Eight studies show information seeking, practical guidance and commercial research converging inside multi-turn conversations.
What are people actually doing when they search with AI?
The answer depends on whether the interaction is really search. Asking ChatGPT to rewrite an email, summarise meeting notes or generate an image is an AI interaction, but it has limited relevance to Answer Engine Optimisation. Asking for the safest midsize SUV for a family, the best CRM for a 40-person company or a comparison of three holiday destinations is different. The user is not only requesting information. The user is asking the system to help interpret criteria, compare options and move towards a conclusion.
Eight studies using platform data, representative surveys, experiments and commercial research point in the same broad direction. People use conversational AI to learn, solve problems, discover options, compare alternatives, narrow choices, request recommendations and check a decision. Those activities do not always occur in sequence, and they do not mean the AI makes the final choice. They do suggest that the commercially important unit in AI search may not be the initial query. It may be the <strong>decision context that develops across the conversation</strong>.
<h2>Real ChatGPT use is dominated by guidance, information and writing</h2>
One of the strongest behavioural datasets comes from <em>How People Use ChatGPT</em>, a September 2025 National Bureau of Economic Research working paper written by researchers affiliated with OpenAI, Harvard, Duke and NBER. The team used a privacy-preserving automated pipeline to classify random samples from logged-in ChatGPT Free, Plus and Pro consumer plans. Its principal classified-message window ran from May 2024 to June 2025, with some trends through July 2025.[1]
Practical Guidance, Seeking Information and Writing collectively accounted for nearly 80% of consumer conversations in the study. The paper also classified messages as Asking, Doing or Expressing. Across messages, 49% were Asking, 40% Doing and 11% Expressing. Asking meant seeking information or clarification to inform a decision; Doing meant requesting an output or task; Expressing meant sharing views or feelings without seeking information or action.[1]
This does not make 49% of ChatGPT messages an AEO opportunity. Asking includes non-commercial advice, education and many activities that do not involve brands. The finding establishes something narrower: information seeking and personalised guidance are central consumer uses of ChatGPT, and Asking was growing faster than Doing in the paper's observation period.
The study is also an OpenAI-linked NBER discussion paper rather than a peer-reviewed publication. It excludes enterprise, education, logged-out, under-18, deleted and opted-out conversations, and its categories were assigned by validated automated classifiers rather than human readers. Those boundaries matter, but the scale and behavioural design make it more informative than a small self-reported use-case poll.
<h2>Survey evidence shows AI being used to find information and work through choices</h2>
Pew Research Center surveyed 5,119 US adults from 17 to 23 February 2026. It found that 49% reported using AI chatbots and 42% of all US adults said they used chatbots to search for information. Information search was the most common chatbot use in Pew's multi-select question.[2]
McKinsey's August 2025 AI Discovery Survey used a representative panel of 1,927 US consumers and a broader definition of AI-powered search that included standalone assistants and Google AI Overviews. More than 70% of AI-powered-search users said they asked top-of-funnel questions about a category, brand, product or service. McKinsey's cross-training-shoe example then showed a possible progression: learn which attributes matter, identify brands associated with those attributes, consult reviews, introduce a budget or training goal and refine the recommendation.[3]
Yext's March 2026 survey of 1,120 US adults focused specifically on local-business search. It reported that 47% used AI for quick recommendations, 47% for thorough pre-decision research, 43% for category discovery and comparison, and 41% for deals and value. The categories could overlap, so they should not be added or treated as a distribution. Their value is the narrow range. No single use case dominated the survey.[4]
Together, these studies show why “AI search” is too broad to function as one intent category. A person can seek a fact, request advice, establish criteria, identify brands, compare trade-offs or ask for a preferred option. The commercial meaning changes even if every interaction looks like a prompt followed by an answer.
<h2>One conversation can contain several purposes</h2>
Consider a simplified CRM conversation. It might begin with “What is the difference between CRM and marketing automation?” That is learning. “Which platforms do both well?” introduces discovery. “Compare HubSpot with Salesforce” requests comparison. “We have 40 employees and no IT department” adds a decision criterion. “Given that, which two would you shortlist?” narrows the set. “Which would you choose?” asks for a recommendation. “What is the main reason I might regret that choice?” moves into validation.
That is one conversation, but it does not have one stable purpose. The answer can reveal criteria the person had not articulated at the beginning, and the next prompt can change the evidence required for a credible response.
Research on Naver Cue provides direct evidence that conversational follow-ups do more than repeat the first query. The study combined an in-lab evaluation with 72 South Korean users and 250 conversational tuples with 2,061 real-world Cue tuples from January and February 2024. Researchers developed a taxonomy of 18 follow-up patterns. Frequent actions included requesting more information, substituting conditions and adding or specifying conditions. Clarifying, excluding and substituting conditions correlated with lower satisfaction as rated by trained external evaluators.[5]
The study analysed consecutive query pairs rather than complete customer journeys, and its classifier achieved 73% accuracy. It does not prove that every conversation becomes a funnel. It does show that users react to answers, refine conditions and alter what should count as a satisfactory result.
