Blog / AI & search /
What an AI assistant cannot infer from an analytics export
The export has one row for a collection page: fewer clicks, lower click-through rate. The AI assistant announces that customers no longer trust the brand.
Dramatic. Also nowhere in the file. It’s a plausible story wrapped around a handful of observations. An analytics export can help an assistant group, calculate and raise questions. It almost never establishes the motives or causes behind a satisfying narrative.
Before I ask for interpretation, I decide what the data can support. The best assistant is the one that shows its uncertainty and helps pick the next investigation, not the one that fills every gap with an explanation.
An event does not reveal a person’s motive
A click records an action under one measurement definition. It doesn’t say whether the visitor was interested, confused, comparing options or fixing a mistake. The same behaviour can come from different intentions.
Give the assistant the event definitions and the page context. Ask it to describe the patterns without assigning motives, then list the questions customer feedback or a usability review could answer.
Say visitors keep opening a delivery section. That could be sensible checking or unclear information elsewhere. The export alone can’t choose. Treat both as hypotheses until another source gives evidence.
Missing data does not establish missing activity
Collection is shaped by consent, configuration, implementation faults and the scope of the export. A missing event may mean an unmeasured route, not an action nobody took.
Document which systems and pages the file covers. If the assistant doesn’t know the coverage, it mistakes the edge of the export for the edge of the business. A report limited to one device type isn’t the whole store.
Check for tracking changes before interpreting anything. If the event definition changed mid-period, the comparison may describe a measurement fix as a shift in customer behaviour.
A sequence does not automatically prove causation
A page edit followed by a traffic rise is a useful observation. It doesn’t show the edit caused the rise. Promotions, demand, stock and other releases overlap the same period.
Ask the assistant to separate the timeline from the explanation. It can identify the dates and compare groups where the file supports that work. It should also identify plausible competing explanations and the evidence needed to assess them.
I avoid prompts that say “explain why” before the analysis has established what changed. That wording invites a finished story, when the more useful output is a short list of open questions.
Aggregate rows cannot reconstruct every journey
A table grouped by page or channel throws detail away. The assistant can’t reliably rebuild individual customer paths from aggregates. It shouldn’t claim someone arrived from one source and later bought through another unless the authorised data supports that.
Keep definitions clear when joining sources. Search Console clicks and analytics sessions describe different measurement systems. Matching page names does not make the rows a single record of the same people.
Use each source for the evidence it provides. Combining them can improve questions and context, but it should not manufacture precision that neither source contains. A useful join needs an explicit basis and known limits.
Business context cannot be guessed from product names
A model may guess from a title that an item is seasonal, expensive or popular. Those guesses aren’t facts about your store. Supply verified context where it matters, such as availability changes or a documented campaign.
Avoid sharing unnecessary private information to make the context richer. Use approved tools and proportionate data. The assistant needs enough to answer the question, not every customer detail the business holds.
If a field is ambiguous, ask the responsible owner to explain it. A column called “value” could represent several things. Letting the assistant choose a definition can make every later calculation internally consistent and still wrong.
A recommendation needs a decision framework
The largest percentage change is not automatically the highest priority. The assistant needs information about customer impact, business purpose, confidence and effort to propose useful next actions.
Keep those factors visible rather than hiding them inside a request for the “best opportunities”. Ask for a rationale and an uncertainty statement beside each candidate. A small but consequential compatibility issue may outrank a large movement in an unimportant report.
Human review should challenge the recommendation against the actual page and workflow. The export can point to a place worth inspecting. It cannot decide whether the proposed repair is accurate or safe to publish.
Ask for analysis that can be checked
Request clear calculations, definitions and row references where supported. Have the assistant identify excluded data and assumptions. Recheck important arithmetic and inspect the source rows behind consequential findings.
End with a distinction between observations, hypotheses and next evidence. That structure makes the output usable without pretending the file contains every answer. It also gives the team a route to learn more.
Ask the assistant what extra evidence each consequential hypothesis would need. That turns an attractive explanation into an investigation plan. If the suggested evidence can’t tell the competing explanations apart, refine the question before you spend time collecting it.
Related reading
- A practical test for AI-written product descriptions
- How to document human review in an AI content workflow
- AI for internal linking suggestions: a workflow
Take one explanation an AI gave you from your last analytics export. Which part do the rows directly support, and which part needs evidence the file never held?