Multi-touch Attribution

Deciding which touchpoint gets the credit for a conversion was for a long time shaped as one touchpoint taking everything, made more so by ad tools defaulting to giving the last touchpoint 100 percent.

Under that default the last touchpoint is usually a search for the company name. Looking at one $540 contract at Pagedeck that way, the brand-name search ad took the whole $540. That person had read a blog post two months earlier, compared the product against a competitor three weeks before that, and searched the name at the end — and the first two scored zero on the report.

Multi-touch splits the credit instead. Dividing three touchpoints 30-30-40 gets both the activity that brought the person in and the activity that closed them onto the board.

Five decisions.

1. Write down what counts as a touchpoint. Clicks only, or impressions too. Include impressions and the number of touchpoints multiplies, and the channel ranking shifts with it.
2. Set the window. If the attribution window — how many days after a touchpoint still count — is seven days, somebody who buys ten days after seeing the ad becomes an unrelated conversion. Products with long consideration periods get 30 days or more.
3. Set the weights. Split them evenly, or give more to touchpoints nearer the conversion. There is no objective basis for this. The team decides and writes the numbers into a document; that's the whole step.
4. Look at the same month through other models. Putting first-touch and last-touch side by side separates which channels bring people in from which ones close.
5. Record the date and reason for any change. Changing the rule flips the channel ranking on the same month's report, so without a record there's no way to compare quarters.

The limits follow from that. Change 30-30-40 to 20-30-50 and channels move a place each on identical data. When first and second are close together, this report doesn't get used as a budget argument. Plenty of touchpoints leave no record at all: a colleague recommending the product in conversation appears in no model.

"By last touch, brand search is first. Let's run it through multi-touch once."

Doubting a report where one touchpoint takes everything.

"Weights are 30-30-40 and it's written down. We aren't changing it until next quarter."

Fixing a value that has no basis, rather than tuning it.

"If you don't note the date the model changed, we can't compare with last quarter."

Asking for a change log on the rule.

At Pagedeck, Ravi builds this report and Grace reads it in the budget meeting. Any channel that ranks differently under the two models keeps its budget that month and gets another quarter of observation.