The next e-commerce personalization layer is search

David Annez, SVP of Engineering and Product at Sanity, argues that AI-powered search is the next layer of e-commerce personalization, enabling storefronts to understand natural language queries like "earthy linen" while applying deterministic rules on price and inventory. He explains that AI reduces the infrastructure costs of personalization, allowing teams to build targeted experiences without the "combinatorial explosion" of caching and data-fetching that previously limited experimentation.

The next e-commerce personalization layer is search Building a search function that understands "earthy linen" and knows what sizes are left is more approachable than it used to be. David Annez SVP of Engineering and Product Published Table of Contents For a long time, e-commerce personalization meant building different experiences for different audiences. A returning customer sees one hero, while a high-intent cohort sees another offer, and a loyalty segment gets a different landing page. That work matters, but it is still only an approximation. The shopper is mapped to a group, and the experience is only as personal as that group is specific. Personalization can move much closer to the moment of intent today. A customer can describe what they want in their own words, and the storefront can use AI to find relevant products while still applying deterministic rules around price, inventory, size, location, and availability. This is a more useful framing for commerce teams rather than the usual question of whether chat will replace the website. Websites still matter. The real question is whether the storefront can understand intent as well as it understands filters. That was the core thread in the first episode of Sanity's Structured Series webinar /events/structured-series-ai-commerce , where I joined Dan Laugharn, Senior Solutions Architect at Vercel, to talk about personalization, AI, and what has actually changed for e-commerce teams. Personalization used to hit an infrastructure ceiling I’ve lived the old model first-hand. Before Sanity, I worked at loveholidays, an online travel agent, where relevance and speed both translated directly into conversion. Historically, getting personalization right required serious machinery and resources. We had an entire data science team determining the best way to personalize and target our content. Even with that level of investment, the targeting was still relatively vast or wide. The issue was not a lack of ambition or planning. It was the shape of the problem. Every new cohort, experiment, or page variant creates another branch in the experience. I call this the "combinatorial explosion." The more specific the experience becomes, the harder it is to cache. The harder it is to cache, the more often a page has to fetch fresh data from several systems before it can respond aka the slower the experience becomes for the customer . At loveholidays, the team was always trying to find the upper bound of experiments they could run on a page without harming load times. Sanity's Content Lake /content-lake can serve content with p95 latency under 80 milliseconds, but once a page depends on four or five services chained together, the total experience can still slow down. That is how teams end up with suboptimal compromises: four variations when they want twenty, broad audiences when they want specific intent, and a site that is technically personalized, but still not displaying the right item for the right person. Dan described how his previous employer, Made In Cookware, deliberately chose an engineering-heavy culture to build the experiences it wanted. For a DTC company with a small headcount, that was a real trade-off. Every idea started with the same questions: can we do this, should we do this, and how many engineers will it take? AI makes building personalized experiences faster and cheaper Dan described the shift as moving from telling to showing. "We can show and not tell that this is what is possible," he said. Teams do not have to spend weeks debating whether a personalized experience can be built. They can build enough of it to react to, test, and improve. I see the same change in the data and content layer. AI can help identify smaller cohorts, generate targeted variations, and update structured content at a scale that used to require a dedicated team. Here’s a practical example: connect a Model Context Protocol https://modelcontextprotocol.io/ server to a BigQuery dataset, ask for the top five customer cohorts, and start building five landing page variations from structured content. That is useful, but it does not make messy systems magically reliable. Generated variants still need a source of truth. Product recommendations still need live inventory. Price and availability still need deterministic answers. Customer-facing experiences still need governance, performance, and clear boundaries around what the system is allowed to say or do. AI reduces the cost of creating the experience as long as your content and product data is structured well enough for AI to use. Commerce is not becoming one big chat window There is also a popular narrative circulating right now that shoppers will move into general-purpose AI assistants and skip the website altogether. I am skeptical of that, especially for commerce. Customers may use chat to research broad categories, narrow down options, or understand what kind of product they want. But shopping is