Fewer visits can signal a problem. They can also mean part of the journey takes place elsewhere. Without additional data, neither explanation should become a verdict.
Likewise, a brand appearing in an AI response does not prove that the response produced a sale. Measurement needs to separate observation from attribution.
Four levels, four questions
| Level | Useful question |
|---|---|
| Visibility | Where and in what context does the offer appear? |
| Interaction | What visits, inquiries, or actions do we observe? |
| Commercial outcome | Which orders are paid, or which inquiries become customers? |
| Economic quality | What remains after variable costs, cancellations, and returns? |
No level replaces the next. A useful report connects them and identifies where the evidence is insufficient.
A traffic change does not explain the outcome
In a fictional scenario, a store drops from 10,000 to 8,000 sessions while total orders across all channels rise from 200 to 210. We have not shown that AI increased sales. We cannot calculate website conversion rate by dividing all orders by website sessions if some purchases finished elsewhere.
We would separate channels, duplicates, cancellations, and changes in pricing or budgets before evaluating possible explanations.
What each source can support
The order system can show what was purchased and whether the transaction completed. Website analytics describes observed interactions there. Platform data must be interpreted using its own definitions and limitations.
Search Console’s Generative AI performance report shows impressions in Search’s AI features. We would check the property’s available data. Impressions do not establish attributed sales. [1]
A sample of AI responses can help identify errors about the offer. It cannot, by itself, establish market share or revenue attribution.
How we would build the evaluation
At NOT MORE., we would start with the business question: more profitable sales, suitable inquiries, or fewer order errors?
We would define metrics, sources, shared identifiers, and deduplication rules. We would keep a baseline period and a change log. Comparisons would account for product availability, budgets, and seasonality.
We would decide in advance which outcome supports expanding a pilot, which error requires stopping it, and which missing data limits the conclusion. We would not invent a universal threshold for every store.
More data helps when it clarifies a decision. A new metric is not a strategy on its own.
Do your reports capture the complete commercial outcome?
We can clarify what you measure, what you can attribute, and which questions remain open.
Sources
[1] Google Search Console: Generative AI performance report (Search)