Credit is a model. Impact is a question.

The path records the interactions your system can include. The model decides how credit is distributed across eligible interactions. A classroom first-touch rule can give the full value to the first recorded click, while a classroom last-touch rule gives it to the last. A linear illustration can split credit equally. These explain the idea; they are not a list of current choices in every product.

Google’s current Analytics documentation describes data-driven attribution and specific last-click models, and says first-click, linear, time-decay and position-based models are no longer available in GA4 attribution reports. Its documented direct-traffic rules also matter. Check the actual report and settings rather than copying a textbook model list into a current account tutorial.

Record the outcome definition, model, eligible channel scope, lookback window and reporting period. A click outside the window may not be eligible even if it was important to the person. An offline recommendation or activity on an unconnected device may be absent from the observed path. ‘First recorded touch’ need not mean the first time someone encountered the business.

Keep comparison settings consistent or explicitly record their changes. If two platforms claim credit for the same booking, do not add their credited values as if they were separate bookings. They may apply different windows, identity rules and models to overlapping activity. Reconcile the actual business record and preserve the different reporting perspectives before drawing a budget conclusion.

A model output tells you how that model allocated credit under its method and available data. It can help identify patterns and questions. It does not, on its own, prove that removing the credited channel would remove the same amount of business. Likewise, an increase after a page edit can coincide with other changes in traffic, demand, offer or collection.

Data-driven models can use sophisticated methods; calling something algorithmic does not remove the need to understand its scope and assumptions. Treat the reported credit as one piece of evidence. If your question is whether a specific change adds useful outcomes, choose a comparison designed for that question rather than treating a before-and-after attribution chart as a complete answer.

Write the hypothesis, the single intended change and the measure before reading a result. For example, test whether clearer request-button wording improves the defined completed-request outcome. Keep the offer and measurement stable, use an appropriate comparison setup and record any other changes. Google’s campaign-experiment guidance similarly emphasizes a clear hypothesis, preselected measures and isolating the variable being tested.

Check whether the observations are comparable and sufficiently mature for the decision. Missing records, small samples, delayed outcomes or a changed audience can limit interpretation. The useful result can be ‘not enough evidence yet’ or ‘repair measurement and repeat.’ Record what was observed, its limitations and the next decision, with an owner. This lesson plans a test; it does not run one or establish a statistically significant effect.