Does AI believe what a brand says about itself, or what others say about it?
Flavia Barbat and Thomas Marzano in conversation. Legible and Lovable, episode 8. Published by Brandingmag, 2026-11-12. 66 minutes.
Episode notes by Thomas Marzano.
Neither on its own. AI now infers a brand from all the evidence it can reach: what the brand publishes, what customers and the press leave behind, and what it knows about the person asking. Where the promise and the evidence part, it says so, and the gap is closed inside the company, not on the website.
Inferred from the evidence
When most of us first met AI as consumers, it was a chatbot on a website, and what it did was summarise: you asked, and it gave you a digest of what it had been trained on, mistakes included. Then it began to remember us, conversation after conversation, and to read the live web next to its training data. That is when inference started. Ask it for three running shoes to consider for a marathon, and it filters for what is sold where I live and for what it knows about how I run, before I say any of it. Then it checks which of those shoes people like me recommend, and comes back with a shortlist and one pick on top: the Shortlist Effect, seen from the inside. Flavia has a name for what this does to brands: the inference problem.
I always start with the incentive. The labs that build these assistants need us to trust the answer; if we cannot rely on it for a purchase or a decision, the technology fails, and so do they. So the system is built to check a brand’s promise against the evidence and to surface where the two part.
That is where the costume brand is exposed: a brand managed on the surface, in its assets, campaigns and website, rather than governed in what the company does. Being on brand was always about cohesion, and cohesion means saying no to what you do not stand for. Every act of a company now leaves a digital trace, from customer care to the supply chain to how it treats the people who work there. When those traces do not echo the promise, the delta is what the manifesto called the integrity gap, and an agent reads it in milliseconds.
My own thinking has moved on since the manifesto. Brands have always averaged their customers into a primary and a secondary audience and a handful of personas. The agent holds more context about the person it serves than any company ever held about a customer, so a brand that says more about itself, and says it richly, can be inferred for people it never targeted. In Copenhagen I asked for a restaurant within a fifteen-minute walk where I could comfortably eat alone. The answer came with a reading of the menu against what it knew about me, which I had not mentioned. I went where it recommended. A better restaurant may have been round the corner; it never made itself visible enough for the agent to find.
Flavia has pushed brands for years to say how they work, from sourcing to hiring. Her verdict on air: it is bad business to hide, and once you have said it, you have to live by it, because a lived experience becomes memory, then something shared, then the digital footprint the agent reads as evidence: the proof layer. On Oxford Street I asked ChatGPT’s voice mode, in my ear, for a shop that sells a Peaky Blinders hat. It found one, then a cheaper one, guided me there, and in the shop told me what to look for in the fit. I never saw a website or any branded surface. Being inferred, helpfully and authentically, at the moment of need: that is the inference challenge.
Flavia asked the question every research report on AI visibility asks: does the AI believe the identity a brand declares, or the one it infers from what others say? Her answer is that brands have far more power over their own narrative than they use. Mine is that reputation is the outcome of acting with integrity. Nobody can order five more metres of it. Integrity is a decision.
That is why the gap cannot be fixed on the surface. If a restaurant calls itself the best in town and the reviews say the steaks are poor and the room is cold, more praise on its website changes nothing. Ask why five times, find the root cause, change it. A press release can contain the damage; it cannot repair it. For the last ten years of my career as a head of brand, much of the job was managing the cohesion of behaviour deep inside the company, which is why a brand leader needs the mandate to govern there.
Most companies I speak to now track their AI visibility. Much of what it shows is spring cleaning, and it is table stakes. The step change comes when the tracking shows a pattern rather than a blip, in quality, supply, service or employment: then the brand leader takes it to peers, finds the root cause and funds the fix. On how much weight the score itself deserves, Flavia and I disagree; both sides are in the questions below.
Flavia closed on four things a brand needs: accuracy, consistency, substance and proof. I added one: know what role your brand plays for your company. In categories that agents will buy almost entirely on the person’s behalf, that is the future to plan for.
