AI Overviews and falling clicks: investigating the effect on sales

Analyse stable page groups, comparable periods and conversions without confusing simultaneous changes with proof of AI’s business impact.

Start with a testable question

Choose your own page group and period instead of asking whether AI removes everyone’s traffic. One hypothesis concerns guides that remain visible but attract fewer visits. Examine offer pages separately. The analysis should identify a content or sales-path decision rather than merely confirm an opinion about AI.

Do not combine incompatible counters

The dedicated generative AI report in Search Console describes impressions. Dividing all Google clicks by AI impressions does not produce AI CTR. Search Console clicks and Analytics sessions are also different events. Compare visibility, overall organic traffic and outcomes without implying unsupported user-level attribution.

Build stable page groups

Separate informational guides, comparisons and offers. Keep the same URLs within each group throughout the period and show new publications separately. Align country, device, language and dates. Publishing growth can hide deterioration in an older cluster. A fixed group reveals composition effects but cannot establish their cause.

Investigate other explanations

Record title changes, redirects, indexing problems, seasonality, campaigns and offer changes. Compare relevant earlier periods and pages with similar intent. Do not automatically call those pages a control group: the same phenomenon may affect them. Otherwise a redesign or reduced demand may be incorrectly labelled an AI effect.

Evaluate outcomes as well as visits

An illustrative 1,000 visits and 20 enquiries gives 2% conversion; 700 visits and 21 enquiries gives 3%. Traffic fell 30%, but enquiry volume did not. This does not prove an AI benefit. Check quality, customers and margin. With small counts, a few events can shift percentages substantially.

Log content tests clearly

Choose pages with a concrete gap: missing conditions, outdated information or no route to an offer. Record the date and scope. Avoid simultaneous changes and allow reprocessing plus the buying cycle. Improvement after an edit remains observation rather than clean causal measurement without an appropriate study design.

State the confidence of the conclusion

Report what changed, what cannot be separated and which decision the evidence supports. Weaker informational traffic may coexist with stable leads; declining sales may need investigation. Do not automatically write that AI removed 30% of customers. Management needs an honest assessment and next test.

FAQ

Do rising AI impressions and falling clicks prove cannibalisation?

No. They show simultaneous changes. Check alternative causes, data scope and business outcomes.

Is lower traffic always worse marketing?

No. Evaluate qualified enquiries, customers and margin without assuming improved visitor quality unless measured.

Sources

Information checked: 26 September 2026. Feature access may depend on the account.