths for a Markov chain model, and around 500 conversions with 50 paths for Shapley value. Treat those as order-of-magnitude gates, not precise cutoffs.

Run your own volume first, then pick from this table.

Monthly conversions	Model to run	Why it holds up in review
Under 100	Two-touch: first touch and last touch reported side by side, never blended into one figure	Every number traces to a single CRM record. A CFO can audit any row in under a minute.
100 to 300	Rule-based W-shaped: 30% first touch, 30% opportunity creation, 30% closed-won, 10% distributed across the rest	The weights are a stated policy decision, documented and defensible, rather than an output nobody can interrogate.
300 to 500	Markov chain, reported alongside the rule-based model for at least two quarters	The parallel run is what earns the switch. Finance sees both, sees the delta, and approves the change knowingly.
500 and above	Data-driven or Shapley, with a documented fallback model	Volume supports it. The fallback matters, because algorithmic models fail quietly when data quality slips.

Two rules apply at every volume band. Never report sourced and influenced pipeline as one combined figure, because that is the single fastest way to get a number rejected. Always run a new model in parallel with the old one before you switch, because a methodology change that lands as a surprise reads to finance as marketing moving the goalposts.

The honest position for most mid-market teams: a simple model that finance can audit outperforms a sophisticated model that nobody can explain. The organizations that get a sophisticated model approved are usually the ones that spent two quarters proving out a simple one first.

Step 2: Anchor on the four numbers finance already believes

Finance has a mental model of the business and it does not contain MQLs. The fastest way to make attribution legible is to report against numbers finance already tracks, then let marketing’s detail hang underneath.

The number	How to state it	What finance does with it
Customer acquisition cost, fully loaded	Total sales and marketing spend, including salaries and tooling, divided by new customers acquired in the period	Compares against gross margin and payback assumptions in the operating model
CAC payback period	Months of gross profit needed to recover fully loaded CAC	Feeds cash-flow planning directly. This is the number a CFO cares about most.
Marketing-sourced pipeline as a share of the coverage requirement	Sourced pipeline over the pipeline coverage the sales plan requires	Answers whether marketing is delivering the top-of-funnel volume the plan assumed
Marketing-sourced closed-won revenue	Revenue from opportunities where marketing created the first qualified touch, reconciled to bookings	The only marketing number that can be checked against the general ledger

Report those four at the top. Put channel-level detail below them, clearly labeled as diagnostic rather than financial. The separation matters more than it looks: it signals that marketing understands which numbers are audited and which are directional, and that single act of self-classification does more for credibility than any model upgrade.

One definition has to be settled before any of this works. Write down, in one sentence each, what your organization means by marketing-sourced and marketing-influenced, get sales and finance to agree to both sentences, and put them somewhere permanent. Teams that skip this step re-litigate it in every quarterly review.

Step 3: Reconcile to the general ledger, on finance’s calendar

This is the step almost nobody does, and it is the step that converts a marketing report into a finance-grade one.

Reconciliation means producing a documented bridge from marketing’s reported revenue contribution to the revenue finance recognizes. The two will not match. They are not supposed to match. What earns trust is being the team that explains the delta before anyone asks.

A working bridge has five lines:

1. Marketing-reported sourced revenue for the period, straight from the attribution model.
2. Less deals not yet closed-won in the finance system, because attribution frequently books an opportunity that finance has not recognized.
3. Less revenue-recognition timing differences, covering multi-year contracts, ramped deals and anything where bookings and recognized revenue separate.
4. Less deals reclassified after the fact, such as an opportunity re-tagged from marketing-sourced to partner-sourced during deal review.
5. Equals reconciled marketing-sourced revenue, the number that ties to the ledger.

Run this on finance’s close calendar rather than marketing’s reporting calendar. If the finance close is the tenth working day, marketing’s reconciled number is due on the tenth working day. Reporting on a different cadence than finance guarantees the two sets of numbers are never comparable, which is a large part of why they never agree.

Expect the gap to be substantial the first time. Octane11’s first-party analysis across its enterprise client base, which is vendor analysis rather than independent research, puts the gap between marketing’s self-reported influenced pipeline and CRM-verified pipeline at roughly two to four times. A first reconciliation that surfaces a gap of that size is not a failure. Publishing it before finance finds it is the point.

