Ask a marketing team how they know their spend is working and you’ll usually get a dashboard. Last-click here, a multi-touch model there, a tidy chart that assigns credit to every channel down to two decimal places. It looks like science. It is mostly theatre. The uncomfortable truth, increasingly accepted by serious operators, is that the majority of the modern buyer journey happens in places your analytics tools will never see — and the dashboards that claim otherwise are confidently wrong.
This isn’t a counsel of despair. It’s an invitation to grow up. Measurement still works; it just doesn’t live in the attribution panel anymore. It lives in incrementality tests, in a revived discipline called marketing mix modeling, and in the unglamorous act of simply asking buyers how they found you. Here’s how the old model broke, why the metrics most teams still trust mislead them, and what a measurement system that survives contact with reality actually looks like.
How attribution quietly broke
Attribution didn’t collapse in a single dramatic event. It eroded from three directions at once.
First, the technical floor gave way. Third-party cookie deprecation, Apple’s privacy changes, server-side blocking, and a general tightening of consent regimes have steadily blinded the tracking that user-level attribution depends on. The pixel that once stitched together a person’s path across sites now drops touchpoints, double-counts others, and quietly models the gaps. Every “data-driven” attribution model today is, to a meaningful degree, an educated guess dressed as a measurement.
Second — and more importantly — buyer behaviour moved somewhere the pixel was never invited. The so-called dark funnel is where decisions actually form: peer conversations in Slack and WhatsApp groups, review sites like G2 and Capterra, LinkedIn comment threads, podcasts, communities, in-person events, and now AI search. When a prospect asks ChatGPT or Perplexity to compare vendors, reads the synthesized answer, and arrives at your site already half-sold, none of that influence appears in your analytics. They land as “direct” or “organic,” and your last-click model hands the credit to whatever happened to be the final clickable step. The Smarketers estimates that an enormous share of the B2B buyer journey — on the order of 70 to 80 percent — now happens in these untracked dark-funnel channels. Whether or not that exact figure holds, the direction is undeniable: the part of the journey you can see is the small, late, lowest-value tail of it.
Third, the conceptual model itself was always partly fiction. Multi-touch attribution assumes you can observe every meaningful interaction and assign it a fractional weight. But if most influence is invisible, the maths is being run on a fraction of the inputs. A 40/30/30 split across three tracked channels tells you nothing about the eleven untracked conversations that did the real persuading. Multi-touch attribution adoption has climbed — Improvado, citing Digital Applied, puts it near 47 percent of marketing teams in 2026, up from 31 percent in 2023 — yet most teams still fall back on last- or first-click in practice. That gap between sophisticated tooling and primitive habit is the whole problem in miniature: we bought better instruments to measure the wrong thing more precisely.
The metrics that mislead
If the system is broken, the individual metrics it produces deserve specific scrutiny. Three in particular cause the most damage because they look like progress.
Last-click. Still the default in countless reports, last-click attribution rewards the channel closest to the conversion — usually branded search or direct traffic. This systematically overcredits the bottom of the funnel and starves everything that created demand in the first place. Brand search converts well precisely because something earlier made the buyer search for your brand. Last-click sees the harvest and ignores the planting, which is how teams end up cutting the very activities that fill their pipeline.
Vanity reach. Impressions, reach, follower counts, and “engagement” feel reassuring and correlate with almost nothing that matters. Reach without a mechanism for influence is just noise you paid for. A post that reaches a million indifferent people is worth less than one that reaches a thousand in-market buyers and changes their thinking — but the dashboard makes the million look like the win.
MQL theatre. The marketing-qualified lead remains one of the most gamed metrics in B2B. Teams optimise toward an MQL target, lower the threshold to hit it, and celebrate volume while sales quietly ignores most of the handoffs. The MQL count goes up and to the right; revenue doesn’t move. It’s a metric designed to make a department look busy rather than to describe whether the business grew.
A measurement model that survives reality
The alternative isn’t more dashboards. It’s a smaller set of methods that tolerate invisibility instead of pretending it away. Three pillars matter.
Incrementality testing and holdouts. The only honest question in marketing is: what happened because of this spend that wouldn’t have happened anyway? Incrementality answers it directly. You hold out a region, an audience, or a time period from a campaign and compare outcomes against the exposed group. If conversions are identical with and without the spend, that spend was harvesting demand you’d have captured for free. Geo holdouts, paused-channel tests, and randomised audience splits are cheap relative to what they reveal, and they cut straight through attribution arguments. You stop debating which model assigns credit and start measuring lift you can actually bank.
Marketing mix modeling revival. MMM is old — older than the cookie it now outlives — and it’s having a deserved renaissance precisely because it never relied on user-level tracking. By analysing aggregate spend and outcomes over time against external factors like seasonality and promotions, MMM estimates each channel’s contribution without following a single individual. Privacy changes don’t dent it. The dark funnel can be partly inferred from it. Once available only to enterprises with statisticians, MMM is now accessible through open-source libraries and lighter tooling, putting it within reach of mid-market teams willing to invest in clean spend and revenue data.
Self-reported attribution. The lowest-tech method is often the most revealing. Add a “How did you hear about us?” field to demo requests and high-intent forms. It’s messy, it’s qualitative, and it captures exactly what your pixels cannot: the podcast, the peer recommendation, the AI search answer, the conference hallway. Used alongside incrementality and MMM, self-reported data triangulates the dark funnel — not with false precision, but with directional truth you can act on.
Putting it to work
A measurement system is only useful if a team can run it without a data-science department. Here’s a practical rhythm.
A weekly read. Once a week, look at the few things that resist manipulation: total pipeline created and revenue, blended efficiency (all spend against all new revenue, no channel credit games), the trend in self-reported attribution sources, and any live holdout results. The question is always directional — is the machine producing more efficient revenue than last month — not “which channel deserves credit for this deal.”
What to ignore. Stop reporting last-click channel splits as truth. Demote impressions and reach to operational diagnostics, never headline KPIs. Treat raw MQL volume with suspicion unless it’s tied to downstream conversion quality. If a metric can be improved without the business improving, it doesn’t belong in the leadership deck.
- Ignore decimal-point attribution percentages — they imply a precision the data can’t support.
- Ignore week-to-week reach fluctuations; they’re noise.
- Ignore any “ROI” figure built purely on last-click conversions.
Where AI helps versus where it just speeds up bad measurement. AI genuinely helps with the heavy lifting of real measurement: building and maintaining MMM models, cleaning and reconciling messy spend data, clustering free-text self-reported answers into usable categories, and flagging anomalies worth a closer look. Where it does harm is when it’s pointed at the broken model — generating ever-richer last-click dashboards, fabricating plausible attribution paths from incomplete data, and producing confident charts faster than anyone can question them. The test is simple: if AI is helping you measure incremental impact, it’s an asset. If it’s helping you produce more attractive attribution theatre, it’s just accelerating the lie.
None of this is a fashionable answer. It’s harder than reading a dashboard and accepting the number it shows. But the dashboard moved into a building it can no longer see inside. The teams that win the next few years won’t be the ones with the prettiest attribution model — they’ll be the ones honest enough to measure what they can prove, comfortable enough with uncertainty to test their way to truth, and disciplined enough to ignore the metrics designed to make them feel good.
