The End of the Segment
What a cat picture, a coupon flyer, and Alexa's new shopping brain have in common
Before the main story, a few quick takes worth your attention.
Google is pushing AI shopping toward agentic commerce. Its new Universal Cart can pull products from across Search, Gemini, YouTube, Gmail and retailers, then help track prices, surface deals and move shoppers toward checkout. For marketers, products now need to be discoverable by AI agents, not just people. More from Search Engine Journal.
Amazon relaunched Alexa for Shopping directly in the search bar. Customers can now ask questions in Amazon’s main search bar, generate product comparisons, see price history, build shopping guides and automate deal-finding or routine purchases. This is another sign that retail search is becoming conversational and assistant-led. More details below in the main story.
YouTube is becoming a more AI-native discovery and creation platform. Google’s Gemini Omni will power new multimodal creation across products, including YouTube Shorts, making video creation and remixing faster and more conversational. For brands, content strategy will need to account for AI-assisted creation, not just distribution. More at the YouTube Official blog.
Etsy is testing conversational shopping inside ChatGPT and on Etsy. Its ChatGPT app is live in beta, while Etsy is also testing a conversational search experience for gift discovery. The broader signal is clear: shopping is shifting from keyword search to intent-rich conversation. More here.
When the Page Becomes a Conversation
In 2012, a team at Google X fed about 10 million unlabeled YouTube thumbnails into a neural network running across roughly 16,000 processors and a billion connections. Nobody told the model what a cat was. Nobody labeled “cat” as a category. The model just watched, and watched, and watched. And one day, somewhere inside that pile of math, a neuron lit up reliably every time a cat appeared on screen.
The resulting research paper was a turning point in the world of AI. It didn’t just prove that deep learning worked. It made a much stranger claim: that with enough data, enough compute, and enough patience, recognition itself could emerge without explicit instruction. Intelligence as a side effect of scale.
Hold onto that idea for a moment, because it’s exactly what is happening to marketing right now. Let me explain.
For three decades, personalization has meant rules. If a customer is in segment A, send creative B. If they abandoned a cart, fire email C 24 hours later. We hand-coded the cat. The new generation of systems doesn’t need us to. They watch what people do across watching, browsing, and buying, and the relevant patterns light up on their own. The unit of optimization has stopped being the segment. It’s becoming the decision: what to say, show, offer, or quietly not send, in this exact moment, to this exact person.
That shift is bigger than it may appear on the surface. And it brings back an old ghost that those of you who grew up in digital marketing may remember.
The story marketers never quite shake
A father walks into a Target outside Minneapolis holding a mailer addressed to his teenage daughter. It’s full of coupons for cribs, baby clothes, and maternity wear. He’s furious. “She’s still in high school, and you’re sending her coupons for baby clothes and cribs? Are you trying to encourage her to get pregnant?”
A few days later he calls back to apologize. It turns out his daughter was pregnant. Target’s model knew before he did.
That story, which Charles Duhigg made famous in The New York Times Magazine more than a decade ago, has become marketing’s version of a campfire warning. The lesson wasn’t that the model was wrong. The model was eerily, painfully right. The lesson was that being right is not the same as being welcome. Target had figured out the math of predicting pregnancy from shifts in shopping behavior using predictive analytics. What it hadn’t figured out was the social cost of showing its work.
We are about to walk into that same room again, but with vastly more powerful technologies. The 2012 cat neuron showed that pattern recognition can emerge from scale. The 2026 personalization stack shows that decisions can emerge from scale. And every one of those decisions, like the one inside that Target envelope, sits somewhere on a line between useful and creepy.
Gartner’s most recent numbers tell you exactly where we are on that line. Fifty-eight percent of consumers say brands fail to understand their needs. Only fifteen percent of CMOs acknowledge a meaningful gap. Nearly half of consumers describe personalized digital communications as irrelevant, creepy, or both. Personalized customers are 3.7x more likely to purchase more than they originally intended. The upside is real. So is the trip wire.
