Picture a shopper who found you through ChatGPT. They asked for a carry-on that fits Ryanair’s cabin rules, got three suggestions, and one of them was yours. They click through and land on your product page. They like what they see: the finish, the price, maybe the brand itself.

Now they have questions. Does it also pass on easyJet? Is the four-wheel version heavier? What does the warranty actually cover? Is there a matching bag for the rest of the family?

Where do they go? They could open ChatGPT again and ask there: about your product, on someone else’s surface, in someone else’s words. On your site, there’s nowhere to ask. There’s a menu, an FAQ page buried in the footer, maybe a contact form that answers in 48 hours. And a search bar that was never built for questions: it expects a product name and returns a grid of thumbnails, the same interaction as in 2012.

That’s the paradox of the moment. The shopper who arrives from an assistant has just had a conversation. Then they walk into your store and there’s no one to talk to. The most frustrating step of their journey is now the one you own.

Our last piece argued that the middle of the funnel, interest and consideration, is the one part a brand can still own. This one is about what owning it actually takes, starting with the most basic job of all: helping people find.

ChatGPT got you discovered. Showing them around is your job.

Let’s give the assistant its due. It did exactly what it was supposed to do: it put you in front of someone who may not have known you yet, or only by name. That’s discovery in the sense we used in our last two pieces, being discovered, and it increasingly happens somewhere you don’t own.

And those visitors are worth a lot. According to Adobe, traffic from AI assistants to US retail sites converted 54% better than other traffic in May 2026, a complete reversal from a year earlier. Once on site, these shoppers spend 53% longer per visit and browse 23% more pages. They arrive pre-qualified, curious, and ready to dig in. [1]

What happens next is a different job: making them discover you. Your range, your expertise, what makes you different from the two other names the assistant listed. In a store, that’s what a good associate does in the first five minutes. Online, the shopper who has just spent twenty minutes asking ChatGPT anything expects to keep asking anything.

That expectation isn’t new, by the way. In Baymard’s usability testing, roughly half of participants turn to search as their preferred way to find products, and 34% of participants in one study typed non-product questions into it: return policy, cancelling an order. On two-thirds of sites, those questions get a grid of products back. [2]

So the shopper wants to ask anything, like in ChatGPT. You want the answer to come from you: your words, your catalog, your rules. That’s the whole brief.

Then why not let ChatGPT answer for you? In a sense it already does: that’s how the shopper found you. But ChatGPT is a generalist assistant, not an e-commerce product: shopping is one use case among many, next to homework, writing and code. In OpenAI’s own study of 2025 usage, questions about products made up about 2% of ChatGPT messages, less than tutoring (10%) or programming (4%). [5] It isn’t built to be a sales associate for any store, let alone yours. It sees your catalog through a product feed, not through the attributes, content and rules your team built, and it will always answer for its own user rather than for your brand.

There’s also what the exchange itself contains. The way shoppers describe their needs, the constraints they add, how they answer the clarifying question: that is the richest intent data a brand will ever collect. When the conversation happens on the assistant’s surface, that data stays with the assistant. When it happens on your site, it’s yours, to sharpen the answers, fix gaps in your catalog, and learn what people want that you don’t sell yet. That’s the part of the exchange you need to own.

Keep the search bar. Stop asking it to understand.

Let’s be clear: lexical search isn’t going anywhere. When someone types a product name or a reference, keyword matching is fast, cheap and exactly right. Amazon, whose keyword search is the benchmark, drew the same line: since May 2026 its search bar also takes plain-language questions, answered by Alexa for Shopping, the assistant that replaced Rufus. [6]

The problem is everything else. Baymard’s 2026 benchmark rates 56% of sites “mediocre or worse” on search, and the failures cluster where shoppers describe rather than name: a use case, a compatibility need, a problem to solve. The most telling example is a Home Depot search combining a refrigerator’s style and capacity. It returns zero results, while the site’s own filters, applied by hand, find 45 matching products. [2] The catalog knew the answer. The search bar couldn’t ask it the question.

That is the brand agent’s first job: turn a vague intent into a structured query, wherever it can. Not to replace your catalog structure, but to use it. Your categories, attributes and filters are years of merchandising work. The agent’s job is to translate a sentence into them.

