Brand positioning for AI agents: who writes your brand?

Where a shopping agent reads your store, brand positioning is not a text but a missing field. We queried nine Dutch Shopify brands at the public address where their store answers AI agents. Six different brand questions produced exactly the same block of text at all nine stores. Four questions that actually matter commercially, such as "what is your premium line" and "why you and not the other one", produced nothing at all at all nine stores. And the brand story the agent does read out was not written by those brands.

There is a way out, and that is why this article exists. Since the end of May, every Shopify store can override the document agents read first with a template in the theme. That is in Shopify's own changelog of 28 May. Of the nine brands we measured, zero had put their own version in place.

This piece shows what an agent says about a Dutch brand today, where that text comes from, and which three layers you can fill in yourself. We do this work from the listing side: six years of building product pages for brands on marketplaces and in their own webshops, and recently a second reader that reads the same page very differently from a human.

What does a shopping agent actually ask your store?

Not your shop window. An agent working on behalf of a buyer knocks on a separate address. On Shopify that is /api/mcp on your own domain, with a second, newer variant at /api/ucp/mcp. Shopify openly documents those standard storefront tools: search the catalogue, update a cart, and search policies and frequently asked questions.

That last one is the brand layer. When a buyer asks what this brand stands for, that question does not go to your about page but to that address. The answer comes back as plain text, without styling, without imagery, without your typography. Everything a brand identity normally leans on has been stripped off. What remains is the content of the sentence.

This is not theory and not a beta. You can reproduce it yourself with one command on your own domain, and the answer you get back is the same answer a shopping agent reads out to a customer. That makes this layer different from the rest of the AI story: there is nothing to predict, you can simply read it off.

We described the shift itself earlier in agentic commerce: how AI agents buy products via Shopify and in agentic search and your webshop traffic. This article goes one layer deeper, to the question left open there: if the agent says something about your brand, who came up with that?

Nine brands, eighty-one brand questions, three answers

We took nine Dutch Shopify brands from coffee, fashion, lingerie, underwear and household goods. We asked each store nine brand-focused questions first, and then eight more. All in English, because that is the language in which this layer answers reliably.

The result is easier to summarise than expected. Across all brand questions together there are exactly three unique generated blocks, and those three have to cover nine brands.

Six brand questions, one answer. Two operational questions, differentiated answers. That contrast is the entire finding. The platform has pulled your logistics apart into separate, usable blocks and compressed your brand into a single paragraph that gets served up again for every question.

For a brand that spent years building a story, that is a painful outcome. For a brand that has put nothing in place yet, it is a draw. Precisely the wrong way round.

The four questions you get selected on return nothing

There is a second set of questions that weighs more heavily, because a buyer asks them at the moment of choosing. We asked four, at all nine stores:

• Which line is your premium line?
• What is your most exclusive collection?
• How do you position yourselves against cheaper alternatives?
• Why should I choose this brand and not another?

Result: zero answers. Not one store out of nine returned anything on any of these four questions. No empty block, no referral, no "that is on our site". An empty list.

These are the four questions a brand agency normally builds an entire engagement on. They are about ranking within your own assortment and about your place next to a cheaper supplier. In the agent layer they do not exist. An agent firing these four questions on behalf of a buyer gets nothing at all nine brands and then goes on to compare on what is there: price, availability and delivery time.

That is how a conversation about brand choice falls back to price comparison. That is no longer a technical detail but a commercial problem, and it is the same problem we described in why cheap is not enough: as soon as the difference is not legible, the choice falls back to the one axis that is.

Nobody at your company wrote the brand story the agent reads out

The block that does come back reads well. Description, about us, target audience, categories, popular search terms. At a coffee roaster it carries the founding year, the in-house roastery, the chain from sourcing to packaging. Factually it is correct.

Except nobody at that company wrote or approved that text. That is measurable, and we measured it. We cut the generated profiles of four brands into fragments of five consecutive words and searched for those fragments in the full text of their own websites. Four to six hundred thousand characters per site, every page in the sitemap.

Of the 625 word groups we checked this way, zero came literally from the merchant. Zero out of 625. The facts are derived from about, team and footer pages, but not one formulated sentence is the brand's own. The voice is entirely the platform's, down to lines like "looking for toilet paper that's good for your bum and the planet".

Where it goes wrong is where the source text is thin. At one brand, a second name appeared under the heading "Owner/Founder" that does not belong there. The only public person with that name is the founder of a British lingerie brand with no connection to this store. At another brand, the parent company from the footer was presented as the founder.

That is not minor noise. A brand rebuilding trust after a difficult period gets an invented co-founder read out to every buyer who asks. We do not name the brands here, because this is not a reckoning and the error is reproducible with one command on your own domain.

Shopify's own app for managing these answers sits at an average of 3.2 points in its own app store, with a telling verb in that app's description: you can view the generated questions and answers, and add your own alongside them. Editing is not on the list.

There is a free button, and nobody presses it

Now the part that is in your hands. Every Shopify store serves a file at /agents.md, and that file literally describes itself as "the canonical, agent-facing description of the store". So it is the document an agent reads first to understand who it is dealing with.