Spotify: personalization as an operating system
The easiest way to see what good looks like is to open the app most of us already use to fall asleep.
Spotify’s recommendations don’t come from one model or one engine. They come from a system. The catalog is tagged by genre, era, tempo, mood, and a long list of features most of us never see. The app watches what you play, when you play it, what you skipped at twelve seconds, and what surfaced you in the first place. Behind that, a constant churn of micro experiments runs on slices of the user base, with human curators shaping playlists at the top of the funnel.
What’s changed in the last year is what sits underneath. Spotify engineers describe the older stack as “TradRecs”: classic multi-stage pipelines doing candidate generation, ranking, and scoring. The new layer translates both content and user histories into vector embeddings and semantic IDs, then fine-tunes LLMs on Spotify’s internal catalog and interaction logs so the model speaks the same language as the data. The headline shift, as ML engineer Shivam Verma framed it publicly, is from analytical recommendations to steerable recommendations: a user can hand the system listening history plus a plain prompt like “give me an episode I could listen to next,” and the model returns a personalized answer it can also explain. Tools like Taste Profile let users push back on the model directly. Discovery becomes a conversation, not just a feed.
The result is a user base and revenues that have each grown roughly 1,000% over the past decade, to more than 600 million users and $14 billion. The more useful number for marketers is the one nobody puts on a slide: very few people unsubscribe from Spotify because the recommendations feel invasive. The discovery feels like the product. The personalization is the product.
That’s the bar. Not “we A/B tested two subject lines.” A living operating system where signal quality, content intelligence, human judgment, and now LLM-native reasoning all feed each other.
Pandora: when gen AI compresses the calendar
Pandora, the global jewelry brand, shows what happens when generative AI changes both the content calendar and the customer conversation.
On content, Pandora has used generative AI to tailor messaging and imagery to individual customers, including backgrounds and model images. What once took 12 to 14 months can now be compressed to roughly 10 days. That is not just personalization. It is a collapse in the cost and time required to create variants.
On customer experience, Pandora has put AI agents closer to the front line. Clara, its customer service agent, now resolves a meaningful share of inquiries without escalation. Gemma, its sales agent being piloted in Australia, is the more interesting bet: an effort to recreate some of the emotional texture of in-store jewelry selling online. Instead of simply filtering by product type or price, the agent can ask who the gift is for, what the occasion means, and what memory the piece is meant to hold.
That matters because jewelry is not just a product category. It is an emotional category. Pandora is testing whether AI can do more than answer questions or recommend SKUs. It is testing whether AI can help scale assisted selling.
And that is the deeper marketing lesson. When the production cost of variants and conversations both approach zero, the constraint stops being production and starts being judgment. You no longer ask, “Can we afford to make a version of this for that segment?” You ask, “Should we?” That is the harder question, and the one Target got wrong.
Ulta Beauty: retail media as commerce intel
Sitting underneath all of this is something less glamorous but more consequential: retail media. BCG estimates the global retail media advertising business is growing at 25% a year and will pass $100 billion worldwide by 2026. Onsite retail media on a retailer’s own site can carry gross margins north of 85%. For some retailers, data-driven advertising now generates as much revenue as the products on the shelf.
The headline names are Amazon, Walmart, Target, Kroger, and Instacart. But the category specialists are doing something instructive. Take Ulta Beauty which ADWEEK recently highlighted. CMO Kelly Mahoney runs a loyalty program with 46.7 million members that now drives roughly 95% of company sales: the kind of first-party data foundation most marketers, in her words, would die to have. On top of it Ulta has built its own AI engine, the Beauty Graph, which has hundreds of automated “recipes” deciding what each shopper sees next across site, app, email, and paid channels.
