A client's dashboard reported 41 conversions from paid social last month. The platform reported 88. Their sales team could identify 12 deals that had actually come from that channel by asking people.
Three numbers, three methodologies, one reality. None of them was lying, and none of them was right.
This is the normal condition now, and the mistake is trying to reconcile the numbers. You cannot. What you can do is build a measurement stack where each layer answers a question it is genuinely capable of answering.
Why the old model broke
Deterministic click attribution required a persistent identifier following one person across sites and sessions. That identifier is gone or degraded across most of the ecosystem: third party cookies restricted, mobile app tracking gated behind a permission prompt most people decline, link decoration stripped, and cookie lifetimes on some browsers cut to a week.
Every platform responded by modeling. When they cannot observe a conversion, they estimate one based on aggregate patterns. That is a reasonable engineering answer and it means platform reported conversions are now partly a prediction rather than a count.
Which produces the specific pathology of the moment: every channel claims credit, the sum of claimed conversions exceeds actual conversions, and the marketer in the middle is asked to explain the discrepancy.
Stop trying. Build this instead.
Layer one: self reported attribution
The highest value measurement change I have made in five years is one field on the form: "How did you hear about us?" Open text or a short list, required.
It is unfashionable because it is not technical and it is imprecise. People misremember, they say Google when they mean a podcast that made them search Google, and the data is messy.
It is also the only source that captures the channels nothing else can see: a recommendation in a private conversation, a podcast, a conference, a forum thread. Those channels are frequently the most valuable and completely invisible to every analytics platform.
The pattern I see repeatedly is that self reported data credits brand and word of mouth channels far more than any dashboard does, and that gap is roughly the size of the thing your dashboards systematically undervalue. I would rather have directionally right data about all channels than precisely wrong data about half of them.
Ask it at the point of highest commitment, not the first touch. On the demo request, on the checkout, on the intake call.
Layer two: disciplined tagging
Everything you control should be tagged consistently. Email links, social posts, partner placements, QR codes, podcast URLs, newsletter sponsorships.
This is not a technology problem, it is a discipline problem. The failure is always that three people tag things three different ways and six months later "facebook," "Facebook," and "fb" are separate rows.
Write a convention: lowercase, defined source values, defined medium values, campaign names that encode date and initiative. Then make everyone use a UTM builder rather than hand editing URLs, because hand editing is where the inconsistency enters.
This layer is not truth. It tells you what people clicked immediately before arriving, which is one fact about the journey, not the journey.
Layer three: server side collection
Moving your tracking from the browser to a server you control recovers a meaningful share of the data lost to blockers, restrictions, and short cookie lifetimes. It also gives you control over what you send to whom, which matters increasingly for privacy compliance.
It is not a magic restoration of the old world. It improves data quality, it does not resurrect cross site identity. But it is the most reliable technical improvement available and it has the side benefit of taking script weight off the user's device, which helps the metrics in Core Web Vitals in practice.
The practical version, including what it does and does not fix, is in server side tracking basics.
Layer four: holdout testing
The only method that establishes causation rather than correlation.
Turn a channel off in a defined geography or audience segment for two to four weeks. Compare total conversions in that segment against a matched control. The difference is the channel's real incremental contribution.
This is uncomfortable because it means deliberately not spending money in a market. It is also the only way to answer the question every executive asks, which is what happens if we stop.
I have run this and found channels contributing roughly what they claimed, and channels contributing close to nothing while reporting healthy numbers, mostly by taking credit for people who would have purchased regardless. Branded search is the classic case.
Size the test properly. An underpowered holdout produces a confident wrong answer, so check that your volume can actually resolve the difference you are looking for before you spend a month on it.
Layer five: the aggregate correlation
At the top, the crude one. Weekly spend by channel against weekly total conversions and revenue, looked at over a long enough window.
Not a model, just a chart. Over 12 to 18 months, real relationships become visible: the channel where spend increases and total revenue does not, the channel with a four week lag, the seasonal pattern that explains what you thought was a campaign effect.
Sophisticated teams formalize this into media mix modeling. Most businesses do not have the data volume for that to be meaningful and should just look at the chart honestly. The analytical instincts required are the ones in data driven marketing decisions.
What to tell the person asking for one number
You will be asked which channel drove revenue. The answer that keeps your credibility intact is to give a range and a confidence level rather than a false precision.
"Paid social contributed somewhere between 20 and 35 percent of new customers last quarter. Platform reporting says the top of that range, self reported data says the bottom, and our holdout test in Florida suggested the middle." That is an honest, useful, defensible answer.
The alternative, which is to report the platform number as fact, works until someone checks, and then every number you have ever reported becomes suspect.
Pick the metric that survives measurement collapse
The strategic conclusion I have reached: build your business on a metric that does not depend on attribution at all.
Total new customers per month against total marketing spend per month. Blended customer acquisition cost. Payback period. These do not require knowing which touch mattered, they cannot be inflated by platforms competing for credit, and they are the numbers that actually determine whether the business works.
Then use the five layers to allocate within that total, accepting that allocation is a judgment call informed by evidence rather than an output of a report.
Add the self reported field to your forms this week. It takes ten minutes, it costs nothing, and within one quarter it will tell you something about your business that no dashboard you currently own has ever shown you. That has been true for every single client I have convinced to do it.