not only a question-answer flow. It is visual, current, contextual, and tied to the brand's own systems. An assistant can help you decide what kind of handbag you might want. It cannot always tell you whether the exact item is in stock near you, show the color in the right context, reflect the current price, apply the right promotion, and carry you through the checkout experience. "Your URL.com is still your digital storefront," Dan said. Instead of building features that try to move every shopper into a chat UI, consider bringing in more intelligence into the storefront itself – especially at the point where customers are already expressing intent: search. Search is where personalization gets practical Traditional personalization starts with a business question: which audience segment does this visitor fall into? But modern search starts with a customer question: what am I trying to find right now? A search query already contains personal context. It has the shopper's language, constraints, taste, and intent. Most e-commerce search functionality has historically forced the shopper to translate that intent into the site's taxonomy. A shopper does not naturally think, "category equals tops, color equals beige, size equals M, material equals linen." They think, "I need linen pieces in muted earth tones, size M," or "show me autumn sweaters under $150 that work with black jeans." Classic filtering makes the customer do the translation. Pure semantic search can understand the vibe, but it is not enough on its own for commerce. A storefront also needs hard answers: is it in stock, is it available in this size, is it under the price limit, can it ship to this location? Sanity Context /docs/ai/agent-context is useful in this use case because it lets teams combine semantic and deterministic search. The AI received a scoped, structured way to work with content and product data in the Content Lake. Semantic understanding deciphers the fuzzy parts of the request, while GROQ /docs/groq handles deterministic queries over structured data. The shopper can search in their own words, and the system can still respect the facts of the catalog. I demoed what this looks like in an example website I built where I searched for "linen pieces in muted earth tones, size M." The results included products that matched the style of the query, checked availability in size M, and understood that cream could be close to "earthy" even though the schema did not explicitly define cream that way. That is meaningfully different from swapping a hero banner based on a cohort. The experience becomes personal because it responds to the individual's stated intent. The search results remain reliable because the structured parts of the answer still come from governed product and content data. The infrastructure should not become the product Teams don’t need a bespoke infrastructure project before they can test this kind of experience. Anyone working in Sanity can configure Sanity Context with a content filter for products and an instruction prompt for how the agent should behave. Instead of dumping product data, markdown /glossary/markdown files, and business logic into a separate chatbot system, the context sits over the content model /glossary/content-model teams already use. On the frontend, Vercel's AI SDK https://ai-sdk.dev/docs allows developers to build the experience inside the application. A personalized search experience should feel like part of the storefront, not a third-party widget sitting above it. Dan said that the thing teams used to evaluate vendors for, negotiate contracts over, and provision quarterly is now closer to a Jira ticket. That does not mean the work is trivial. Teams still need clear content models, accurate product data, good UX, sensible prompts, and a performance architecture that keeps static and dynamic parts of the page fast. But instead of building a sprawling personalization platform before learning whether the experience works, teams can prototype and iterate inside the systems that power the storefront. From audience segments to intent The next phase of e-commerce personalization will not just be more variants for more cohorts. AI will make that easier, and it will be useful, but it is not the whole shift. The more important move is away from audience-based personalization to intent-based personalization. When a shopper searches, they are telling you what matters to them right now. If your storefront can understand that request, search the right content, apply deterministic business rules, and return a fast visual experience, personalization stops being a separate layer bolted onto the site. It becomes part of how the site works. Sanity and Vercel fit together here. Sanity gives teams structured content, Sanity Context, and deterministic access to the data that matters. Vercel gives teams the frontend primitives to make the experience fast, native, and shippable. Modern commerce teams are evolving beyond building as many audience segments as possible to building experiences where the storefront can understand what an individual customer is looking for in that precise moment. How many audience segments can our team support before the system becomes too slow or too complex? I think we should instead be asking how well can the storefront understand what this customer is looking for? If you’d like to hear the entire conversation and see the full demo, watch the Structured Series webinar /events/structured-series-ai-commerce .