How do you measure marketing when you only close 30 deals a year?

A CFO asks a reasonable question. Of the four channels we fund, which one is actually producing customers? The honest answer is that his data will support a conclusion in about six years — not because his tracking is badly configured, but because of how many deals his company closes.

Almost everything written about marketing measurement is written for businesses that close thousands of transactions a month. Applied to a company closing thirty deals a year at six figures each, that advice does not merely underperform. It cannot function. The difference is worth stating precisely, because the precision is what makes the decision obvious.

The arithmetic

Take a company closing 30 deals a year across four funded channels. Suppose one channel converts leads to customers 50% better than another — 4.5% against 3%. That is an enormous difference. If you knew it were real, you would move budget the same afternoon.

To establish that difference at conventional confidence — 95% certainty, 80% statistical power — you need to observe 189 closed deals. At 30 deals a year, that is 6.3 years. At 20 deals a year, 9.4 years. Reduce the effect you are hunting to a 20% difference, still large enough to matter to a budget, and the requirement rises to 918 closed deals. Thirty years.

Two consequences follow, and the second ends the argument.

First, the numbers you already have carry no information. Of 30 deals, suppose 8 are credited to one channel. That looks like a 26.7% share. The 95% confidence interval around it runs from 14.2% to 44.4%. You cannot distinguish between "this channel produced a seventh of the business" and "this channel produced nearly half." Both are consistent with the same data. Suppose two channels show 11 deals and 7 — one appearing 57% stronger. Fisher's exact test returns p = 0.399. That is not a weak signal. It is the result you would expect roughly two times in five from two channels performing identically.

Second, the answer arrives after the question has expired. Six years of data describes a market, an offer, a price, a competitive set and a channel mix that no longer exist. Your positioning changed. Two of the four channels were replaced. Buyer research behaviour was rewritten by a technology that did not exist when the measurement started.

Three structural breaks, briefly

The volume problem is the fatal one. Three others compound it, and they are worth knowing because every attribution vendor already concedes them.

The measurement window closes before the deal opens

Meta's maximum click-attribution window is seven days. LinkedIn defaults to thirty. Google Ads reaches ninety. Google Analytics 4 caps acquisition events at thirty days and everything else at ninety. Against that, the median B2B sales cycle runs 84 days, 6Sense puts the average buying cycle at roughly ten months, and deals above €100,000 routinely run 90 to 180 days or longer.

A Meta click is eligible for credit across about 2% of a ten-month deal. The first touch on such a deal falls outside every attribution window that exists on any platform. Not measured imprecisely — not eligible to be measured at all.

Attribution tracks devices; decisions are made by committees

Gartner puts a typical buying group for a complex solution at six to ten decision-makers, each arriving with four or five pieces of independently gathered research. Forrester's 2024 figure is higher: thirteen stakeholders, with 89% of decisions crossing multiple departments.

The number that should stop a CFO is this: Gartner finds buyers spend around 17% of their total purchasing time meeting vendors, and that fraction is divided across every vendor under consideration. Roughly four percent of the decision process happens where you can observe it. The tracked click belongs to whichever committee member happened to click. It is not the decision. It is a fragment of one person's contribution to a decision taken in a room you were not in — the same private, invisible cut that decides which suppliers reach the shortlist at all.

The statistical models say so themselves

Marketing mix modelling is the standard recommendation when click attribution fails. Google's own Meridian documentation works through a hypothetical: two years of weekly data gives 104 observations, which against a realistic channel and control specification yields about four data points per estimated parameter. Google's stated conclusion is that this is too low to estimate the model reliably. Meta's Robyn sets the same floor — two years of weekly data minimum, and roughly ten observations per variable.

A mid-market firm running three channels with stable allocation does not clear those thresholds. Both vendors publish the criteria that disqualify their own tools for your business.

