Start with the planning question
A claim about AI search market share is incomplete until it explains what is being divided by what. Visits, searches, people, and assistant interactions describe different things. Before comparing a conversational product with a search engine, decide which behavior matters to the decision you are making.
A media planner may care about reachable audiences, while a website owner may care about referred visits. A researcher studying information discovery may care about question sessions. None of those questions is answered automatically by a platform’s total traffic or a count of all messages.
Name the numerator and denominator
Write the metric in a sentence before calculating it: assistant-attributed visits to the observed websites divided by all observed visits from the defined discovery channels. This is a referral share within that dataset. It is not the assistant’s share of all information seeking.
Alternatively, you could divide observed search-like assistant sessions by all observed search-like sessions across a specified panel. That requires a defensible classification of search-like behavior. Drafting a letter and asking where to buy a replacement part should not be silently treated as equivalent searches.
Align units before comparing platforms
One person can make several visits, and one visit can contain several questions. A long conversation may mix research, writing, and personal assistance. Comparing message counts on one service with visits on another produces a ratio whose interpretation depends more on instrumentation than behavior.
Choose a common unit where possible and disclose the compromise where it is not. Define whether a follow-up question starts a new event, whether embedded assistants count, and whether mobile applications are observed. Excluding an environment is acceptable; leaving the exclusion unstated is misleading.
Use a transparent worked example
Imagine a deliberately synthetic worksheet with forty assistant-attributed referral visits and one hundred sixty conventional-search referral visits. Dividing forty by two hundred gives twenty percent of those two observed referral categories. These numbers are invented solely to explain arithmetic and are not market evidence.
- Write the included channels and unit beside the calculation.
- Keep unknown-origin visits separate instead of assigning them without evidence.
- Repeat the calculation under an explicitly different scope to reveal sensitivity.
- Label synthetic examples so they cannot be mistaken for observed statistics.
If another hundred eligible referral visits are added to the denominator without changing the numerator, the share changes. That does not mean assistant usage fell. It means the comparison set changed. A useful chart makes that change visible rather than narrating it as a behavioral trend.
Inspect the collection blind spots
Referral analytics cannot see every answer read inside an assistant, and an answer without a click produces no website visit. Missing referrers can obscure arrivals. A browser panel may miss application activity, while a survey can depend on recall and wording. Each method has a different blind spot.
Look for sampling details, collection environment, weighting, geographic coverage, and bot handling before accepting an estimate. If those details are absent, narrow the conclusion accordingly. Do not repair missing methodology with an apparently precise percentage or an unsupported confidence interval.
Publish evidence another person can verify
A publishable statistic needs a retrievable source, the exact definition, the observation period, and enough calculation detail to reproduce the reported value. Preserve the original table or authorized export and document transformations. A citation to a headline alone rarely establishes how the number was produced.
When reporting your own sample, include exclusions and an honest label such as observed referral share. Never replace missing observations with invented statistics to make the story feel complete. If privacy prevents sharing raw records, provide aggregated workings and describe the resulting verification limit.
Match the conclusion to the evidence
Use a narrow referral estimate to plan landing-page work, not to declare the winner of the entire search market. Compare trends only when definitions and collection methods remain compatible. When they change, mark a break in the series and explain what can still be learned.