A multi touch attribution model does not discover anything. It applies a distribution rule that a human selected to a set of touchpoints that a tracking system happened to capture. Change the rule and the answer changes. Change the tracking coverage and the answer changes again.
That is not a criticism of the tooling. It is a description of what the tooling does. The damage comes from the sentence that follows the dashboard: this channel drove 34 percent of revenue.
Three things attribution structurally cannot see
- Anything without a click. The podcast, the conference conversation, the Slack community recommendation, the answer an assistant generated. All real, all invisible, all reallocated to whichever channel happened to catch the click.
- The counterfactual. Attribution tells you which touchpoints preceded a purchase. It cannot tell you whether the purchase would have happened anyway, which is the only question a budget decision actually depends on.
- The long tail of consideration. B2B cycles routinely exceed the cookie and session windows the model relies on, so early influence is systematically undercounted and closing activity is systematically overcredited.
The third one is getting worse, not better. Win rates have fallen and committees have grown, which means more people touch a deal over a longer window before anyone signs. The share of that window your tracking can observe is shrinking while the reporting stays just as confident.
Attribution answers what happened before the purchase. Budgeting requires knowing what would not have happened without the spend. These are different questions.
The same deal, three rules, three different answers
This is the demonstration I run when a team tells me their model is accurate. Take one closed deal with a normal touch history and apply three standard rules to it. Nothing about the deal changes. Only the arithmetic does.
Worked example
One 60,000 dollar deal, six tracked touches
A mid market deal closes after six tracked touches over five months: a podcast mention that drove a branded search, an organic guide visit, a paid social click, a webinar registration, a second organic visit, and a direct visit to the pricing page. Credit is allocated three ways.
- First touch
- Organic search takes 60,000. Paid social takes nothing.
- Last non-direct
- Organic search takes 60,000 again, for a different reason.
- Linear, six touches
- Paid social takes 10,000. Organic takes 20,000 across two visits.
- Untracked in all three
- The podcast that caused the first branded search takes nothing.
Result: Paid social's contribution to this deal is 0 dollars or 10,000 dollars depending on a dropdown setting. If a channel's budget survives on that difference, the budget decision is being made by the dropdown.
Now scale that to a quarter of deals. The spread between models on a single channel is routinely two to three times. That is wider than most of the performance differences teams are arguing about in the meeting, which means the argument is usually about the rule rather than the market.
What attribution is genuinely good for
I still build attribution reporting for every client, because used as a management instrument rather than as evidence it does real work.
- 1
Detecting change
The absolute value is arbitrary. The direction is not. If a channel's attributed contribution halves month on month under an unchanged rule, something moved and it is worth investigating.
- 2
Settling internal disputes cheaply
A shared, stated convention stops the weekly argument about whose channel deserves credit. That is worth a lot of organisational time even when the numbers are approximate.
- 3
Finding the pages that appear in winning journeys
Not as proof of causation, but as a shortlist. Pages that recur across closed won journeys are usually worth investing in, and this is a cheap way to find them.
- 4
Catching tracking breakage
A channel that goes to zero overnight is almost never a market event. It is a broken tag, a consent banner change, or a redirect that dropped parameters. Attribution is a good smoke alarm even when it is a poor thermometer.
2 to 3x
Typical spread in one channel's credit across standard models
1 field
Cost of adding self reported source to a demo form
4 to 6 weeks
Holdout window that produces a readable signal in B2B
The three inputs I trust more
| Method | What it tells you | Cost |
|---|---|---|
| Self reported attribution | What the buyer believes influenced them, in their words | One form field |
| Geographic or timing holdouts | The counterfactual, approximately, for one channel | A few weeks of deliberate under-spend |
| Branded demand trend | Whether total market pull is growing, regardless of channel credit | Free, already in your search console |
How to run a holdout without frightening the board
Most teams never run one because pausing spend sounds like sabotage. It is not, if it is scoped. You are not turning a channel off. You are turning it off in two places for six weeks and reading the difference.
A six week regional holdout
Pick the pair
Two comparable regions with similar historical pipeline per week. Not your largest market.
Freeze everything else
No new creative, no landing page changes, no outbound surge in either region during the window.
Cut one, hold one
Spend goes to zero in region A. Region B keeps its normal budget as the control.
Read pipeline created
Compare pipeline created, not closed revenue. Closing lags too far behind to read in six weeks.
Restore and decide
Restore spend, then size the gap against the money saved. That ratio is your incrementality estimate.
Worked example
What a holdout actually told us
A paid social line item was credited with roughly a third of attributed pipeline under a last non-direct rule. Two comparable regions were paired. Spend went to zero in one for six weeks while the other continued at its normal weekly level in the low five figures.
- Attributed share before the test
- About 33 percent of pipeline
- Pipeline created, control region
- Flat against its trailing eight week average
- Pipeline created, held out region
- Down by roughly a fifth
- Implied incremental share
- Nearer 20 percent than 33 percent
Result: The channel was real and it was overcredited by around a third. We kept it, cut the budget by the difference, and moved that money into the creative testing budget. No model would have produced that decision.
The strongest case against this
Holdouts cost real pipeline, and self reported data is famously biased. Attribution is at least consistent, applied to everything, and available every morning without anyone running an experiment.
That is fair, and it is why I keep the dashboard. Consistency is genuinely valuable for reading change. What consistency cannot do is turn a rule into a finding. A holdout costs a few weeks of spend in two regions, once a quarter. Getting a channel's budget wrong by a third for a year costs more than that in the first month. Run the cheap experiment on the biggest line item and keep the consistent dashboard for everything else.
