What the model is asking for
Attribution is sold as the answer to an old question, which is where the money should go next. A person clicks an ad, comes back later to read a page, opens the emails that follow, and then fills in the form. A multi-touch model gives each of those steps a share of the customer and adds the shares up by channel, so the channel with the biggest total gets more budget next quarter. That is a better answer than crediting the last click. It is also a model built for a volume of customers that most small companies do not have.
The arithmetic underneath it needs volume. A share is a fraction of a customer, and a fraction of a customer only means something once there are enough of them to add up to whole people in every channel, every month. The models that work out the weights from the history itself need enough conversions to tell a pattern from a run of luck. Some of the tools will not switch that model on below a stated number of conversions a month.
Where a team's enquiries arrive a few at a time and a handful of those become customers, that condition isn't met. It is not close to being met. We have sat in reviews where a whole quarter's customers fitted on one side of a page, with the source written next to each name. That page was a better report than the attribution dashboard open on the other screen.
What the report shows at low volume
The report still produces a number, because the models with fixed weights never refuse to run. It divides the customers it can see between the touches it recorded and prints a percentage next to each channel. When the month's total is a handful of customers, any one of them can swing a channel's share on their own, and the order of the channels turns on which of them happened to search first. The chart draws that as a trend.
Nobody reading it is a statistician, and nobody was meant to be. They see paid search drop down the order and ask what went wrong with it, and somebody trims the budget on a movement that one customer produced. The next month that customer's channel is back where it was. We have watched budgets move on that basis, and the decision cost more than the report did.
Underneath the report, the tracking is thinner than it looks at this size. A good share of the enquiries a small company gets arrive by phone or through a referral from another customer, and neither leaves a trail the model can read. Consent banners take out more. So does a person who read the site on a phone and enquired from a laptop. What the model divides is the fraction of customers it could follow, and at low volume that fraction is whoever happened to click, which in most of the businesses we sit with is a different person from the one who usually buys.
The two numbers we take
We take two numbers per source and no model. The first is the first-touch source, meaning where the person first met the company, written onto the record at the moment the enquiry arrives. Where the form or the ad platform caught it, the field fills itself. Where it didn't, which is most of the phone calls, the person who answered asks and types the answer in the caller's words, even when the answer is that a friend mentioned us.
The second is one conversion count. We agree with the client what the win is before the first report goes out, and it is one thing, a booked meeting or a paid job, whichever the business is actually run on. Each month we count how many enquiries from each source became that thing. The report is two columns with one row per source. The total at the bottom has to match the number of customers the finance side already knows about.
That table can be read at any volume. It counts people. When it says that most of the month's customers came by referral and the rest from paid search, that is a fact the owner can check by remembering who those people were. Nobody has to trust a weighting. When the numbers are small the table says so, and when they move the reason is usually a person somebody in the room can name.
It also takes a short sitting each month. A team that has to maintain a model stops maintaining it around the time the campaign names drift and nobody reconciles the platform's conversions with the CRM's. A two-column table whose total has to match the invoice count gets checked, because whoever reads the invoices can see a mismatch.
The record attribution will need later
None of that is an argument against attribution. It is an argument about order. The record comes first, and the model comes once the record has enough in it. The table is the report, and alongside it we set up the parts of the record a model will read later, because they can't be added to the history afterwards. An enquiry whose source was never captured in the month it arrived has no source, whatever the team remembers about it later.
Every enquiry becomes a record when it arrives, through whichever channel, and the record carries the source and the time. Any link the team controls carries a tag. The tag names the campaign, and the campaign names follow one convention, written down, that does not change when a new person takes over the ads. The form and the first call both ask where the person heard of the company. That answer is stored in a field of its own, next to whatever the tracking caught, because the two disagree often enough to be worth keeping apart.
We keep the lost enquiries in the record as well. A model built later on a database from which the losses were tidied out will learn that every source converts, which is the one thing it must not learn. The platform's count of conversions gets reconciled against the CRM's count every month, in the same report. By the time there is enough volume to model, the two counts have agreed for long enough that somebody trusts them.
The monthly count at which a model becomes worth switching on is also written down, with the client. It comes from the tool the client already pays for, where the tool states one, and otherwise from what the client would accept as a fair sample. It sits at the foot of the monthly report where it gets looked at. Most months it is not reached. Writing it down means that when it is, nobody has to argue about whether the time has come.
