← all writing

· 6 min read

How do you know which channel worked when you are the whole team?

Attribution models were built for companies with analysts. What a one-person team can measure, the two questions worth answering, and what to stop tracking.

Marketing attribution is a discipline invented by companies that could afford to argue about it. It has models, decay curves, incrementality tests and an entire vendor category, and almost none of that transfers to somebody running four channels between other work. The useful version for a one-person team is much smaller than the literature, and getting there means deliberately abandoning most of it. (Disclosure: published by BidSurvivor, which sells ad slots and publishes each slot's numbers — the last section says what our own attribution can and cannot tell you.)

Why the standard models do not apply

The models you have read about — first touch, last touch, linear, time decay, data-driven attribution — all solve the same problem: dividing credit for one conversion across several touchpoints. They differ in how they divide it, and they all require enough conversions that the division is stable.

That last condition is the one that fails. With four hundred conversions a month, the choice between last-touch and time-decay moves real budget and the difference is measurable. With eleven conversions, every model returns a different answer, all of the answers are within noise of each other, and picking one is a preference dressed as an analysis. The sample size problem does not go away because the tooling is sophisticated.

There is a second failure that is less often admitted. Attribution models assume the touchpoints were observed. Podcast mentions, a link in someone's newsletter, a screenshot in a group chat, a search for your name after hearing it somewhere — these produce direct traffic with no referrer, and at small scale they are frequently the majority of what actually worked. A model that divides credit among the channels it can see will confidently allocate 100% of it to the wrong places.

The consequence: for a one-person team, attribution is not a modelling problem. It is a question of which two or three things you can observe cleanly enough to act on.

The two questions actually worth answering

Everything else is optional. These two decide what you do next week.

Did this specific spend produce more than it cost? Not "which channel deserves credit" — whether the fifty dollars came back. This is answerable without a model if you buy things that resolve fast: spend on one placement, count the visits it sent, multiply by what a visit is worth to you, compare. The answer is crude and it is enough, because the decision it feeds is binary.

Is the total going up? Weekly visits, weekly signups, weekly revenue, as three lines. If you added a channel and the total did not move, the channel did not work regardless of what any dashboard attributed to it. If the total moved and you cannot say why, you have still learned that something is working and can find it by subtraction — stop one thing for two weeks and watch.

Those two questions are answerable with analytics you already have and arithmetic you can do in a notebook. Neither requires resolving the credit-division problem, because at your scale you should be running few enough things at once that credit is not genuinely contested.

The setup that is worth twenty minutes

Tag every paid or placed link. One UTM parameter per source, spelled consistently. This is the entire difference between "some visits came from somewhere" and a countable line in your analytics, and it costs nothing per link once you have a naming habit. Inconsistent spelling is the most common way small teams destroy their own data.

Ask, in one field, at signup. "How did you hear about us?" as an optional free-text box catches everything UTMs structurally cannot: the podcast, the group chat, the friend. The responses are messy and self-reported and they routinely contradict the analytics — which is the point. Two disagreeing sources are more informative than one confident one.

Keep a dated log of what you did. A single file with one line per action: what you shipped, where you posted, what you paid for, on what date. When a spike appears you will otherwise spend an hour reconstructing the week from memory, and memory reliably credits whatever you were most proud of.

Run few things at once, on purpose. This is the highest-leverage attribution decision available to a one-person team, and it is not a tool. Two channels at a time makes credit obvious; six makes it unrecoverable, which is the argument behind the two-channel rule.

What to stop tracking

Multi-touch paths. You do not have the volume, the touchpoints are partly invisible, and no decision you make this quarter changes based on the answer.

Impressions and reach. They are denominators you cannot verify and they will not tell you whether the spend came back. Vanity metrics are not defined by being fake — most are accurately measured. They are defined by being unable to change a decision.

Last-click revenue by channel, reported to two decimal places. With eleven conversions, that precision is a presentation choice, not a measurement. Round it to whole conversions and the false confidence disappears with the decimals.

Anything you have never once acted on. A metric you have watched for three months without a single decision following from it is a habit. Delete the chart; you will not miss it.

What our own numbers can and cannot tell you

Every slot on our board publishes two figures: how many times a brand's card was opened and how many people it sent to that brand's site. Both are on the front page and on each brand's page, before anyone bids.

What that gives a buyer is the first of the two questions above, answered cleanly: this placement sent this many people, you paid this much, the arithmetic is yours to finish. What it cannot give is the second. We see visitors leaving for a brand's site; we do not see what they did after arriving, and we deliberately do not follow them there. So our numbers are a clean input to your attribution and never a substitute for it — if a slot sent you sixty visits and none of them converted, we will still be reporting sixty, and you are the only party who knows the placement failed.

Two slots a day, twelve hours each. The first brand into an empty slot pays nothing, a held slot starts from a $1.00 floor capped at $5,000, and every visitor starts with $100 of house credit. The counting is deliberately simple because at this scale simple counting is the only kind that survives.

BidSurvivor sells advertising in twelve-hour blocks, at auction. The first brand into an empty slot pays nothing, every account starts with $100 of house credit, and every brand's click-throughs are public before you bid.

See the board

Read next