We fetched that file at all nine brands and laid the versions side by side. Every file is eighty-six lines long. At eight of the nine brands, exactly eight lines differ from the rest, and those eight contain only the brand name and its own web addresses. At the ninth brand it is thirteen. Beyond that the file is identical down to the space.

More than that: of those eighty-six lines, nine are about Shopify's own checkout button and about why an agent should use it. About your brand, there is your name. Nothing else.

And it does not have to be that way. The changelog of 28 May says it literally: put templates/agents.md.liquid under Online Store > Themes > Edit code, and if no template is present the system falls back on the default file Shopify generates. That is a free, documented entry point into the first thing an agent reads about your brand. It has been open for more than two months.

Of the nine Dutch brands we measured, zero had put their own version in place. That is not a reproach to those brands. There is no supplier billing you for this and no dashboard turning red, and that is exactly why this kind of thing sits untouched for years.

One thing we keep honest: whether your own agents.md also overrides the generated brand profile block is something we cannot establish from the outside. We test that on our own store before offering it as a service. What is certain is that the first document an agent reads could be yours, and that right now it is nowhere.

What you do steer today: tags, collections and your about page

The product layer gives you little. We ran the catalogue query again on a Dutch coffee roaster and got nine fields back per product: id, title, description, url, price range, variants, options, media and categories. The newer address returns fourteen, adding collections and gift card among others. Neither contains a field for brand, manufacturer or vendor.

The only text fields in that product layer you control are your tags and your collections. That changes their status. A collection name was until recently a navigation choice in your menu. Now it is one of the few places where you determine how your assortment arrives at an agent. "Sale" and "New" say nothing there. "Hand-roasted single origin" does.

That produces a short list of what you steer today:

The bottom three rows are simply work you already know. The difference is that your about text is no longer an atmosphere page but raw material. What is there gets picked up. What is not there gets filled in by a system that does not know your brand.

That makes brand copy verifiable for the first time in a way that goes beyond taste. You write the text, you ask the question again, and you read back what the machine made of it. Before and after, in the same week. Anyone defining brand values now has a measurement point instead of a conviction.

Why bol and Amazon do precisely the opposite

There is another reason to separate this per channel, and it is sharper than expected.

At the same moment, from the same machine, we measured five Dutch Shopify stores. All five returned an error on an ordinary request for their robots.txt, because a bot wall sits in front of it. All five answered the brand profile at /api/mcp at that same moment without any barrier. No login, no check, no limit. The address serving your brand story and your owner's name is more open than your own store.

On the marketplace side this simply does not exist. bol returns an error on every agent address we tried. Amazon does the opposite of Shopify: the robots.txt of amazon.nl contains 101 lines with a user-agent, and the shopping agents of OpenAI, Anthropic, Google and Perplexity are each fully excluded there.

Two platforms, two opposite choices, the same client. Shopify pushes the merchant's brand story outward and fills it in itself if the merchant does not. Amazon shuts the agents out and keeps the brand contact to itself. Anyone running both channels therefore needs not one brand strategy but two, and that is exactly the dividing line between listing work and brand strategy for your webshop.

What brand positioning for AI agents means for a branding engagement

A branding engagement used to deliver a positioning, a visual identity and a brand book. That stays. What is added is a layer that has nothing to do with styling: the text a machine returns about your brand to a buyer who does not know you yet.

In practice that shifts the order of the work. Positioning remains the first step, because without a sharp positioning you have nothing to put in any layer. See brand identity and revenue in e-commerce and, for the phasing, building an e-commerce brand. But the deliverable gains a component: the text that has to be machine readable, in the language that layer answers in.

We approach that the way we approach any branding engagement, only with an extra checkpoint. First the command on the client's domain, live, while they watch. Then advice. At one brand there is a tidy profile with a correct founding year. At another there is a name that appears nowhere on the site. That saves half the conversation, and it is the same approach we take at elvou on the listing side: first show what is there now, then talk about what should be there.

What we do not do is promise that this gets you into ChatGPT. Nobody can. What we do is make sure the answer an agent already gives is correct and yours.

What you do this week

No engagement needed. Four checks you can run in a morning, with an outcome you can show your team.

1. Read back your brand profile. Send one request to yourdomain.com/api/mcp asking for the story behind your brand. Is the description right, is the founder right, is the audience right?
2. Open yourdomain.com/agents.md. Is there anything of yours in it, or only your name and your policy pages? If the second: that is the file you can build your own template for.
3. Ask the four choice questions. Premium line, most exclusive collection, position against cheaper suppliers, why you. If you get nothing back, that is exactly the gap.
4. Look at your collection names. Do they describe a property or a promotion? Anything called "sale" or "new" is empty to an agent.

The order is not arbitrary. The first two let you read back what is being said about you now, and step three shows where that story stops. Only the fourth is something you can change this afternoon without calling anyone.

After that it is simply craft. About and policy pages that are complete and factual, in English too. Product titles with brand and type up front. Collections that name your assortment instead of your promotions. And a webshop in which those texts sit where the platform reads them out, not only on the page where your customer sees them. Anyone already taking a structured approach to webshop SEO has half of this work in place.

Curious what an agent makes of your store? Book a call and we will run the command together, on your domain, while you watch.