<h2>Ordinary users do not necessarily begin with elaborate prompts</h2>
A March 2026 <em>Scientific Reports</em> study adds another boundary for AEO measurement. A nationally representative sample of 937 US adults completed a multi-turn information-seeking task with ChatGPT on assigned health, science or policy topics. Participants were required to interact for at least five turns, and researchers analysed 747 valid conversation URLs.[6]
Only 19.1% of the full recruited sample used at least one of the study's eight identifiable prompting strategies. Those strategies included providing context, requesting external references and specifying response style. ChatGPT's communication style also changed with topic controversy and aspects of the user prompts.[6]
This was a constrained experiment, primarily using GPT-4o, rather than natural commercial browsing. It does not prove how most shoppers prompt. It does challenge prompt libraries dominated by elaborate constructions. A real person may begin with “What is a good CRM for a small company?” and introduce constraints only after seeing the first answer.
That means repeated measurement should not focus only on carefully engineered single prompts. It should also test natural openings, follow-up questions and the way consideration changes as new conditions enter the conversation.
<h2>Recommendation is not the same as choice</h2>
The research consistently draws a line between assistance and surrendered authority.
Gartner surveyed 322 US consumers in January 2026. Willingness to let AI make purchase decisions topped out at 11% across lower-stakes categories such as personal care and household supplies. Willingness to let AI narrow choices was higher: 31% for household supplies and 28% for personal electronics.[7]
Product.ai's April 2026 survey of 1,463 US online shoppers found that 43% had used AI for product research in the prior 90 days. Among the 623 AI product-research users, 86% said they verified an AI recommendation through another source before buying. The subgroup split was 45% who always verified, 41% who sometimes verified and 14% who did not.[8]
Yext reported a similar local-search pattern. More than 93% of the AI-user subgroup took at least one verification step before acting. After an AI recommendation, 62% searched Google, 58% visited the business website and 52% clicked a source cited in the answer. Seven per cent said they acted without additional research.[4]
These are commercial, self-reported surveys, and the samples and shopping contexts differ. They should not be pooled. They do support a clear measurement boundary: <strong>Mention, Consideration and Recommendation can be classified from an AI response. Choice and behaviour require downstream human or business evidence.</strong>
<h2>A working taxonomy for AI-mediated decisions</h2>
Traditional informational, navigational, commercial and transactional intent families remain useful. Conversational AI adds a more fluid layer because one session can move among several purposes.
<div class="overflow-x-auto"><table><thead><tr><th>Zone</th><th>Conversational purpose</th><th>What the person is trying to do</th></tr></thead><tbody><tr><td>Understand</td><td>Learn</td><td>Build basic understanding or identify relevant criteria</td></tr><tr><td>Understand</td><td>Solve</td><td>Resolve a problem or determine a workable approach</td></tr><tr><td>Decide</td><td>Discover</td><td>Identify available categories, products, services or brands</td></tr><tr><td>Decide</td><td>Evaluate</td><td>Assess whether an option is credible or appropriate</td></tr><tr><td>Decide</td><td>Compare</td><td>Understand differences and trade-offs among alternatives</td></tr><tr><td>Decide</td><td>Narrow</td><td>Reduce the consideration set using personal constraints</td></tr><tr><td>Decide</td><td>Recommend</td><td>Ask the system to advocate for a preferred option</td></tr><tr><td>Verify or act</td><td>Validate</td><td>Challenge, corroborate or stress-test a provisional choice</td></tr><tr><td>Verify or act</td><td>Act</td><td>Obtain the information or pathway needed to execute a decision</td></tr></tbody></table></div>
This is an AEO Updates working taxonomy, not a sequence measured consistently across the eight studies. Not every conversation contains every purpose. Some conversations begin with comparison, stop after learning or return from validation to discovery. The framework is useful because it makes the evidence requirement explicit.
A Learn prompt may require a clear category explanation. A Compare prompt may depend on consistent specifications, independent testing and credible reviews. A Recommend prompt introduces a stronger preference threshold. A Validate prompt may surface negative evidence, exclusions or reasons not to choose the brand.
<h2>The information ecosystem becomes a verification loop</h2>
Websites, reviews, social content, expert sources and journalism remain relevant when a consumer asks AI for guidance. Those sources may contribute evidence to an answer. They also become destinations where the person checks what the answer said.
The loop can be described cautiously as: <strong>evidence may influence the answer; the answer may influence the person; the person may then verify against additional evidence.</strong> The studies do not reveal every source a model used, and a visible citation does not prove causality. They do show that post-answer verification is common in the surveyed shopping and local-search contexts.
