Who Actually Shops on ChatGPT

Purchase research used to live in tabs. A buyer comparing two skincare serums in 2019 opened the brand site, two review aggregators, a Reddit thread, maybe an Amazon listing, and walked away forty minutes later with a verdict. The same buyer in 2026 opens one conversation, asks a question, gets an answer, and makes the decision in under five minutes. The tabs are gone. The forty minutes are gone. The verdict is the same kind of verdict, arrived at through a structurally different process.

If you are running paid acquisition today, the obvious next question is who this buyer is, what they want, and how to reach them. We will answer that as honestly as it is answerable, which is partially. Some of the demographic and behavior data on ChatGPT shoppers is publicly available and worth knowing. A lot of it is not yet legible to anyone outside the platform. This post separates what we can say with confidence from what is still a hypothesis, and offers three archetypes worth holding as working models while the population resolves.

The two behaviors that matter

Demographics are the wrong starting point. The ChatGPT user base is not a homogeneous segment, and trying to summarize the population in a sentence about age, income, or geography hides more than it reveals. The two behaviors that actually predict commercial intent are research consolidation and conversational replacement.

Research consolidation is the substitution of a multi-tab comparison with a single-question summary. A buyer who would have spent forty minutes across eight sources now spends five minutes across one. The substitution rate is highest in categories where the buyer feels uncertain and where the cost of asking the wrong question is low. Skincare, supplements, software, consumer electronics under three hundred dollars, anything in the 'I want a recommendation, not a research project' lane.

Conversational replacement is the structurally different one. It is the buyer who never opens a browser tab in the first place. The product question is refined inside the conversation; the actual transaction completes on the merchant site after the buyer clicks through. There is no search query, no organic listing, no comparison page. This is the behavior that breaks legacy analytics, because the conversion path begins inside a surface with no UTM and lands on a merchant page that looks like direct traffic unless a Conversions API pixel or proper UTM plumbing is in place. We covered the broader attribution problem in the $400B dark funnel piece.

Both behaviors are present in the same user base; some buyers exhibit one, some both. The mix varies by category and by buyer mode. The point of separating them is that an ad creative that works for research consolidation often does not work for conversational replacement, and vice versa. The user is the same; the moment is different.

What we know from public sources

By OpenAI's own disclosures alongside the ad pilot launch, ChatGPT had seven hundred million weekly active users and roughly fifty million shopping queries a day passing through the conversation. Independent industry research from Idea Grove in 2026 reported that thirty-eight percent of generative AI shoppers use ChatGPT specifically for product ideas. Those are the load-bearing numbers; everything else in this section is interpretation.

Beyond raw users, the public commentary from OpenAI and from independent reporting points to several consistent traits. ChatGPT skews younger than the average internet user but not dramatically; the median ChatGPT user is closer to the median Google user than to the median TikTok user. Usage extends across consumer and professional contexts, with consumer commerce queries emerging as a growing share of conversation volume. The engagement pattern is multi-turn rather than single-query, which means a typical commerce question is the start of a conversation, not a one-shot lookup.

Independent surveys of agent shopping behavior, including work we did ourselves in the 10,000 agent conversations analysis, point at a few stable patterns. Reviews carry disproportionately heavy weight in model recommendations. Brand authority signals matter more than ad-channel signals. Price sensitivity is real but ranks below trust signals in the recommendation calculus. These are observations, not laws; they reflect the model behavior we have measured to date and will continue to shift as the platforms iterate.

Where reliable third-party data ends, we step into the territory marked 'early signals.' That territory is large, and being honest about its size is more useful than pretending the data is firmer than it is.

What we do not yet know

Four open questions matter for anyone trying to make budget decisions today.

The first is purchase conversion rate. We do not have public data on what fraction of commerce-intent ChatGPT conversations end in a purchase, and the brands that do have this data internally are not sharing it. The closest analog is early Amazon Sponsored Products, where conversion rates were three to five times higher than competing surfaces, but that analog is suggestive, not proof.

The second is platform stickiness in commerce. ChatGPT today is the dominant AI conversational surface by user base, but the category is competitive and the user behavior of asking a model for a recommendation will outlive whichever specific platform leads. A brand betting on the channel needs to know whether they are betting on a platform or on a behavior. Our reading is that it is the behavior, but reasonable people disagree.