The untracked layer is growing, not shrinking
Ten years ago the invisible part of the journey was word of mouth and events. It still is. What has been added is a layer that answers buyer questions before the buyer ever reaches a site: assistant answers, summaries at the top of a results page, and community threads that get read far more often than they get clicked.
A buyer who reads a comparison inside an assistant, forms a shortlist, and then types your brand name into a search bar arrives as branded organic. Every model in common use will file that deal under organic search. The work that created the shortlist is invisible, unpaid, and unbudgeted, which is precisely how it stays underfunded.
| Where the influence happens | What the model records | What gets underfunded |
|---|---|---|
| Assistant answer or AI summary | Branded organic, or direct | The content that made you quotable |
| Peer recommendation in a private community | Direct, or a branded search | Community presence and customer advocacy |
| Podcast or conference conversation | Whatever caught the next click | Founder and operator visibility |
| Sales follow up on an old inbound lead | The original first touch, months earlier | Nurture and re-engagement |
I am not arguing you should try to track these. Most of them cannot be tracked and the attempt burns quarters. I am arguing you should stop letting a report that cannot see them decide which of them gets funded.
When sales and marketing claim the same deal
This is the argument attribution was invented to end, and it is the one place the model earns its cost. An outbound sequence and an inbound guide visit both preceded a deal. Sales says the sequence created it. Marketing says the buyer was already reading. Both are partly right and the argument cannot be resolved with the data either side is holding.
- 1
Agree the rule before the quarter, not after the deal
Any rule is acceptable if it is agreed in advance. A rule chosen after the numbers are visible is a negotiation, and everyone in the room knows it.
- 2
Report sourced and influenced separately
One number for who opened it, one for who touched it. Two honest numbers beat one contested number, and nobody has to lose to make the report add up.
- 3
Let the buyer break the tie
The self reported field settles more of these arguments than the model does. When a buyer writes that they had been reading for months before the call, that is the end of the debate.
The failure mode to avoid is a compensation plan wired directly to a modelled number. The moment a rule decides someone's bonus, the rule gets gamed, and the reporting stops describing the market at all. Pay on pipeline and revenue. Report on credit.
How to talk about it without losing credibility
Marketers lose the room when they present a modelled number as a measured one and then get asked a question the model cannot survive. The fix is to change the sentence, not the dashboard.
Says more than the data supports
- Paid social drove 1.2 million in revenue
- Content has a 4x return
- This channel is our most efficient
Defensible, and still useful
- Under our last non-direct rule, paid social touched 1.2 million in closed revenue
- Content appears in 60 percent of closed won journeys, and buyers name it unprompted
- When we paused this channel in two regions, pipeline fell measurably in both
Verdict: The right column survives a CFO's follow up question. The left column is why marketing budgets get cut in the first bad quarter.
Honesty about the limits is not a weaker position. A team that says 'we do not know precisely, and here is the experiment that will tell us' is far more trusted over time than a team with a confident chart and no counterfactual.
The measurement stack I would actually run
- ✓One stated attribution rule, written down, unchanged for at least four quarters.
- ✓A free text self reported source field on every high intent form, read monthly by a human.
- ✓One holdout per quarter on the largest line item, scoped to two regions.
- ✓Branded search volume tracked as the demand trend line, separate from any channel report.
- ✓Pipeline created as the primary read on any test, with closed revenue reviewed a quarter later.
- ✓A quarterly note recording which decisions the measurement actually changed. If none, simplify it.
What I would do Monday
- 1Write down which attribution rule you use and who chose it. If nobody remembers, that is your finding.
- 2Add a self reported source field to every high intent form.
- 3Run one geographic or timing holdout on your largest channel this quarter.
- 4Stop presenting attributed revenue as a fact and start presenting it as a convention with a stated rule.
Common questions
- Is B2B attribution accurate?
- Not in the sense the word implies. An attribution model reports which tracked touchpoints preceded a purchase and splits credit using a rule a person chose. It cannot observe untracked influence and it cannot observe what would have happened without the spend. It is accurate about touch order and nothing else.
- What is the difference between attribution and incrementality?
- Attribution divides credit for deals that already closed. Incrementality asks how many of those deals would not have closed without a specific spend. Attribution needs a rule. Incrementality needs a control group, usually a region or a time window where you deliberately stop spending. Only the second answers a budget question.
- Which attribution model should a B2B team use?
- Pick one, write it down, and keep it. Last non-direct is the easiest to explain and the hardest to game in a meeting. The model matters far less than the consistency, because you are reading the change over time, not the absolute value on any single month.
- Is self reported attribution reliable?
- It is biased and still worth more than most dashboards. Buyers under-report ads and over-report the last memorable thing. What it catches that nothing else does is the untracked layer: communities, podcasts, peer recommendations, assistant answers. Read it as a list of what exists, not as a share of revenue.
- How often should we run a holdout test?
- One per quarter on the largest line item is enough for most teams under fifty people. Any more and you are spending more attention on measurement than on the work. Hold the channel out in two comparable regions for four to six weeks, then compare pipeline created against the regions that kept running.
Who wrote this
Avishai Sam Bitton
Founder, DemandBox
Avishai runs demand generation programs for B2B SaaS companies across performance marketing, SEO, and answer engine optimization. He works directly with the teams he advises, with no account managers in between.
Connect on LinkedInThe long version
How to Build a Pipeline Model Your CFO Believes
Every formula for a pipeline model finance will sign off on: worked models at three ACV bands, 2026 win rate and CAC payback benchmarks, and stress tests.
Want this argument applied to your numbers?
No deck, no discovery sequence. Tell us what you are spending and where the pipeline stalls, and we will tell you what we would change first.
Talk to an Expert