What Attribution Should Tell You When Nothing Adds Up

Your reported ROAS does not match your revenue because the decision is no longer happening anywhere a pixel can see it. The fix is not better tracking. It is a measurement stack with three layers: ask buyers directly at the form, prove lift with holdout tests, and reconcile everything against closed revenue in the CRM, with platform data demoted to a creative and bidding diagnostic.

Marketing leaders still arguing last-click versus multi-touch are optimizing a map while the territory moves. Both models assume the journey is visible. It is not.

Why doesn't my reported ROAS match the revenue in my bank account?

Platform-reported conversions describe the last visible touch. That is all they have ever described. What changed is how much of the journey is now invisible, which turned a tolerable distortion into a structural one.

Three forces did it. Privacy limits cut the signal platforms can collect. Dark social moved conversations into closed rooms. And AI assistants now build the shortlist before the buyer ever types your brand name into a search bar.

The direct-traffic bucket is the tell. Messaging apps, private communities and in-app profile clicks routinely strip referral data, so the visit lands in "direct" with no source attached. None of that is a tracking bug you can fix with another UTM convention. It is the architecture of how people now share things.

Where the decision actually happens now

Picture the real path. A VP hears your category discussed on a podcast. Six weeks later she asks ChatGPT which vendors handle her specific problem, gets four names, and yours is either in that list or it is not. She posts the shortlist in a private Slack community and two peers reply with experience. She searches your brand name, clicks a branded ad, and fills out a form.

Your dashboard records one thing: branded search converted.

Everything that actually decided the outcome, the podcast, the assistant, the peer replies, left no trace. This is why software-only attribution systematically over-credits demand capture and under-credits demand creation. The channels that create future demand look worthless in-platform precisely because they work early.

The assistant layer is the newest and least understood part of this. The overlap between what ranks first on Google and what AI engines actually cite has collapsed from roughly 70% to as low as 17%. These are different audiences, different signals and a completely different optimization discipline. If you are invisible in ChatGPT, Perplexity, Gemini and Claude, you are losing buyers at the shortlist stage and your analytics will never show you the loss. It shows up as fewer branded searches, and you will blame the wrong channel.

What platform data is still good for

Keep it. Use it correctly. Platform numbers are excellent at the question they are built for: which creative is working, which audience is cheaper, where cost per click moved this week. That is a diagnostic. It is not a verdict on whether a channel deserves budget.

Branded search is the clearest case. It usually posts the weakest incremental return of any line in the plan, because it largely cannibalizes clicks that were already coming. The direction matters: the channel your dashboard crowns is often the one adding least, and the upper-funnel line you are about to cut may be the reason pipeline exists at all.

So the rule is simple. Platform data decides creative and bids. Incrementality decides budget.

The three-layer replacement

Layer one: ask at the form

Put a "how did you hear about us" question on every form and every post-purchase flow. It is the only instrument that can see a podcast, a peer referral or an AI assistant.

Design it carefully, because sloppy design produces confident garbage. Dropdowns pull answers toward the first option, so randomize option order, keep the list short, and include an open text field. Answers also carry recency bias, and placement in the flow shifts them. Pick one placement and hold it so your trend line is comparable.

Layer two: hold out and measure lift

Run geo or audience holdouts on your two largest spend lines. This is no longer an enterprise luxury. Meta open-sourced GeoLift, which builds a synthetic counterfactual from untreated geographies, and minimum viable test budgets have come down with it.

The governance matters as much as the math. Name the budget decision first, then design the test to answer it. Write the success threshold down before launch. Record who assigned the control group, how large the holdout was, and how wide the confidence interval came back.

Layer three: reconcile to revenue

Your source of truth is the checkout and billing record, Shopify or Stripe or your invoicing system, not the ad platform. Build a view in the CRM that carries first touch, self-reported source, pipeline created and closed revenue on the same row. That is where you find the channel with a mediocre cost per lead and the best win rate, which is the single most common hidden winner in B2B.

The mix shifts by business model. Ecommerce leans harder on geo holdouts and post-purchase surveys. B2B considered purchase leans on self-reported attribution at form fill plus pipeline and win rate by source, because the sample is usually too small for clean geo splits.

Who owns what

Split it, deliberately. RevOps owns the CRM revenue view. Media owns creative and cost diagnostics. Finance co-signs blended CAC and payback period.

And one line to put in your next agency conversation: the team spending the money should not assign its own control group or grade its own lift test. Separate the buying from the measurement. The measurement layer is the cleanest place to draw that line.

The cadence that ends the weekly panic

Review the wrong thing at the wrong frequency and you manufacture the Monday morning crisis.

• Weekly: creative performance and cost. Nothing else. No budget decisions.
• Monthly: pipeline and win rate by source, reconciled against closed revenue.
• Quarterly: incrementality results, and the budget reallocation that follows them.

Then the board number. Defend blended CAC against payback period. One number, total spend against total new customers, measured against how long it takes to earn the money back. Channel detail is supporting evidence, never the headline. The moment you present channel-level ROAS to a CFO, you have agreed to argue about a model instead of the business.

Attribution was never going to come back. What replaces it is a system you can defend, and most teams need help building it once before they can run it themselves. That build means mapping where real demand is created, including the AI assistants deciding your shortlist, and standing up measurement that reports revenue instead of pixels.

Frequently asked questions

Why did my attribution break all at once?

It did not break all at once, it just crossed a threshold. Privacy limits cut platform signal, dark social moved conversations into closed rooms, and AI assistants now build shortlists before buyers ever search your brand. Platform conversions still describe the last visible touch, but far less of the journey is visible than it used to be.

Should I stop looking at platform ROAS entirely?

No, demote it. Platform data is good at diagnosing creative performance, audience cost and bidding decisions, which is what it was built for. It should not decide budget allocation, because it systematically over-credits demand capture and under-credits the channels that create demand earlier in the journey.

Is self-reported attribution reliable enough to act on?

It is reliable if you design the question properly and treat it as one layer of three. Ruler Analytics found 47% of leads selecting the first dropdown option were inaccurately attributed, so randomize option order and include open text. It also carries recency bias, which is why you reconcile it against holdout tests and closed revenue in the CRM.

We are not big enough to run incrementality tests. What then?

Minimum viable test budgets have fallen, partly because Meta open-sourced GeoLift, which builds a synthetic counterfactual from untreated geographies. If your sample is genuinely too small for clean geo splits, which is common in B2B, lean harder on self-reported attribution at form fill plus pipeline and win rate by source.

What single metric should I take to the board?

Blended CAC against payback period. Total spend against total new customers, measured against how long it takes to earn the money back. Keep channel detail as supporting evidence, because presenting channel-level ROAS to a CFO turns the conversation into an argument about a model rather than the business.