Questions this conversation answers
Does AI believe what a brand says about itself, or what others say about it?
Neither on its own. An AI assistant infers a brand from all the evidence it can reach: what the brand publishes about itself, what customers, employees and the press leave behind, and what it knows about the person asking. Where the brand’s promise and that evidence agree, the brand is easier to recommend; where they part, the assistant says so. Flavia Barbat’s point is that brands have far more power over that evidence than they use: they can say much more about how they work and which needs they serve. Mine is that what others say follows from what the brand does. Reputation is not an input a team can order more of; it is the outcome of acting with integrity, and integrity is a decision. The chain from lived experience to the agent’s evidence is the brand’s proof layer.
What changed when AI stopped summarising and started inferring?
The early assistants summarised what they had been trained on. Two changes turned summary into inference: they began to remember the person across conversations, and they began to read the live web next to their training data. With the model’s own knowledge, the person’s context and live sources together, an assistant can triangulate. Ask it for marathon shoes and it filters for what is sold where you live and what it knows about how you run, then checks which of those people like you recommend, and comes back with a shortlist and a pick. Brands are no longer only summarised. They are inferred, from sources they only partly control. Flavia Barbat calls what this does to brands the inference problem. In the sequence the field guide describes, from summary to inference to rendering, inference is the present.
Listen from 01:50 · Why do AI assistants recommend some brands and not others? (Ep. 4)
Why would an AI expose a brand that breaks its promises?
Because its makers need people to trust its answers. If an assistant cannot be relied on for a purchase, a decision or a task, people stop using it, and the businesses behind it fail. So the system is built to check a brand’s promise against the evidence and to surface the inconsistencies: whether the brand keeps its promise to the people it was made for, and for whom the promise does not hold. The answers are already considerably better than they were a year or two ago; I read that as the incentive working. For a brand, the consequence is plain: the gap between what it says and what it does is now part of how it is described. What happens once that gap is exposed is the subject of episode 1.
Listen from 06:59 · What happens when AI exposes the gap between what a brand says and what it does? (Ep. 1)
Can a brand fix what AI says about it with better content?
Only where the problem is on the surface. If customers, staff and the press say the product, the service or the employer falls short of the promise, more content on the website changes nothing: the assistant reads them as well. Ask why five times until you reach the root cause, and fix that. A press release can contain the damage; it cannot repair the gap. That is the weakness of the costume brand, a brand managed in its assets and campaigns by a leader with authority only over the surface: it struggles once an agent reads everything else. Closing the gap needs the mandate to govern the brand inside the company, which is what a Brand Constitution is for.
Should a brand say more about how it works, even if competitors can copy it?
Yes, says Flavia Barbat, with the obvious exceptions: legal matters, intellectual property, customer data. For years her advice to brands has been to show their approach: their sourcing, their ingredients, how they hire, the steps they take. Many hold back for fear of being copied. Her experience is that a brand that speaks up wins over a better one that stays quiet, and an AI assistant raises the stakes, because it uses every detail it can find to match a brand to a person’s need. In her words, it is bad business to hide. Saying it also binds you: once you have said how you work, you have to work that way, because the assistant will check.
Can AI recommend a brand to customers it never targeted?
Yes, if the brand gives it enough to infer from. Brands have always averaged their customers into a primary and a secondary audience and a few personas. A personal assistant holds more context about the person it serves than any company ever held about a customer, and it uses that context unasked. In Copenhagen I asked for a restaurant within a fifteen-minute walk where I could comfortably eat alone; the recommendation came with a reading of the menu against what it knew about me, which I had not mentioned. That is the opportunity: a brand that describes what it offers richly, and for whom it works, can be inferred for people outside its personas. The risk is the mirror image. A restaurant that might have suited me better never appeared, because it had not made itself visible enough. This is where my thinking has gone since the manifesto.
