This article is based on a Q&A with Klaviyo Chief Marketing Officer Jamie Domenici and Greg Kihlström for The Agile Brand Guide’s Expert Mode series.
Most marketing teams have decided that AI assistants are a visibility problem. Get the product into the answer. Budgets are shifting toward generative engine optimization in almost exactly the shape they shifted toward search a decade ago. Being absent from the result is worse than being third in it, and everybody running a brand knows it.
That instinct is correct as far as it goes, and it stops early. Jamie Domenici, Chief Marketing Officer at Klaviyo, thinks the real contest is happening a layer down, over the customer record that any answer gets built from. Her question for marketing leaders is whether a brand can still see its own customers once a model sits between the two.
Getting Into the Answer Is a Different Job Than Knowing the Buyer
Domenici doesn’t dismiss the GEO push. She thinks teams are right to chase it. Her objection is that optimizing for a shopping agent and understanding a buyer are separate pieces of work, and marketing keeps collapsing them into one.
Teams are already optimizing for GEO the way they optimized for SEO a decade ago, and that instinct is right. But getting into the AI’s workflow is a different job than understanding the person on the other end of it. The purchase decision is still made by a human who’s weighing things AI doesn’t fully see: trust, timing, how a product fits into their life… A CMO still owns the data that feeds the answer. Product info, reviews, and purchase history are what the AI is actually pulling from to tell a customer what to buy. Get that right, and you become the answer the assistant gives. Get it wrong, and you’re invisible before the customer ever reaches your site.
What she says a CMO gives up is the presentation layer. Layout, tone, the order a model decides to list things in. That belongs to someone else’s system now, and for a team that has spent years running landing page tests and rewriting subject lines, it’s a real loss of craft. What replaces it is unglamorous. Product titles. Attribute completeness. Review coverage on the SKUs that actually sell. Whether the size chart is machine-readable or a JPEG. None of that gets a case study written about it, and all of it decides whether a model can describe the brand accurately when a shopper asks a vague question at eleven at night. So the practical move this quarter is to find out who owns first-party product and customer data inside the company. Usually that’s merchandising or engineering. Get into that room.
Context Means a Record That Updates Itself
The phrase “customer context” gets used loosely enough to mean almost nothing. Asked what it means at the level of a single record, Domenici got specific.
At the record level, context means real, individual customer data. When does this customer usually buy? Do they respond to texts, or ignore them and read email instead? What’s sitting in their cart that they keep coming back to but haven’t bought? That’s the detail that separates a B2C CRM from an LLM. An LLM can hold a document about a customer and generate content from it, but that reasoning goes stale fast — every new purchase or support conversation has to be manually fed back in to keep the picture current. What works is a live record that updates itself every time the customer clicks, opens, or buys. Without it, a company sends messages that sound plausible but land wrong: the right offer on the wrong day, personalization that reads as generic because it’s guessing at what the customer wants.
“Plausible but land wrong” names a failure most measurement doesn’t catch. A campaign built on a stale snapshot doesn’t produce an obvious error. It produces a slightly-off offer that gets ignored, and ignored messages look like a creative problem in the dashboard. They’re a freshness problem. Domenici ties this to Klaviyo’s September announcement that it’s opening the platform through MCP tools and APIs, so outside agents can read and write to that same live record. Klaviyo sells the record, of course, which makes “context beats the model” a fair description of the company’s own moat. That doesn’t make it wrong. The operational point holds for brands that will never buy Klaviyo. If a marketing team can’t say how many hours old its customer data is at the moment an agent acts on it, the real number is probably worse than anyone wants to admit.
The Gatekeeper Test Is Whether the Brand Can See Inside
Klaviyo warns that LLMs shouldn’t become gatekeepers of the customer relationship, and Klaviyo builds agents that sit between brands and their customers. We asked Domenici to argue the skeptic’s case against her own company. She took it head-on.
Any layer that sits between a brand and its customer holds real influence, simply through the knowledge it has access to. But an agent that decides what a customer sees, what gets recommended, what gets pushed to the top, functions as a teammate and not a gatekeeper… Here’s the distinction I’d defend. Gatekeeping means an outside party sits between a brand and its own data, deciding what the brand gets to know about its own customers. Klaviyo’s agents run on the brand’s data, inside the brand’s account, visible to the brand at every step.
She’s drawing the line at visibility, and visibility is at least testable. Set her description against an assistant that owns the interface and the reasoning and shows a brand none of it. Nobody can say why one product surfaced over another. Nobody owns the conversation history. Whether or not you accept Klaviyo’s framing, it hands a CMO three questions for any vendor pitching an agent. Can we see the reasoning behind a recommendation. Do we own the conversation log. Can we change the guardrails ourselves without filing a ticket. Yes to all three describes a supplier. No to any of them describes a partner with real leverage over your customer relationship, which is a different kind of contract.
Service Is Where the Experience Claim Gets Tested
“Customer experience is the competitive advantage” has been said at every conference since about 2014. We asked for a case where the difference showed up in a number, and Domenici pointed to Naked Wardrobe, an apparel brand running Klaviyo’s Customer Agent on the same data as the rest of its business.
They put our Customer Agent to work for service, running on the same customer data as the rest of the business, so the agent already knows a shopper’s order history, sizing preferences, and past conversations before it responds. In the first 90 days, it resolved 86% of customer queries and 94% of product recommendation queries on its own, no human needed.
Those figures come from Klaviyo’s own case study, but are compelling nonetheless. The structural claim underneath matters even more than the percentages. Service and marketing were reading the same profile, so the agent knew what the shopper had already bought before it recommended anything. Most brands can’t do that. Service sits in one system, marketing in another, and the two reconcile overnight if at all. Testing this doesn’t take a pilot. Open a support ticket as a customer and see whether the reply reflects your last order. The answer arrives in about four minutes, and it’s usually no. That gap is where retention leaks, and it costs nothing to check.
Which leaves the harder question of what to fund. Domenici’s advice is to hold off on the roadmap conversation until the job is defined. Autonomous marketing and autonomous service are different products with different data requirements, and most pitches lead with the capability and leave the mission as an afterthought. Then fund whatever sits “closest to your actual bottleneck.” Drowning in campaign execution on top of messy data? Fix the data first, because an agent built on bad records just makes bad decisions faster. Losing customers after the first purchase? Fund service.
Her filter for everything else is blunt, and it’s arguably the most useful advice she shared, in a very insightful conversation: “Skip anything pitched purely as a capability, with no outcome attached to it. Ask what metric moves, by how much, and over what timeframe. A vague answer is the red flag.” Six months from now you’ll be defending this line in a room where nobody cares which model it runs on. They’ll ask what it fixed. Fund the thing that’s actually broken now and that meeting goes fine.