One shopper is shown an exclusive in a category she already buys; another gets nudged across categories; another sees a replenishment prompt. Humans don’t pick the recipe; the model does. Beauty has unusually rich personalization signal in the first place: replenishment cadence, skin and hair profile, gifting patterns, season. Ulta is wiring all of it into a closed loop where every transaction makes the next recommendation smarter, and pairing it with a pod operating model that puts marketing, product, IT, and channel owners on the same customer goal.
That phrase, commerce intelligence, is the right way to think about retail media in the AI era. It isn’t a placement. It’s an input to the personalization engine itself.
Alexa: the agent that actually knows you
Now to the moment that should make every one of us sit up.
Alexa for Shopping is now powered by Alexa Plus, Amazon’s LLM-driven assistant, and it lives front and center in Amazon’s main search bar, the app, and the Echo Show. It pulls from Amazon, the open web, and what it already knows about you.
Picture a parent typing “what did I get my mom for Mother’s Day last year so I don’t repeat it” into Amazon’s search bar. The old answer was a list of generic mom gifts. The new answer is the actual silver bracelet from May 2025, three coordinating pieces the agent thinks she’d like, and a reminder that her birthday is three weeks away. That is what changes.
This is what “the page becomes a conversation” actually means. For thirty years, the brand-to-customer interface was a page: a search results page, a product page, a checkout page. Marketing’s job was to win attention on that page and optimize the path through it. With Alexa for Shopping, the page collapses into a dialogue, and the dialogue is mediated by an agent that knows the customer better than most brands do.
If the agent decides what gets considered and what gets bought, the meaningful question for a brand becomes, “Am I legible to that agent?” In practical terms, that means clean structured product feeds with attribute-level metadata, machine-readable claims, replenishment cadences exposed in the data, return policies written so a parser can interpret them, and a brand voice present in the reviews the model summarizes. If you’re not legible at that level, you’re not in the consideration set.
What’s actually behind the search bar:
Cross-device memory. A conversation on your kitchen Echo carries into your laptop search. Ask Alexa about a kid’s science project on a Show, then type “show me what I need to buy for my science project” on Amazon.com, and the model already has the context.
Standing instructions. “Add my regular dog treats” works. So does “add this sunscreen if it drops to $10 and I haven’t bought it in two months.” So do birthday gifts, recurring orders, and refill schedules. Alexa watches and acts, with a full year of price history per product behind it.
Off-Amazon reach. Via the agentic “Buy for Me” feature, Alexa completes purchases on third-party retailers on your behalf.
Rufus, the precursor that handled product questions, served over 300 million customers in 2025. Alexa for Shopping is the next move, free to all U.S. Amazon customers, no Prime required, no Echo required.
This is also where the Target lesson comes back hard. An agent that knows a customer is pregnant, or grieving, or trying to quit drinking, or hiding a purchase from a partner, holds a level of trust no marketing channel has ever held. Mishandle it once and you don’t lose a campaign. You lose the relationship.
What this means for marketers
If you take the cat neuron paper, the Target mailer, Spotify, Pandora, Ulta, and Alexa and stack them on top of each other, a few practical implications fall out.
Signal quality is the new media advantage. As Google Performance Max and Meta Advantage+ automate more campaign controls away from human hands, the thing you actually control is the richness, recency, and relevance of the signals feeding the model. First-party data, retail media signals, offline outcomes, call and service data. The platforms will do the targeting. You feed them better food than your competitors do.
Context beats audience. Identity alone is no longer enough. The full stack is identity plus mode plus moment plus channel plus memory. A “beach vacation seeker” is a segment. “Family of four, school-break window, comparing cancellation policies, came back from mobile search at 11pm” is a context. The second one is where the next decision actually lives.
Creative has to become model-readable. Stop thinking about creative as a finite set of executions. Think of it as a system of modular building blocks governed by machine-readable brand rules and assembled dynamically. Yum Brands sent 200 million AI-generated communications and saw up to 5x more incremental performance than traditional approaches. The volume is real. The risk is creative wear-out and brand drift, which is why curation matters more, not less.
