For most of the internet era, shopping has started the same way: with a product in mind. You know, or think you know, what you need. You type a keyword into a search box, open a million tabs, compare options, and slowly work your way toward a decision. Research like this has taken us hours. For years. But now, AI is changing that. And on the surface, it might seem like it’s just making the old process faster. It is making things much more efficient for the shopper, but it’s also creating a brand new behavior.
What makes this shift so important is that people do not naturally think in terms of products. They think in goals, constraints, and life context. They think,
“I’m trying to survive a red-eye flight and still look alive for a morning meeting.”
“My toddler refuses to eat vegetables and I’m running out of tricks.”
“I want our apartment to feel cozy before winter hits.”
“I want to feel confident walking into this job interview.”
“I need something that will help me focus while studying for finals.”
“My dog gets anxious every time we leave the house.”
This way of thinking has always existed but, in a search-driven world, we have had to manually translate these needs into a series of product searches. However in an AI-driven world, where a conversation is happening, the system can translate the goal, ask follow-up questions, and decompose that goal into actual product recommendations.
That moves commerce from the old world where shoppers searched for products and demand was captured, to a new world where demand is created for the products AI recommends. These could be recommendations for products the shopper didn’t even know they needed.
That framing matters because it helps explain why so many people misunderstand what AI recommendation actually is. Many assume the model simply takes key words from the user prompt and then gives a list of products that talk about those keywords. AI shopping is not keyword ranking. It is a recommendation system built on semantics, constraints, and authority.
Using my real world picky toddler example: “My toddler refuses to eat vegetables and I’m running out of tricks.”, AI isn’t taking the “vegetables” keyword and then recommending a cookbook. It will follow up and ask things like:
By the end of our conversation, the agent has a much richer understanding of the situation than a keyword search ever could. And then it can go out and do the proper research to build a solution set. Things like kid-friendly recipes, hidden-vegetable meal ideas, toddler plates and utensils that make eating fun, cookbooks for picky eaters, or even strategies pediatric nutritionists recommend.
I didn’t have to come up with the idea to search for “vegetable recipes for toddlers,” “toddler plates,” and “picky eater tips” and then do the research. All of this got compressed into one conversation and AI gave me recommendations I never would have thought of.
This evolution underpins one of the biggest mental shifts brands need to make to succeed in agentic shopping. AI does not retrieve brands. It retrieves solutions. The unit of recommendation is not your brand awareness or prestige. It is the product-level match.
Large brands can actually be disadvantaged here because their brand representation is broad and diffuse across many product types and categories. The model is not looking for a halo. It is looking for the right SKU.
So how does AI actually decide? Getting your product recommended by AI requires passing two stages. In the first stage, the model has to decide which products even belong in the room. This is called building the consideration set and it can contain a long list of options.
In this stage the model narrows into a product category neighborhood, applies the shopper’s constraints, and filters for attribute-level fit. If someone asks for the best shampoo for a sensitive scalp under a certain budget, the model is not starting with every brand in the market and sorting from one to one million. It is first figuring out what category this is really about, what constraints matter, and which products have the right attributes to satisfy the task. That means category fit, context fit, and attribute completeness matter enormously. Is the product clearly represented as belonging in the right category? Does it satisfy the shopper’s stated or implied needs? Does the underlying data actually include the ingredients, claims, exclusions, sizing, materials, or compatibility details that the model needs to make the match? If not, the product is filtered out. Brand prestige does not save it.
In the second stage, the model has to decide which of those products selected into the consideration set deserve to rise to the top. This is the ranking stage. And the key is to be ranked in at least the top 10 because AI is typically only recommending the top 3 to 8 products. In this stage the model starts asking a different question: which of these options is most trustworthy, most substantiated, and most likely to satisfy the user? This is where trust signals come in. Certifications, third-party testing, credible reviews, media mentions, consistent product data across brand sites, retailer feeds, and marketplaces. Repeated corroboration from authoritative sources.
Most brands obsess over this second stage and completely miss the first. That is a mistake because no matter how authoritative your product is, if it’s filtered out of the consideration set it won’t be recommended. In most cases, products aren’t being left out because they are outranked. They are left out because they were never considered relevant in the first place.
Another important dynamic happening under the hood is how the model actually conducts research in these two stages. When a consumer enters a prompt, the LLM does not simply run a single search and return the top results. Instead, it initiates a process often referred to as fan-out querying. The model decomposes the conversation into many smaller research questions and launches a large set of queries across the web and structured data sources. It then synthesizes what it finds, evaluates whether the information is sufficient, and often launches another round of queries to fill gaps or validate claims. Each round of these fan-out queries can look quite different from the original prompt and from one another, because the model is translating the user’s intent into the many signals it needs to evaluate products. Some queries are looking for relevant product attributes that satisfy the user’s constraints (as part of the selection stage), while others are looking for authority signals such as certifications, expert reviews, or credible third-party validation (as part of the ranking stage).
This is also why focusing only on prompt optimization is a weak strategy. A shopper can express the same need in countless ways:
“warm jacket”
“insulated coat for commuting”
“winter layer that packs small”
And while these queries may look like different prompts to a human, they collapse into the same underlying intent. Winning recommendation performance is not about chasing an infinite long tail of prompt phrasing. It is about aligning to the stable semantic architecture underneath language. What matters far more is whether the underlying product data and trust signals exist across the web in a way that those fan-out queries that the model itself is generating as it conducts research, can reliably find and verify your data.
The right way to understand AI recommendation is not the old SEO-lens of “did I appear for this prompt?” It is “why was I filtered out, and what would move me into eligibility?” That is a much more useful question. If a model cannot confidently place your product in the right category neighborhood, if it cannot map your product to the shopper’s constraints, or if it cannot find structured evidence to support your claims, you are effectively invisible.
The model needs explicit signals about what a product is, who it is for, what constraints it satisfies, and what evidence backs that up. If you want to be recommended for a weekend camping trip, “20,000 BTU stove” is not enough. The data needs to tell the machine that it works for car camping, serves two to four people, and boils water quickly. The same is true across beauty, food, home, wellness, and beyond. The brands that invest now in structured, verified product data will not just show up more often. They will be the ones AI trusts to recommend.
Which brings me to ChatGPT ads. The lazy comparison is to Google Ads. Conversational ads are structurally different. OpenAI says ChatGPT answers remain independent and unbiased, conversations stay private, and ads are selected separately based on relevance to what is being discussed in the current chat. If a user chooses personalized ads, OpenAI may also use signals like past chats and ad interactions to improve relevance over time. Advertisers do not get to alter the answer itself. They are competing for relevance in a separate system layered around the conversation. TNW’s early advertiser analysis also points to a world where success depends less on classic keyword bidding and more on understanding conversational intent and the AI’s relevance logic (The Next Web).
So the problem starts to look very familiar: relevance is a data and ranking problem. The brands that win will be the ones with product and claims data structured in a way that makes both the recommendation system and the ad system understand, with confidence, when that product belongs in the conversation. In other words, the future of ChatGPT ads is not about gaming the model. It is about making your product legible to the systems around it. The same work required to help a product get selected and recommended is increasingly the work required to help an ad get selected and shown. That is why I do not think the real story is “ads are coming to ChatGPT.” The real story is that AI commerce is forcing brands to build better data infrastructure. And the ones who do will have an advantage everywhere these systems mediate choice.
