A buyer can use Google, ChatGPT and your website during the same decision. Counting which platform gained a visit tells you less than understanding which part of the decision moved. My starting point for AI search migration is the buyer’s task: what they need to understand, compare, verify or do.
Evidence map: compare like with like
| Source / design | Sample and period | What it measures | Interpretation limit |
|---|---|---|---|
| OpenAI | 1.5M conversations; May 2024–June 2025 | Message purpose inside ChatGPT | Not share of search |
| Pew | 900 US adults; March 2025 | AI summary incidence by query length | US panel; reconstructed results |
| Semrush adoption cohorts | 260B-row base; January 2024–June 2025 | Google sessions before/after adoption | Analyzed participant N not disclosed |
| Semrush clickstream | >1B US lines; October 2024–February 2026 | Search-enabled queries and outbound referrals | Global panel size is not US analytical N |
What the usage data actually measures
OpenAI’s September 2025 study analyzed 1.5 million consumer ChatGPT conversations from May 2024 through June 2025, using privacy-preserving automated classification. Its summary grouped messages into Asking (49%), Doing (40%) and Expressing (11%). These are message categories inside ChatGPT, not shares of web search or evidence that the same users abandoned Google. OpenAI, How people are using ChatGPT, September 15, 2025. Full working paper and methodology
That distinction changes the marketing question. An assistant may explain a category, draft a specification or help evaluate criteria. Only some of those tasks involve discovering a supplier. A large usage figure cannot tell your team which commercial question deserves the next page.
Search behavior also varies with the question
Pew Research Center examined 68,879 unique Google searches from 900 US adults’ browsing in March 2025. AI summaries appeared for 8% of one- or two-word searches and 53% of searches containing ten or more words. Results were reconstructed in April, so the study does not capture every result exactly as the participant originally saw it. It describes one US panel, not today’s Spanish search market. Pew Research Center, Google users and AI summaries, July 22, 2025
The useful observation is that the shape of the question matters. It is not evidence that long questions belong exclusively to ChatGPT. Google’s own documentation describes AI Mode as supporting reasoning and complex comparisons. A simple split between Google for transactions and assistants for research would miss that overlap. Google Search Central, AI features and your website
Build a task map before choosing a channel
I use four stages as a planning framework, not a measured law of buyer behavior. Understand: what problem is this and what alternatives exist? Compare: which criteria separate the options? Verify: what evidence, limitations and conditions support the choice? Act: what must happen to request a demo, book a service or complete a purchase?
For a hypothetical marketing team buying an analytics platform, understanding might require an explanation of incrementality. Comparison needs a clear distinction between attribution and experiments. Verification needs documented integrations, data requirements and credible studies. Action needs a precise implementation route. One generic category article will not complete all four tasks.
Map an existing page and an evidence owner to each task. Record where the answer is available, whether its conditions are clear and which question remains unresolved. Keep the platform field open until you collect evidence. This prevents a fashionable channel assumption from becoming the content brief.
Validate the map with three separate observations
First, inspect your own search queries, landing pages and qualified conversions by task. Classify what is visible; an analytics report does not reveal every question asked before a visit. Second, ask willing prospects where they researched and what helped them decide, preserving their words and allowing multiple sources. These accounts are useful but subject to recall and selection bias.
Third, run a fixed set of buyer questions across relevant search and assistant experiences. Save the wording, language, market, mode, date, answer and visible sources. This shows how those experiences answer the task under the recorded conditions. It does not establish how many buyers use them.
Look for agreement between these evidence streams. A missing comparison page matters more when prospects ask for the same criteria and current answers lack them. If only the model panel reveals a gap, treat that as a hypothesis for investigation, not a forecast of lost revenue.
Prioritize the task your team can improve
Choose a question that affects a real decision and has evidence you can responsibly publish. Improve its answer in an existing page where possible. Add selection criteria, relevant exceptions and the next action. Assign an owner from product, research or operations to verify claims before the editorial team rewrites them.
A useful decision record names the task, the evidence gap, the page, the owner and the review date. Success begins with a complete, accurate answer and an easier buyer decision. Qualified engagement and commercial outcomes require later measurement; neither is guaranteed by publication.
If your team already answers the material tasks clearly, preserve that work. Expand the map when customer evidence reveals another decision, rather than manufacturing a new article for every variation of a prompt.
The next move
Take one commercial decision and map its four tasks this week. Compare the questions with the evidence your site supplies. Fix the most consequential gap before reallocating budget on the assumption that an entire audience has changed platforms.
