Marketing attribution is not a truth machine. It’s a decision-support tool, and mid-market companies keep buying it as the former and abandoning it as the latter.
Here’s the pattern. A company between $5M and $50M spends across paid search, paid social, SEO, email, events, maybe radio or direct mail. Two or three agencies each report success in their own dashboard. Every dashboard shows a win. Total pipeline doesn’t move the way the dashboards say it should. Leadership concludes the measurement is broken, buys a more sophisticated attribution model, watches nobody use it, and goes back to gut feel.
The problem was never the sophistication of the model. The problem was the question being asked of it. Attribution can tell you which recorded touchpoints preceded conversions. It cannot tell you what caused revenue. Only deliberate testing answers that. Once you accept that boundary, attribution becomes genuinely useful — as one input into the only marketing question that matters at this size: what do we keep, cut, or test?
What is marketing attribution actually for?
Attribution exists to inform keep, cut, and test decisions. It does not exist to produce a perfect map of credit, and it never will, because a perfect map of credit is not achievable with the data any mid-market company actually has.
That distinction matters more than any modeling choice. When attribution is treated as decision support, the standard is “does this help us make a better budget call than we’d make without it?” That’s a clearable bar. When attribution is treated as a truth machine, the standard becomes “is this number correct?” — and it never is, so the whole apparatus gets discredited and abandoned.
There’s a second failure mode worth naming. The moment attribution becomes a scorekeeping exercise between agencies, it stops informing decisions. If your paid search agency and your paid social agency are each arguing for credit on the same closed deal, the argument consumes the meeting, the CEO adjudicates a dispute instead of allocating a budget, and the actual question — should we spend more or less here next quarter — never gets asked. Attribution pointed at vendors produces politics. Attribution pointed at decisions produces decisions.
So before touching a single attribution model, decide what decisions it needs to serve. For most companies at this size, the list is short: which channels get more budget, which get less, which get a real test, and which get shut off. Everything else is reporting theater.
What are the attribution options, honestly compared?
For most mid-market companies, the right answer is last-touch attribution plus triangulation. Not because last-touch is accurate — it isn’t — but because its biases are knowable, and a simple model read honestly beats a complex model read credulously.
Here’s the honest comparison.
Last-touch gives all credit to the final recorded touchpoint before conversion. It’s simple to run and simple to interrogate. Its bias is predictable: it flatters channels that close, especially branded search, which frequently catches demand that some other channel created. If you know that going in, you can read the report with the right discount applied.
First-touch gives all credit to the first recorded touchpoint. Same simplicity, mirror-image bias: it flatters channels that open conversations and ignores everything that moved the deal afterward. Useful as a second lens. Dangerous as the only lens.
Multi-touch models spread credit across touchpoints by formula — linear, time-decay, position-based, or an algorithmic weighting. This is where mid-market companies most often go wrong, because multi-touch looks like precision. It isn’t. The formula is a choice, not a discovery. Someone decided the first touch gets a certain share and the last touch gets another share, and that decision is exactly as arbitrary as picking last-touch — except now it’s buried inside a model nobody in the room can explain. A last-touch report read with known blind spots beats a multi-touch model nobody understands. If your team can’t articulate why the model weights touchpoints the way it does, the model is making your budget decisions for you, and you can’t audit it.
Triangulation isn’t a model; it’s a discipline. You put three imperfect sources next to each other: your attribution report, your direct-response data (what happens when a specific campaign runs against a specific offer), and self-reported answers to “how did you hear about us?” asked on every form and every sales call. Each source is flawed differently. Attribution misses offline and word-of-mouth. Self-reported answers are fuzzy on specifics but surprisingly honest about origins. Direct-response data is clean but narrow. When two of the three agree, you can act with reasonable confidence. When they contradict, you’ve found exactly where to run a test.
Last-touch plus triangulation costs almost nothing, requires no new software, and produces something a multi-touch platform rarely does: a marketing measurement practice the CEO actually understands and therefore actually uses.
Where does attribution lie to you?
Every attribution model lies in four predictable places. Knowing them is what “reading a simple report honestly” means in practice.
Assist channels look worthless. A channel that warms buyers but rarely takes the final click — content, email nurture, most paid social — scores poorly in a last-touch report even when it’s doing real work. Read that low score as “didn’t close” rather than “didn’t contribute,” and hold the cut decision until a test confirms it.
Harvest channels look heroic. Branded search is the classic case. Someone hears about you at an event, from a podcast, from a colleague — then types your name into a search engine and clicks the paid ad above your own organic listing. The attribution report credits paid search with a conversion that was already decided. The channel harvested demand; the report says it created demand. This single distortion probably misallocates more mid-market budget than any other.
Offline and word-of-mouth are invisible. Radio, direct mail, events, referrals, a rave review over dinner — none of it leaves a recorded touchpoint unless you build a way to capture it. An attribution model doesn’t score these channels low. It doesn’t score them at all, which reads as zero. This is why “how did you hear about us?” belongs on every form and in every first sales conversation. It’s the cheapest attribution instrument you will ever deploy, and it’s the only one that sees word-of-mouth.
