Attribution and marketing mix modelling, compared.

Attribution works from the bottom up. It follows individual people, sees which ads they touched, and assigns credit for what they did. It answers the question of which touchpoints appeared on the path to a sale.

Marketing mix modelling works from the top down. It takes spend by channel over time, plus everything else that was happening, and estimates what each channel contributed in aggregate. It answers the question of what would have happened with more or less spend.

Both are legitimate. Neither is a version of the other, and they will not produce the same answer.

The failure modes are opposite, which is why teams that use only one keep hitting the same wall.

• Attribution cannot see what it cannot track

  Anything that happens off the client's site, in an app, on a different device or under a privacy control is invisible to it. Channels that influence people without a trackable click get systematically undercredited.

• Attribution assigns credit by rule, not by cause

  Whether the last click or the first click gets the sale is a rule somebody chose. Changing the rule changes the answer without anything real having changed.

• Modelling needs a lot of history

  It works on variation over time, so it needs enough months of spend, and enough change in that spend, to separate one channel from another. A new account does not have that yet.

• Modelling cannot tell you which ad to pause

  It works at channel level. Useful for deciding a budget, useless for deciding a creative, and treating a channel-level estimate as a verdict on one campaign is a common misuse.

When attribution says a channel is strong and modelling says it is weak, the usual explanation is that the channel is good at being present near a conversion and less good at causing one. When the reverse happens, the channel is probably doing work that cannot be tracked.

The unhelpful response is to pick whichever number supports the decision you already wanted. The useful one is to treat the disagreement as the finding: it tells you where the tracking is blind, or where the credit rules are generous. It is also a better conversation to have with a client than defending one number against the other.

Neither works on inconsistent data, and this is the part that gets skipped. Modelling needs spend by channel by day, in one currency, with the channel names meaning the same thing every month. Attribution needs the tags on the client's links to be right, because they are the join between the ad and the behaviour.

A client whose metric definitions and campaign naming are not settled is not ready for either method, and running one anyway produces a confident number built on a shifting base.

SightShark is the layer underneath both. Spend, impressions and conversions from each client's platforms land in a warehouse you own and are modelled to one name, one type, one grain and one currency, which is the input format both methods need and neither provides.

History is kept rather than overwritten, so a model built on the last two years is built on what the platforms reported at the time, including their later corrections.