The European aggravation

If your buyers are in Europe, the ground is softer still. Average opt-in for marketing cookies across the EU sits near 46%, down from 54% in 2023, with Germany lowest at around 36%. Northern European markets skew lower, not higher, because privacy awareness is greater. Safari's tracking prevention caps the relevant cookies at seven days. Ad blockers remove roughly a third of users from view globally.

These do not combine into a single tidy loss figure, and anyone who hands you one has invented it. Separately, each removes a different slice of the population from measurement — and the slices are not random. Buyers who refuse tracking are not a representative sample of your market. They are systematically the more senior, more privacy-conscious and more technically literate ones, which in most mid-market B2B is a description of the buying committee.

What the industry recommends instead

The low-volume problem is not a secret. It is documented in the attribution literature: a data-driven model applied to forty closed deals a quarter produces statistically unreliable weights, most high-value low-volume B2B pipelines never reach the required threshold, and running the model below that threshold produces specific-looking percentages that reflect noise rather than signal.

That is an accurate diagnosis. Now look at the standard remedy — use a position-based model instead, U-shaped or W-shaped, because those apply a fixed formula and carry no volume requirement.

Read that carefully. The recommendation is to adopt a model that does not check whether it has enough data, on the grounds that it will therefore never report that it does not. That is not a solution. It is a decision to stop asking. A W-shaped model assigns 30% of credit to the first touch because someone chose the number 30, not because anything in your business produced it. The output is a percentage with a decimal point and no evidentiary content whatsoever — and everyone in the room treats it as a finding.

This is the mechanism by which a company arrives at a confident, precise and entirely fictional account of what its marketing is doing, then allocates real money against it. It is the subject of Reports, Not Revenue, and it is the most expensive habit in mid-market marketing.

The two obvious objections

"Run incrementality tests. Geo holdouts." Correct in principle, and they fail on the same arithmetic. A holdout test is a two-proportion comparison. It needs the same 189 outcomes to detect a 50% effect. Splitting 30 annual deals into test and control halves does not create statistical power; it destroys what little existed. Incrementality testing is an excellent method for businesses with the volume to run it — which is precisely the set of businesses that did not need this article.

"Just ask buyers how they found you." Yes. And this deserves to be said plainly rather than smuggled in, because it is the foundation of how I work. Self-reported attribution is what the measurement literature falls back on when volume runs out. My method is the same instrument applied at a different moment. Instead of asking closed customers how they found us — a sample of thirty, biased toward the buyers who already said yes, recalling events from nine months ago — I ask cold prospects who have never heard of the company whether the offer is worth their time, before any budget is committed.

Same instrument. Better sample, because it includes the people who would have said no. Better timing, because it arrives before the money moves rather than after. If someone tells you this is just a survey, they are half right. The half they are missing is when it runs.

The inversion

Here is what changes once the arithmetic is on the table.

Thirty closed deals cannot separate a strong channel from a weak one at any confidence worth acting on. Thirty structured conversations with cold prospects will tell you, unambiguously and within a fortnight, whether your offer is comprehensible, whether the problem you claim to solve is one they recognise, and whether your price sits inside or outside the range they consider reasonable.

Same number. Two entirely different epistemic situations, because one method requires a large sample to produce signal and the other does not. The choice was never between measuring well and measuring badly. It is between a quantitative method that cannot reach a conclusion inside the lifetime of your current strategy, and a qualitative method that reaches one before you spend. The industry sells the first because it is billable monthly and produces a deliverable. The second produces a decision and then stops.

What this costs if you do nothing

The company that keeps waiting for measurement to resolve does not stand still. It spends. Twelve to thirty months of budget goes out against a channel mix nobody can defend, justified by percentages a statistician would not accept and a CFO cannot audit. At the end of it the honest position is identical to the position at the start — you still do not know which channel worked — except the money is gone and the market has moved. This is the same compounding cost I set out in what declining discoverability costs: the loss is invisible while it accrues and undeniable afterwards.

The dashboard was never the problem. The dashboard was a way of not making the decision.