Every agency dashboard grades its own homework. Each platform and each agency reports using its own attribution settings — its own lookback windows, its own conversion definitions, its own view-through rules. Sum the conversions across three agency dashboards and you’ll routinely get a number larger than the deals in your CRM. No one is necessarily lying. Each dashboard is answering a slightly different question with settings chosen by the party being evaluated. That is precisely why the shared report has to live on your side of the table. (The full anatomy of that disagreement — why analytics, the CRM, and vendor reports never match — is its own post.)
None of these lies means attribution is useless. It means attribution is a witness with known biases. Cross-examine it. Don’t hand it the verdict.
What should you keep, cut, and test this quarter?
The decision framework is three rules, applied in order.
Keep what attribution and common sense agree on. When the report says a channel drives pipeline and your sales team’s anecdotes, your “how did you hear about us” answers, and your own judgment all point the same direction, stop analyzing. Fund it. Agreement across independent signals is the strongest confidence you’ll get, and re-litigating settled channels every quarter burns leadership attention you need elsewhere.
Test — do not cut — what attribution scores low but plausibly assists. This is the rule that saves companies from the classic self-inflicted wound: cutting an assist channel because a last-touch report scored it near zero, then watching the “efficient” closing channels quietly starve two quarters later. Low attribution plus a plausible assist story equals a testing question, not a cutting decision.
Cut only after a deliberate reduction test confirms nothing breaks. A reduction test is simple in principle. Turn a channel down — meaningfully down, not a trim — in one region, one segment, or one defined period, while holding everything else steady. Then watch total pipeline, not channel metrics. Channel metrics will always tell you the channel matters; that’s what channel metrics do. The question a reduction test answers is the one attribution can’t: does the business feel the absence? If total pipeline in the reduced region holds while the rest of the business tracks normally, you’ve learned the channel was harvesting or duplicating, and you can cut with confidence. If pipeline sags, you’ve learned the channel was load-bearing, and the report that scored it low was lying in one of the four ways above.
Give the test enough runway to be legible against your sales cycle. A company with a nine-month enterprise motion cannot read a two-week test. And run one test at a time per decision. The point of first-principles thinking here is isolation: change one variable, hold the rest, read the result against revenue. Two simultaneous reductions produce a result you can’t interpret, which is worse than no test at all.
This is also where attribution and testing snap together into something coherent. Attribution generates the hypotheses — here’s what looks strong, here’s what looks weak, here’s where the sources disagree. Testing renders the verdicts. Companies that skip the first step test randomly. Companies that skip the second step reallocate budget based on a biased witness. You need the loop.
What does a trustworthy attribution setup look like at this size?
It looks like discipline, not software. Four components, none of which require a new platform to start.
UTM discipline on everything inbound. Every link you control — every email, every ad, every social post, every QR code on a direct mail piece — carries consistent, documented tracking parameters. One naming convention, written down, enforced across every agency. This is unglamorous work and it is the foundation everything else stands on. Inconsistent tagging is the most common reason mid-market attribution data is unusable, and it’s entirely self-inflicted.
One shared report instead of per-agency dashboards. Agencies still run their platform dashboards for optimization; that’s their job. But budget decisions get made from a single report you own, built on your definitions: your conversion events, your lookback window, your pipeline stages. Same numbers in front of everyone, every month. This one move ends the scorekeeping wars, because there’s no longer a private scoreboard to argue from.
CRM as the revenue source of truth. The question is never “which channel got conversions” — conversions are a proxy. The question is which channels precede pipeline and closed revenue, and only your CRM can answer it. Every ad platform will claim conversions the CRM can’t find as deals. When the platform and the CRM disagree, the CRM wins. Always. Composing your marketing measurement around CRM revenue keeps every conversation tied to the numbers that actually predict revenue — the ones the board cares about.
A standing quarterly test calendar. Testing can’t be a reaction to a scary dashboard. Each quarter, the leadership team picks one or two channels where the triangulated sources disagree, defines the reduction or expansion test in advance — what changes, for how long, what total-pipeline outcome would trigger a cut or a scale-up — and assigns clear decision ownership for the call at the end. Pre-committing to the decision rule is what keeps the result from being argued into whatever conclusion was already preferred.
That’s the whole architecture. Notice what’s absent: a six-figure attribution platform, a data science hire, a twelve-week implementation. Those can come later, if the decisions ever demand them. Most never do. What the setup demands instead is harder to buy — the discipline to tag consistently, the authority to consolidate reporting, and the patience to let a test finish before reacting.
The bridge from measurement to decisions
At $5M, a founder can feel where revenue comes from. Somewhere on the way to $50M, that feel stops scaling, and the instinct is to buy certainty — a model, a platform, a dashboard that finally settles it. But certainty was never for sale. What’s available is something more useful: a simple attribution model read with honest eyes, triangulated against what customers actually say, and settled by tests that watch revenue instead of channel metrics. Companies that build that loop stop having the agency-dashboard argument entirely, because keep, cut, and test decisions get made from one shared view of the business — holistically, on the founder’s terms, alongside the numbers the board sees. Building that view is exactly the kind of work a fractional marketing leader exists to do: composing the measurement, the test calendar, and the decision ownership so that every marketing dollar has to answer to pipeline. If your dashboards all say you’re winning and your gut says something’s off, that gap is the place to start.