For brands, the objective is therefore larger than generating a mention. The brand needs accurate representation, defensible associations and credible evidence that can survive changing criteria. It then needs to remain relevant in the consideration set, earn a recommendation where warranted and withstand verification on first-party and independent third-party surfaces.
This follows the evidence-environment model in the AEO Updates <a href="/what-is-aeo">What Is AEO? learning pillar</a>: first-party sources establish product and commercial truth; independent sources can provide experience, comparison and corroboration that a brand cannot credibly manufacture about itself.
<h2>Visibility should be segmented before it is weighted</h2>
Suppose an AEO programme monitors 1,000 prompts and a brand appears in 420. A 42% visibility rate is useful only when the prompt universe, AI system, observation conditions and denominator are declared. It does not reveal whether the brand appeared mainly in category definitions or remained present as constraints accumulated towards recommendation.
The first improvement is not an invented universal weighting formula. It is segmentation. Report Share of Visibility, representation, consideration and recommendation separately for Learn, Discover, Compare, Narrow and Recommend prompts. Preserve a platform-neutral view for diagnostic consistency, then connect the relevant stages to observable demand and downstream business outcomes.
A future decision-weighted measure might incorporate prompt purpose, decision proximity, commercial relevance and observed demand. Those weights are not established by the studies in this article. Until they are measured and validated, a weighted score should be treated as a declared planning model rather than an objective market metric.
The safer management chain is:
<strong>Observed demand → Visibility → Brand representation → Consideration → Recommendation → Verification or handoff → Business outcome</strong>
Each stage answers a different question. A brand can be visible but inaccurately represented, considered but not recommended, recommended but rejected during verification, or influential without receiving a directly attributable click.
<h2>AEO Updates Takeaway</h2>
People use AI extensively to find information, but the distinctive value of conversational search increasingly lies in what happens after the first answer. Users can introduce constraints, discover decision criteria, compare alternatives, narrow choices and ask for advice. The evidence also suggests that many consumers still retain final authority and verify recommendations elsewhere.
For marketers, the battle is therefore not simply to rank, appear or earn a citation. It is to ensure the brand is accurately represented, remains relevant as the question changes, enters the right consideration sets, earns preference when the evidence supports it and survives the verification that follows.
The practical measurement change is equally direct. Prompts should be organised by purpose and intent family, not treated as interchangeable rows in one visibility average. Recommendation should be separated from mention. Downstream choice should be measured with behavioural or business evidence rather than inferred from the answer.
AI is not yet replacing the decision journey. In many contexts, it is becoming a navigation and advice layer within it. That creates a different competitive environment: the brand does not merely need to be found. It needs to withstand the reasoning process the conversation makes visible.
<h3>References</h3>
[1] <a href="https://www.nber.org/system/files/working_papers/w34255/w34255.pdf" target="_blank" rel="noopener noreferrer">Aaron Chatterji et al., “How People Use ChatGPT”</a>, NBER Working Paper 34255, September 2025.
[2] <a href="https://www.pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-smart-devices-and-views-on-impact/" target="_blank" rel="noopener noreferrer">Pew Research Center, “Americans and AI 2026: Chatbots, Smart Devices and Views on Impact”</a>, 17 June 2026.
[3] <a href="https://www.mckinsey.com/capabilities/growth-marketing-and-sales/our-insights/new-front-door-to-the-internet-winning-in-the-age-of-ai-search" target="_blank" rel="noopener noreferrer">McKinsey & Company, “New front door to the internet: Winning in the age of AI search”</a>, 16 October 2025.
[4] <a href="https://www.yext.com/resources/consumer-search-behaviors-findings" target="_blank" rel="noopener noreferrer">Yext, “The AI search tipping point has already happened”</a>, 2026 Consumer Search Behaviors findings.
[5] <a href="https://arxiv.org/abs/2407.13166" target="_blank" rel="noopener noreferrer">Hyunwoo Kim et al., “Using LLMs to Investigate Correlations of Conversational Follow-up Queries with User Satisfaction”</a>, arXiv preprint accepted to LLM4Eval at SIGIR 2024, 18 July 2024.
[6] <a href="https://www.nature.com/articles/s41598-026-42465-4" target="_blank" rel="noopener noreferrer">Haoning Xue et al., “Users’ prompting strategies and ChatGPT’s contextual adaptation shape conversational information-seeking experiences”</a>, <em>Scientific Reports</em>, 5 March 2026.
[7] <a href="https://www.gartner.com/en/newsroom/press-releases/2026-05-27-gartner-survey-finds-consumers-want-ai-shopping-help-but-not-ai-purchase-decisions" target="_blank" rel="noopener noreferrer">Gartner, “Consumers want AI shopping help, but not AI purchase decisions”</a>, 27 May 2026.
[8] <a href="https://product.ai/research/trust-in-ai-commerce-report/" target="_blank" rel="noopener noreferrer">Product.ai, “The 2026 Trust in AI Commerce Report”</a>, 23 June 2026.