Step 4: Put a governance layer around the model

A model without governance degrades. Definitions drift, someone changes a lifecycle stage, and six months later the number means something different than it did when it was approved.

The artifact that prevents this fits on one page. Build it once and keep it in a shared location finance can reach without asking.

Attribution model charter

• Model in use and the date it was approved.
• Data definitions: marketing-sourced, marketing-influenced, qualified touch, attribution window, and the closed-won stage that counts.
• Named owner: one person, usually in RevOps, accountable for the model’s integrity. Not a team, not a vendor.
• Approvers: marketing, sales and finance, each named.
• Reporting cadence and its alignment to the finance close.
• Change log: every methodology change, dated, with the reason and the approver. This single field resolves most “why did the number move” disputes before they start.
• Known limitations: what the model cannot see, written by marketing rather than discovered by finance.
• Review date: quarterly.

That last-but-one item is counterintuitive and it is the highest-leverage line on the page. Documenting your own blind spots, including anonymous research, peer influence and everything covered in our earlier piece on the pipeline last-touch attribution never sees, converts a vulnerability into evidence of rigor. A CFO who finds an unstated limitation discounts the whole model. A CFO handed the limitations up front reads the rest as credible.

Separate the two approvals. Sign-off on model design happens before the build, in a working session with finance where weights and definitions get agreed. Sign-off on reported numbers happens every period, as a routine review. Teams that combine these into a single meeting end up defending methodology and results at the same time, which is the hardest possible version of the conversation.

The five objections a CFO will raise, and what answers them

Rehearse these. Each one has a real answer that does not require conceding the model.

The objection	What answers it
“Correlation is not causation. You cannot prove marketing caused this.”	Correct, and attribution does not claim to. Attribution describes contribution. For causal claims, run a holdout or geo-lift test on one channel and report the incremental result separately. Naming the limit of the method is what makes the rest of it credible.
“Why is this number different from last quarter’s?”	Open the change log. Either nothing changed and the business moved, or a definition changed and here is the date, reason and approver. This is the objection the change log exists to end.
“Your pipeline number does not match the ledger.”	Hand over the five-line reconciliation bridge. It is not supposed to match. Here is every line of the difference.
“This model is a black box.”	For rule-based models, show the weights and the policy reasoning behind them. For algorithmic models, provide the parallel rule-based figure alongside it. A model that cannot be sanity-checked against a simple alternative should not be the reporting model.
“Marketing takes credit for deals sales closed.”	Report sourced and influenced separately, always. Sourced means marketing created the first qualified touch. Influenced means marketing touched an opportunity that existed. Both are real, and combining them is what triggers the objection.

A 90-day build sequence

• Days 1 to 15, definitions. Write the sourced and influenced definitions. Get written agreement from sales and finance. Nothing else starts until this is done.
• Days 16 to 30, data audit. Count actual monthly conversions and unique paths. Fix lifecycle stage integrity and UTM discipline. Pick the model band from the Step 1 table.
• Days 31 to 50, build and parallel run. Stand up the chosen model. Run the existing model alongside it. Log every discrepancy.
• Days 51 to 70, first reconciliation. Produce the five-line bridge for the most recent closed period. Walk finance through it before presenting any results.
• Days 71 to 90, charter and sign-off. Complete the one-page charter. Hold the model design approval session. Set the reporting cadence to the finance close and schedule the quarterly review.

Ninety days is realistic when definitions are the first thing settled. Teams that start at the tooling step usually spend the ninety days there and arrive at the same meeting with a better dashboard and the same credibility problem.

Measurement holds when it is run as a system

Buying groups now run from five to sixteen people across as many as four functions (Gartner, May 2025, based on a survey of 632 B2B buyers). No single-model dashboard is going to describe a purchase of that shape cleanly, and pretending otherwise is what erodes trust in the first place.

What holds up is a documented model matched to real data volume, reconciled to the ledger on finance’s calendar, governed by a charter with a named owner and an honest limitations section. That is a durable operating routine, not a report setting.

At Tru Performance, a business growth and operations partner working across 250+ brands and more than $500M in client revenue, measurement of this kind runs inside ConvergeOS™, our operating model that moves work through five stages: Diagnose, Design, Deploy, Operate and Compound. Attribution sits mostly in Operate, because a model is only worth what its monthly discipline is worth. The build is the easy part. The reconciliation that happens every close is what makes finance stop discounting the number.