The multi-touch attribution myth for small teams

The pitch goes like this: your customers touch seven channels before they buy, so crediting only the last click is naive. What you need is multi-touch attribution - a model that observes every touchpoint along every journey and distributes credit scientifically across them. Linear, time-decay, U-shaped, W-shaped, or, for the sophisticated, data-driven models that learn the weights themselves. Then, finally, you will know what your marketing is really worth.

For a large advertiser with millions of sessions and a data team, some version of this can earn its keep. This essay argues something narrower and, we think, more useful: for a small team - a founder, an agency pod, a two-person growth function - the multi-touch project is theater. Not because the math is wrong, but because at your scale the inputs are broken and the sample is too thin for the outputs to mean anything. And the alternative is not "no attribution." It is a humbler instrument you can actually trust.

Strip the vendor language away and a multi-touch model needs three things. First, observation: it must see every meaningful touchpoint, which means tracking users across sessions, devices, and weeks. Second, identity: it must know that the phone that saw the social ad and the laptop that submitted the form are the same person. Third, volume: enough completed journeys that the statistical machinery can distinguish signal from noise when it splits credit seven ways.

Each requirement is heavy on its own. Together they describe an apparatus that large companies staff whole teams to maintain. The question a small team should ask is not "is multi-touch more accurate in principle?" - it is "do I have any of these three inputs in working order?" For most, the honest answer is none of the three.

Start with arithmetic, because no vendor deck does. Suppose your form converts thirty leads a month - a healthy number for a small operation. A multi-touch model splitting each journey across several touchpoints is now allocating fractional credit from a few dozen observations across dozens of channel-position combinations. Most cells in that matrix hold one lead, or none. The difference between "webinars drive 12% of pipeline" and "webinars drive 4%" can be two people, one of whom found you because a friend texted them a link - a touch no model ever saw.

Models fed thin data do not fail loudly. They produce confident-looking percentages with enormous invisible error bars, and month-to-month swings that are pure noise get read as trends. At enterprise volume the law of large numbers rescues the exercise. At thirty leads a month, it cannot.

The deeper failure is upstream of the math: the tracking substrate multi-touch depends on is dissolving. Safari's Intelligent Tracking Prevention aggressively limits cross-site tracking and caps the lifetime of script-set cookies; Firefox ships similar protections; ad blockers strip analytics tags outright; consent banners mean a meaningful share of visitors are never tracked at all. Cross-device stitching without an enterprise identity graph is guesswork.

So the journey data feeding your model is not a complete record with a little noise - it is a fragmentary record with systematic holes, and the holes are biased toward exactly the privacy-conscious and mobile users you most misunderstand. A model is an amplifier of its inputs. Garbage in does not become insight because the model that processed it was sophisticated; it becomes sophisticated-looking garbage.

Why does the ritual persist? Because attribution dashboards perform a social function that has little to do with measurement. A credit-allocation chart looks like science. It gives a marketing function something rigorous-seeming to present, and it diffuses accountability - every channel gets a slice, so no channel's owner is ever plainly wrong. That is precisely what makes it theater: at small scale, the numbers move, the meetings happen, and no decision changes that a simpler measure would not have driven.

The test we would apply to any measurement system is blunt: name a decision you would make differently if the model said 30% instead of 20%. If you cannot - and with a small team's budget concentrated in two or three channels, you usually cannot - the model is a costume, and its maintenance cost is coming out of the same small budget it claims to optimize.

Here is the humbler instrument: capture first touch and last touch on every lead, with near-total reliability, and stop there. Concretely, that means every form submission should land with the campaign parameters that brought the visitor - utm_source, utm_medium, utm_campaign - and the ad click identifiers that make platform reporting exact: gclid for Google, fbclid for Meta, msclkid for Microsoft, and the newer wbraid and gbraid variants.

Two touches per lead sounds impoverished next to a seven-touch journey map. But these two have a property the journey map lacks: they are observed at the moment of conversion, attached to a named lead, sitting in a row you can audit. "This lead arrived from the March pricing campaign" is a fact. "Social assisted 23% of conversions" is, at your scale, an artifact. Reliable facts beat elaborate artifacts.

The catch: even first/last touch is usually captured badly. The standard failure is the multi-page journey - the ad lands on a blog post, the visitor browses, and by the time they reach the form the URL parameters are three pages gone. The fix is mechanical: capture the parameters on arrival, persist them in the browser, and attach them at submission time. Our embed script does exactly this - it stores utm and click-ID parameters in localStorage and appends them to any bound form, which is what makes multi-page attribution work without custom code.

The other half of the bar is where the data lands. Attribution stored in a analytics silo you check quarterly is decoration; attribution in the same spreadsheet row as the lead's name is operational - sortable, pivotable, and joinable to what actually closed. The setup takes minutes, as the gclid capture guide and the ad landing page use case both walk through.

A contrarian argument owes you its costs, so here they are. First/last touch makes assists invisible: the podcast that warmed a lead up gets no credit if the last click came through search, and over time that systematically flatters bottom-of-funnel channels and undervalues content and brand. If you steer spend purely by last touch, you will slowly starve the channels that fill the top of the funnel.

The mitigations are unglamorous but real. Add a "How did you hear about us?" field - self-reported attribution is imprecise, but it surfaces exactly the dark-funnel touches (word of mouth, podcasts, communities) that no tracker sees, and at small scale you can read every answer. Watch branded search volume as a proxy for awareness. And accept that some channel judgment stays judgment. Losing the pretense of precision is not the same as losing knowledge; it is trading a false instrument for a smaller honest one plus your own eyes.

There is one more move available to a small team, and it embarrasses most multi-touch setups: close the loop on revenue. Because each lead's row carries its gclid, you can mark which leads actually became customers and feed those offline conversions back to the ad platform. The platform's own optimization then works from real outcomes - deals, not form fills.

Notice what this is: attribution in service of a decision. The click ID exists to answer "which campaigns produce customers?" and to make the ad platform act on the answer. No fractional credit, no model - a join between a spreadsheet column and a revenue column. For most small teams this single loop is worth more than any attribution dashboard they could buy.

None of this says multi-touch attribution is fake. At sufficient volume, with owned identity and a team to keep the pipes honest, data-driven models can genuinely inform budget allocation - that is a real thing that exists at enterprise scale. The myth is transferability: the assumption that a small team running the same ritual gets a smaller version of the same insight. It does not. It gets the costume without the body.

Our position, then: capture two touches perfectly rather than seven badly. Land them in the same row as the lead. Close the loop with click IDs. Ask people how they heard of you, and keep judgment in the loop for the rest. It is less impressive in a slide deck - and considerably more true. The mechanics of the capture side are on how it works, and the lead attribution glossary entry defines the terms if you want them straight.

What is multi-touch attribution?

Why does multi-touch attribution fail at small scale?

What should a small team measure instead?

Isn't last-touch attribution known to be biased?

How do utm parameters survive a multi-page visit before the form?

What is a gclid and why capture it on forms?

Is self-reported attribution reliable?

When does multi-touch attribution actually make sense?