What should a brand team do with its AI visibility dashboard?
Sort what it shows into two piles. Much of it is spring cleaning: content that disagrees with itself, a knowledge graph out of step, an encyclopaedia entry that says something other than the About page. That work is table stakes and should be done. The second pile matters more: findings about quality, supply, customer care or the company as an employer. When those are a pattern rather than a blip, take them to your peers, find the root cause, and fund the fix. Flavia Barbat adds a point about the prompts themselves: teams tend to test for the good news, such as whether the brand makes the top three, when the most useful prompts look for the problems. And read every score for what it covers: an honest reading names the AI surfaces it was taken on, and the ones it leaves out.
Listen from 42:53 · Is AI visibility tracking still worth it when AI moves onto customers’ devices? (Ep. 3)
Which AI visibility gaps should a brand fix first?
The ones that stand between the business and its goals. My answer is the brand leader’s usual discipline: take a baseline, set a realistic target, fund it, and be clear how much of the gap surface work can close and how much needs the company to change. Then ask whether AI visibility is really what is costing you. If leads and sales hold, a dashboard alarm may not be the priority; if they fall, find the two or three causes behind most of the loss, and if AI visibility is one, fix it, knowing the tools are immature and the models change every week. Flavia Barbat sees it differently. The same prompt can return a different answer minutes later, so for her the value of the tools is in pointing at gaps: close those and the score follows. Her advice: a fixed set of prompts at checkpoints through the year, and the rest of the effort on sharper questions that change as you learn.
Listen from 49:29 · Why do AI assistants recommend some brands and not others? (Ep. 4)
What does a brand need to get right for AI to draw the right conclusions about it?
Flavia Barbat names four things. Accuracy: claims an assistant can check, because it will question the first, the only and the most when it cannot support them. Consistency: the same story across every employee, channel and format, which over time also helps an assistant weigh an old story against recent conduct. Substance: value and meaning in people’s lives, without which no assistant shortlists you and no person loves you. Proof: specificity, because, in her words, inference rewards proof and penalises any generic claim. I add one: know what role your brand plays for your company, which differs by category. Where products will be bought almost entirely by agents on the person’s behalf, plan the brand for that future.
Calls made on air
Published 12 November 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.
- Products and services that become commodities will be bought almost entirely through agents, and brands in those categories need to plan for it. (1:02:51) Direction of travel.
- A voice assistant can already act as a shopping assistant from the search to the fitting, with no website in between. (30:15) Holds now.
In their words
“It’s just bad business to hide.”
“I didn’t go to a website, I didn’t see any branded surface.”
“Reputation is the outcome of your acting with integrity, right? … It’s not an input.”
“It cannot be managed on your website. It cannot be managed with a press release.”
“Inference, I think, rewards proof, and it very much penalizes any kind of generic claim.”
Chapters
- 00:00 Costume brands and the inference problem
- 01:50 From summarizing to inferring: how AI’s role changed
- 06:59 Why does AI have an incentive to expose broken brand promises?
- 13:25 What is the integrity gap, and why does it matter now?
- 15:25 Inference as a strategic lever for brands
- 22:27 Why is it bad business to hide how you work?
- 25:06 The verification chain: from lived experience to the agent’s evidence
- 31:50 Why aren’t brands owning their own narratives?
- 35:10 Why can’t the integrity gap be fixed on the surface?
- 42:53 What should brand leaders do with AI visibility tracking?
- 49:29 How do you decide which AI visibility gaps to fix first?
- 58:38 Accuracy, consistency, substance, and proof
Sources
Where to start
Finding where your brand’s promise and its evidence part, and what it will take inside the company to close the gap, 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 8 is a foundational episode: the two hosts on the inference problem, without a guest.
Where this sits. Agentic Branding is the discipline of making brands intelligible, governable, and desirable in an AI-mediated world. Its governing principle is Thomas Marzano's 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 2026-11-12.