Why do AI assistants recommend some brands and not others?
Flavia Barbat and Thomas Marzano in conversation. Legible and Lovable, episode 4. Published by Brandingmag, 2026-09-30. 63 minutes.
Episode notes by Thomas Marzano.
Because your customer’s AI has one overriding incentive: to keep its owner’s trust. It recommends the brands it can find and verify quickly, and among those, the ones it infers its owner will prefer. Legible to be found, lovable to be chosen, and the machine now judges both in the same query. That is the Legible-Lovable Law at work.
The coffee chain my AI couldn’t find
There is a coffee chain in Amsterdam, Coffeecompany, with sixteen branches in the city alone. The coffee is good, and I love the place. So I asked ChatGPT what time The Coffee Company closes. It had no idea what I was talking about. Which coffee company? There are many. It took several rounds of explaining that I meant a chain, and which one, before it understood. The name is a category descriptor, and to the machine, a brand I love did not exist until I spelled it out.
The opposite happened when I needed a receiver for my wireless microphone. I asked the same assistant for one that would work with my laptop. It found the microphone maker’s own receiver, and another at half the price. I ordered the cheaper one. It works perfectly well, and I couldn’t tell you the brand.
Those are the Law’s two zeroes. Legibility without lovability is a commodity: found by the system, forgotten by the person. Lovability without legibility is a ghost: cherished by those who find it, invisible to the systems through which most discovery now happens. On air, Flavia simply called it invisible.
Why does the machine behave like this? Because the number one incentive of your personal AI is that you can trust what it brings back. If it doesn’t work for you, you stop using it. So it verifies, triangulating what a brand says about itself against what third parties say, and it gets better at this with every generation. Like a person, it judges a brand by whether what it says matches what it does, over time. That is why I don’t believe it can be gamed the way search and social were.
In practice the match runs in two phases. First eligibility: does this backpack meet the request for water resistance, has the brand said so clearly, and has it given proof? Then preference: taste, colour, material, whether this person leans towards a sustainable brand or a fashion one. That second phase draws on the user’s personal context, which sits in no model. It is bespoke to every person prompting, and it is where lovability hides. People feel it; the machine now has to infer it on their behalf.
Does every brand need both? The requirement holds; the stakes vary. In some categories, and for some buyers, nobody cares which brand turns up as long as it does the job at a fair price. I am passionate about cars; plenty of people only need to get from A to B. A brand can compete on specification and price, but that is still a choice to compete as a commodity.
It is not humans versus machines. There is a digital intermediary now in the relationship between people and brands.
Questions this conversation answers
Do the emotional reasons people choose a brand still count when an AI draws up the shortlist?
Yes, and the AI is now the one weighing them. Flavia raised a common misreading on air: that lovability happens only offline, in the human experience. The person still makes the final call, but before they do, their AI infers what they would prefer. My daily assistant knows a lot about me after months of use. When I ask it something, it executes the request and brings in the memory it has built of me, matching on function and on what I find important: my taste, my preferences. That inference is where lovability hides. The AI reads my context and the brand’s footprint, what other people say about it, and approximates something close to preference. The more you use it, the better it gets at this.
What is the difference between what an AI already knows and what it looks up live?
Flavia gave the two their technical names on air: parametric memory and retrieval memory. Parametric memory is what a model learned in training. Periodically, much of the internet is scraped, including everything you publish about yourself and everything third parties say about you. Any mismatch between the two, what I have always called the integrity gap, gets documented and becomes part of the training data. Retrieval memory is what the AI fetches live when you give it a task: whether the restaurant is open, the latest menu, the latest reviews. Then there is a third layer, and I would argue it matters most: the personal context of the person asking. If you and I each ask for a restaurant in Amsterdam, we get different answers. Flavia added the warning worth keeping: retrieval memory owes you nothing, so the durable place is claimed in what models learn, by thinking long term.
Will ads in AI assistants change how brands reach customers?
They add a paid channel beside the answers. They do not buy a place inside them. As we recorded, ChatGPT had just brought ads to its free tiers in Europe, after the United States earlier in the year. OpenAI says they are always clearly labelled and kept separate from the answers, and that they do not influence the response. That separation has to hold. The relationship people have with their assistant depends on its reliability. If the answers could be biased, sponsored or gamed, people would stop trusting the AI, and that would destroy the usefulness of the whole product. So treat sponsored placement as a channel next to the answer, and keep the answer itself earned. Where the AI runs on the customer’s own device, I expect no ads at all.
What happens to a brand AI can read but nobody particularly prefers?
It becomes interchangeable. When all that counts is the problem solved, the specification, the price and the availability, the buyer does not care which brand it is, and neither does their agent. That is what happened with my microphone receiver: half the price, it works, and I couldn’t tell you the brand. How much that matters depends on the category and on the person. I am passionate about cars and how they make me feel; plenty of people only need to get from A to B. To some buyers a rucksack is a tool, and to others it is a status symbol worth 3,000 dollars. The same holds in B2B, where a buying team that picks the newcomer over the incumbent shows what it values: innovation, risk-taking, how it believes it succeeds. The requirement holds for everyone; the stakes vary. Competing on specification and price is viable for some, but it is still a choice to compete as a commodity.