Take our carry-on shopper. “Something that fits Ryanair” holds a constraint no keyword will ever match. Ryanair allows 55×40×20 cm with Priority, and 40×30×20 cm for the free bag under the seat. An agent that knows your catalog turns the sentence into a query: cabin luggage, height ≤ 55 cm, depth ≤ 20 cm. And because the sentence is ambiguous, it does what a keyword engine never could. It asks one question first: with Priority, or the free under-seat bag?

Two things make that translation work. The agent has to understand language, which large models now do well. And it has to know how your catalog is organized: which attributes exist, what they’re called, which values they take, which ones you can trust. That second part is the hard one, and it’s different for every brand.

Working this way also keeps the answer honest. A structured query can be shown back to the shopper as filters they can see and adjust. The shopper sees how they were understood, and your merchandising stays the backbone of the answer.

What doesn’t fit in a filter

Structure only takes you so far. A lot of what shoppers ask has no column in your product database.

Take a skincare shopper at a retailer like Oh My Cream!: “I need a moisturizer that won’t pill under my foundation. My skin is oily by noon but tight after cleansing.” Part of that is structured. Combination skin, dehydration: Oh My Cream! product pages already carry skin types and skin concerns as fields, and the agent should turn them into filters. [3] But “won’t pill under foundation” isn’t an attribute anywhere. It lives in texture descriptions, in the team’s own notes on each product, in usage tips, in reviews.

The same goes for the questions from our opening. What the warranty covers, where the products are made, whether the brand repairs or replaces: none of that is in the catalog. It sits in FAQ pages, brand stories, buying guides and care instructions, the content your team has spent years writing.

So the agent’s second job is to leverage everything unstructured you can give it, and to use it well. Semantic retrieval gets you part of the way: it matches meaning instead of strings, so “won’t pill” can find “leaves no residue”. But retrieval alone just returns paragraphs. The agent has to combine them with the structured query from the previous step: filter on what’s certain, then rank what remains on what the content actually says.

There’s an irony here. Brands are rewriting that same content so ChatGPT can read it. Adobe ranks cosmetics first for AI readability, precisely because of dense ingredient lists, tutorials and how-to guides. [1] If your content is good enough to be read by someone else’s assistant, it’s good enough to power your own.

One rule holds throughout: the agent only says what the content supports. If no page says a cream works under makeup, the agent doesn’t claim it. That’s the responsibility layer we described in our last piece, applied to discovery. [4]

Your rules, not the assistant’s

Understanding the shopper and knowing the catalog still isn’t enough. Two products can answer the same need equally well, and you rarely want to be indifferent between them. That’s the third job: apply your own business rules to what gets shown.

Some rules are simply written down. Push the new collection this month. Hide products under a stock threshold. And for a multi-brand retailer with its own label, prefer the house brand when it’s a genuine match. Think of a beauty concept store selling thirty brands next to its own skincare line: when two moisturizers fit equally well, it wants its own shown first. A generic assistant will never know that preference exists. Your agent should apply it on every answer, visibly or not, and never at the expense of fit.

Other rules are better learned than written. Which products shoppers actually buy together. When suggesting a second item helps and when it annoys. Which pairing converts for a first-time visitor but not for a loyal one. Nobody writes those down, because nobody knows them in advance. An agent that sits in thousands of conversations can learn them, as long as those conversations stay yours.

That last part doesn’t come for free. Writing a rule takes a line of configuration. Learning one takes infrastructure: capturing every conversation and what followed it (the click, the add-to-cart, the purchase, the return), joining it with catalog and order data, and running the models that turn all of it into better recommendations. That’s the classic recommendation toolbox, collaborative filtering and propensity models, now fed with what shoppers say rather than only what they click. A model provider can ship a smarter language model tomorrow. It can’t ship that learning loop, trained on your shoppers. That’s what compounds over time, and that’s what stays hard to copy.

Put the three jobs together and you get something a search bar was never designed to be: a guided tour of your brand, in the words you chose. The shopper asks anything, like they would in ChatGPT. The answer comes from your catalog, your content and your rules.

That’s the foundation. Once the shopper has found the right product, a new question opens up: what else should you suggest, and when should you stay quiet? Cross-sell and upsell have existed forever, and they’ve almost always been clumsy. That’s the subject of our next piece.

Sources