The third is the durability of the trust signal. Models today rely heavily on review density and editorial authority. As more brands optimize for these signals, the rank-ordering will compress, and the signals that differentiate winners will shift. The current playbook for being recommendable will not be the playbook in three years.

The fourth is measurement. Until there is a standard way to attribute revenue back to specific AI conversations, brands will under-credit the channel and over-credit the channels that get the last click. This will correct, but the correction will be uneven across analytics vendors and may take more than a year to settle.

Three buyer archetypes worth holding as hypotheses

Treat these as working models, not as customer personas. They are useful for sharpening creative and measurement decisions; they are not yet rigorous enough to drive segmentation.

The Researcher. This buyer would have opened tabs. They are using ChatGPT to compress the comparison. They want options, tradeoffs, and a recommendation, with a brief explanation of why. The ad surface they respond to is one that earns the model's recommendation through review density and clear product positioning. The conversion path is short but informed, and the buyer is willing to read for a paragraph if the paragraph helps them decide.

The Replacement Shopper. This buyer has a category they have already bought in. They know what they want. The conversation is one or two turns, the ad surface is a recommendation that fits their stated constraints, and the click-through to your site happens fast. Repeat purchase categories are heavily weighted here: skincare refills, supplement subscriptions, replacement parts for hardware they already own. The structural opportunity is that the model can remember the prior purchase and surface the brand without competition.

The Discovery Shopper. This buyer is not looking for a specific product. They are exploring a category, often because someone mentioned it or because they encountered a problem they did not know had a product solution. The conversation surfaces options they would not have searched for. This is the archetype where brand authority and category presence pay the highest dividend, because the model has the most degrees of freedom in what it recommends and tends to anchor on the brands with the most signal density.

These archetypes are not mutually exclusive. A buyer can be a Researcher in one category, a Replacement Shopper in another, and a Discovery Shopper across the rest. The point is that an ad strategy optimized for one archetype will underperform on the others. Differentiate the creative, the offer, and the measurement window by archetype, not by demographic.

What this means for ad creative

The creative that works in a conversation looks structurally different from the creative that works in a feed. Three principles hold across the archetypes.

Before the principles, one finding worth holding through the next year of strategy work. Idea Grove's 2026 study of generative AI shoppers reported that roughly ninety-eight percent of buyers verify the model's recommendation before purchase. They Google the brand, read reviews, check the merchant site. The compressed conversion path inside ChatGPT does not mean buyers are skipping verification. It means the verification happens off-platform, after the recommendation, before checkout. The implication for creative strategy is sharp: your brand authority and review density outside the conversation matter more than any single ad unit inside it. The ad is the introduction; the brand has to survive the cross-check.

Specificity beats emotion. A Meta creative is rewarded for stopping the thumb and earning an attention beat; the conversation does not have a thumb to stop. The creative that performs is the one that contains a fact the model considers worth surfacing, ideally one that distinguishes the product from the next best option. 'Vitamin C serum, 15% concentration, formulated for sensitive skin, third-party reviewed at 4.7/5 across 12,000 verified buyers' is a creative the model can use. 'Glow, radiance, transformation' is not.

Survive paraphrase. The model will rephrase your positioning before it surfaces a recommendation. Creative that depends on a turn of phrase or a clever visual hook does not survive that translation. Creative that states what the product is, who it is for, and why it is the right choice in functional terms survives the paraphrase intact. Write copy that reads correctly when stripped to its sentence-level meaning.

Carry trust signals. The trust signals that matter for conversational ads are not the same as the trust signals that drive performance in a feed. Review density, expert endorsements, third-party verification, and structured product data all compound in this format. The brand that has invested in being verifiable has a creative advantage that does not show up in Meta but does show up here.

All three principles point at the same underlying observation: creative for the new surface is less about hooking attention and more about deserving recommendation. The brands that figure this out first will compound a creative-format advantage that takes their competitors years to copy.

The honest version of the ChatGPT shopper question is that the population is real, the behavior is meaningfully different, the data is incomplete, and the brands that learn fast will compound an advantage during the window when the data is still resolving. The way to learn fast is to be on the surface while it is still being shaped, not to wait until the demographics report drops. If you want to get on, the OpenAI Ads page shows the integration path, and you can start with Cresva for free.