Why can’t AI find a brand its customers already choose?
Because being loved and being legible are separate achievements, and sometimes the name itself gets in the way. Coffeecompany, a chain in Amsterdam I love, is a case in point: when I asked my AI when The Coffee Company closes, it could not tell the brand from the category. The name is a complete descriptor. For anyone working on naming and brand architecture, that is worth noting: a category descriptor in a brand or product name makes it hard for a machine to work out what you are asking for. A lovable brand that isn’t legible is a ghost. Its customers cherish it, and the systems through which most discovery now happens cannot see it. Flavia recognised the risk at once. Brandingmag started life as Branding Magazine, and she explained that it was renamed partly because a descriptor cannot be trademarked. She wondered what an AI would have made of the old name.
What is WebMCP, and does my brand need one?
WebMCP makes a website actionable by an AI. Most people know MCP from the connectors in their AI assistants: a back door to a service, an API made for agents to act on. Without a WebMCP, an agent visiting your site can only use the human surface: open pages, scroll, take screenshots, fill in forms. It works, but it costs compute and time. With one, the agent has its own entrance and acts directly on the functions you have exposed. I see it as the next frontier of legibility, and the brands that offer useful actions there will win that side of the Law. I also expect a split: few people will connect a brand’s MCP before they buy, so WebMCP serves discovery and the pre-purchase relationship, while MCP serves the relationship after purchase. WebMCP is a proposed open web standard, drafted by Google and Microsoft engineers in a W3C community group and in public trial in Chrome since mid-2026. It is not yet a formal standard. Building one is relatively inexpensive; I have already experimented with building one. If you have the bandwidth, start.
Calls made on air
Published 30 September 2026; windows added in October 2026. Each call as I made it, with the window in which I expect it to land: the year it could first happen, and the year by which I think it will have happened.
- WebMCP becomes the next frontier of legibility; sites without one are left to agents clicking through the human interface. (52:05) From 2026, by 2027.
- Few people will connect a brand’s MCP: WebMCP carries discovery before purchase, MCP the relationship after it. (52:52) From 2027, by 2028.
- The major players build AI-native entry points, because speed and convenience win over agents browsing pages. (57:21) From 2027, by 2028.
- Brand Constitutions will be governed by several stakeholders across the organisation, convened by a brand leader. (39:57) From 2026, by 2028.
- Ads in AI assistants must stay clearly separated from the answers, or trust in the assistant collapses. (31:15) Holds now.
- The personal AI keeps getting better at trustworthy recommendations, so it cannot be gamed the way search and social were. (30:34) Direction of travel.
In their words
“In order to exist as a brand in that space, you need to be legible. And subsequently, to be chosen by a human on the other end, or to be recommended by an AI, you need to be lovable.”
“In that inference, this is where the lovability is hiding.”
“That personal context is not in the LLMs. It is bespoke to every single user prompting.”
“We do it by observing a longitudinal cohesion in their action and behavior and values and communication.”
“Retrieval memory, on the other hand, it owes you nothing.”
Chapters
- 00:00 A second foundational episode
- 01:22 The Legible-Lovable Law, defined
- 08:02 The lovability the machine infers (and the offline myth)
- 09:50 Parametric memory vs. retrieval memory
- 20:19 The research on retrieval volatility
- 27:46 First principles: a brand that cannot be gamed
- 32:55 Brand leaves marketing, and the editor-in-chief debate
- 41:05 Legible, not lovable: the interchangeable brand
- 42:15 Category and audience change the stakes
- 47:25 Lovable, not legible: the invisible brand
- 50:25 MCP, WebMCP, and the next frontier of legibility
- 58:50 Closing: both sides, and what’s next
Sources
- Coffeecompany, locations
- OpenAI, “ChatGPT Ads expands across Europe”, 18 August 2026
- W3C Web Machine Learning Community Group, WebMCP, Draft Community Group Report
- Chrome for Developers, “15 updates from Google I/O 2026”, 19 May 2026
Where to start
A readiness assessment measures both halves of the Law: whether agents can find and verify your brand quickly, and whether they can infer why someone would prefer it. That is where the Agentic Brand Readiness Assessment begins.
Explore the engagements · Start a conversation
About the series
Legible and Lovable is Brandingmag’s podcast on agentic branding, hosted by Flavia Barbat, editor-in-chief of Brandingmag, and Thomas Marzano. In Brandingmag’s words: “Thomas identified the field’s dynamics and wrote the initial thesis; Flavia named the field and edited the content.” Episode 4 is the series’ second foundational episode: what each half of the Legible-Lovable Law asks of a brand, and what happens to a brand that has only one.
Where this sits. Agentic Branding is the discipline of making brands intelligible, governable, and desirable in an AI-mediated world. Its governing principle is the Legible-Lovable Law: brands must be legible to machines and lovable to people, with the machine now carrying both halves.
All episodes of Legible and Lovable. Episode notes by Thomas Marzano, updated 2026-10-06.