What a marketing data pipeline is, and what breaks in it.

A marketing data pipeline moves data from the platforms a client advertises on to the place you read numbers. Every version of it, however it is built, has the same four stages.

Extract pulls from each platform's API. Load puts what came back into storage. Transform reshapes it so figures from different platforms can sit in one query. Serve hands the result to a dashboard, an alert or a person.

People mean the first two when they say pipeline and it is the third that decides whether anyone trusts the output.

Each stage has a characteristic failure, and telling them apart is what makes a pipeline maintainable.

• Extract: the platform changed

  A field gets renamed, a permission expires, a rate limit tightens. These fail loudly if you are watching and silently if you are not, because an API that returns nothing still returns successfully.

• Load: it ran twice, or half

  A retry after a partial failure is the usual cause of duplicated rows, and duplicates inflate every figure downstream. A load that stopped early leaves a day that looks like a bad day.

• Transform: it is quietly wrong

  Nothing errors. A join at the wrong grain multiplies figures, a currency goes unconverted, a metric gets summed when it should be averaged. This is the stage that produces a wrong number rather than no number.

• Serve: the report disagrees with itself

  Two dashboards read two different tables, or one reads a filtered view somebody made and forgot. Both look right in isolation and cannot both be.

Extract looks like the hard part because it involves APIs, and it is the part every tool leads with. It is also the part that is genuinely solved: connectors exist, they are cheap, and none of them is a differentiator.

Transform is where the work is, and it is not technical work. It is deciding what a conversion means when three platforms count it differently, which currency each client reports in, and what grain a table should be at. Those are decisions somebody has to make and write down, and no tool makes them for you.

Building it yourself is reasonable when you have someone who will own it. The cost is not the initial build, which is a few weeks. It is that platforms change and the pipeline needs an owner for as long as it runs.

Buying is reasonable when you do not, which is most agencies. The thing to check when buying is not the connector count. It is whether the product does the transform stage, or whether it moves data and leaves that to you. Many do the second and describe it as the first.

All four stages, for every client, into a warehouse you own. Extract from 200+ sources daily, load into your own account, transform in SQL inside that account so you can read how a number was built, and serve the result to client dashboards, alerts and the chatbot from the same tables.

The reason all three agree is that there is only one set of numbers for them to read.