The research question: replacement or redistribution?
There are at least three hypotheses behind the phrase AI search migration. Substitution means the same people do fewer searches after adopting an assistant. Complementarity means they add an assistant while keeping search in their routine. Task redistribution means particular stages move even if total search usage stays stable. Each hypothesis needs a different observation. A market traffic chart alone cannot choose between them.
Semrush’s August 2025 study used a 260-billion-row US clickstream base, January 2024 to June 2025. It compared Google search sessions before and after first ChatGPT adoption in early 2025 and included a non-adopter control. The public summary reports no statistically significant change in daily Google search sessions. It does not disclose the analyzed cohort size or confidence intervals. Rows are not people. Semrush, 2025
That finding weighs against a simple story of total replacement in those cohorts. It cannot establish that individual tasks stayed on Google. Adoption was not randomized, so users who adopted ChatGPT may differ from those who did not. I would need task-level observations and a defensible comparison before making a claim about a specific buyer journey.
A separate Semrush analysis covers more than one billion US desktop and mobile clickstream lines from October 2024 to February 2026. In February, Google received 21.6% of ChatGPT’s outbound referrals and 34.5% of ChatGPT queries had search enabled. The global panel size of 200 million users is not the analyzed US sample size. Semrush, 2026
This is evidence of movement between services. It does not tell us why a person opened Google, whether a supplier was shortlisted or whether the visit converted. I use it to keep the journey map open: an answer can lead to another search, rather than end the research.
Read the difference without changing its meaning
In the Pew observations, 53% minus 8% is a 45-percentage-point difference in AI-summary incidence between the two query-length groups. The ratio is about 6.6 to one. Those are calculations from the published percentages, not new measurements. The denominator is searches within each length group, not all users or all buying decisions.
A long query may contain a comparison, a constraint or a complete question. Length is therefore a useful classification clue, but it is not an intent label. “What is attribution and how does it differ from incrementality?” is an educational question. “Which analytics platform supports our consent requirements and exports?” is a verification question. Both can be long.
For a commercial content brief, preserve the actual constraint. A page that explains the category but omits the buyer’s data residency requirement has not answered the decision. Adding more generic paragraphs would increase length without increasing usefulness.
A collection protocol your team can reproduce
Start with one product or service decision and a defined audience. Keep a question register with the verbatim wording, language, source type, date, task stage and the decision it affects. Remove personal identifiers from the working copy. Mark whether a question came from a prospect, an observed search query or the editorial team. These sources are not interchangeable evidence of demand.
Use a short codebook. Understand questions ask for meaning or alternatives. Compare questions ask for criteria or tradeoffs. Verify questions test a claim, condition or suitability. Act questions ask for the next operational step. Allow a primary and a secondary stage when needed. Record unresolved coding disagreements instead of forcing every question into a clean bucket.
Review a small shared subset with a second colleague before classifying the full register. Compare the labels and resolve ambiguous definitions. This is a proposed quality check, not an inter-rater reliability result from research I have conducted. Preserve the original wording alongside the label so future reviewers can challenge your interpretation.
Then triangulate. A query report provides observable search demand, buyer accounts provide recalled decision context, and a fixed answer panel shows what selected interfaces return. Only the first two provide evidence about your audience’s behavior, with their respective coverage limits. An answer panel supplies a diagnostic of answers under recorded conditions.
For each material question, write the expected answer components before inspecting your own page. For example: supported integrations, implementation conditions, limitations, evidence date and a practical next step. This reduces the temptation to declare the existing page complete simply because it mentions the topic.
Turn the research into an editorial decision
Prioritize a question when it affects a consequential buyer decision, the current answer has a specific gap, and a responsible owner can supply publishable evidence. Record those three judgments explicitly. An opaque score with invented weights will not make weak evidence stronger.
Keep a compact decision record: question, current page, missing evidence, evidence owner, proposed change and review date. For a comparison task, the output may be a criteria table with exceptions. For verification, it may be a documented implementation condition. For action, it may be a clear route to an assessment. The format follows the task.
At the next review, assess answer completeness and qualified engagement separately. A better answer can be established by checking the specified components. A commercial effect needs later outcome data. If the observed questions change, version the register rather than quietly swapping the panel and presenting an apparent improvement.
Sources and scope
Public-source analysis, checked October 12, 2026. Proposed frameworks and examples are not findings from an original study. US data is not extrapolated to Spain.
FA