Synthetic audiences are a planning tool, not a verdict. AI personas can pressure test ideas, unpack insights and reveal blind spots before you spend media. They cannot tell you what will actually happen in market as they maybe too many variables at play. Hold out groups and incrementality tests still decide truth.
Agents are the next interface. Build for that future now. Make your product feeds, claims, reviews, policies and APIs legible to agents. Assume that within the next few years, a non-trivial share of consideration and purchase will route through Alexa, ChatGPT, Gemini or client-side agents acting on behalf of the customer. The old SEO playbook becomes the new AEO playbook, optimizing not just for a search engine, but for an agent making sense of options, constraints and intent. In practical terms, as Bhanu Sharma, CEO of Maker.co says, brands will need two layers: one emotionally rich, immersive and designed for humans, and another machine-readable layer that allows AI agents to understand, compare, recommend and transact with confidence.
Governance is a performance discipline, not a compliance checkbox. The Target story didn’t fail on math. It failed on judgment. Decide explicitly what AI is allowed to decide on its own, what requires a human, where you’ll suppress a message even when the model says fire, and how you’ll explain a recommendation when a customer asks. Trust is now a measurable performance variable. The brands that treat it that way will compound. The brands that don’t will keep generating dad-in-the-Target-store stories, at machine speed and at machine scale.
Closing thought
The 2012 cat paper was the moment we stopped having to tell models what to look for. The 2026 Alexa launch is the moment we stop having to tell consumers what to look for. The connecting thread is the same: scale plus signal plus self-improving systems are quietly absorbing tasks that used to live inside human teams.
Personalization is where AI is going to land first, hardest, and most visibly. Not because it’s the most exciting use case, but because it sits at the intersection of three things every business already has: customers, content, and decisions. The brands that win the next decade will not be the ones that personalize the most. They’ll be the ones that personalize the best: with richer context, sharper governance, smarter signals, and a system designed to earn trust at every touchpoint.
The cat is out of the box. The question is whether your brand becomes the one the agent recommends, or the one a customer turns to a stranger and says, “You won’t believe what they sent my daughter.”
Where I’ve been
Last week, I had the privilege of joining an Alpha Board meeting focused on AI in the boardroom, followed a day later by a National Association of Corporate Directors dinner in Palo Alto. Across both conversations, the through line was clear: boards are no longer asking whether AI matters. They are asking how to govern it, how to understand its risks, and how to recognize where it is already reshaping strategy, operations, talent and competitive advantage. This aligns with my own experience, having served on a large public company board for eight years now. And this morning, I also ran an AI training session on Personalization in the AI Era for 200 marketers earlier today, which helped inspire this piece.
In early June, I’ll be speaking at the Financial Times Board Summit in New York, where these questions will continue to be front and center. From there, I’ll be heading to the Cannes Lions International Festival of Creativity for a series of meetings, a Mastercard-curated panel and potentially an AI Trailblazers gathering. Email me if you’d like to discuss Cannes further.
What I’ve written lately
When Intelligence Enters the Org Chart (May 2026)
A Public Company Goes All In on AI (April 2026)
AI Is Rewriting Who Decides (March 2026)
Fighting Cognitive Surrender (March 2026)
Shiv Singh is the CEO of Savvy Matters, which helps business teams translate AI disruption into practical business and marketing strategies, organizational design, executive-ready roadmaps, and bespoke education programs. He is also the Co-Founder of AI Trailblazers, a vibrant community uniting marketers, technologists, entrepreneurs, and venture capitalists at the forefront of AI.
A former two-time Chief Marketing & Customer Experience Officer and author of Marketing with AI for Dummies (4th print run, translated into five languages), Shiv built his career at LendingTree, Visa, PepsiCo, and The Expedia Group, and serves as a public-company board member of a Fortune 300 company and private investor.