Context not clicks: the insight rewriting the ecommerce playbook

Your customers are having conversations about your products somewhere you can't see. By the time they land on your site, the decision is mostly made. That shift sat at the center of one of the most practical sessions on the main stage at MXP San Francisco this year.

Headlining the insightful conversation was Gandharv Kalra, VP of Digital at Mejuri (a global D2C fine jewelry brand), and Artin Bogdanov, CEO and Co-founder of SUN (an AI-native personalized content startup). Two very different ecommerce businesses, one shared problem: the brands closing the gap between what a customer just did and what you show them next are winning right now.

Here's what they talked about.

Take jewelry as an example. A customer shopping for something to wear to a cocktail party used to search "gold bracelet" and scroll down the results page. Now they upload a photo of their dress to an LLM and ask what pairs well with it. They arrive at your product page with a shortlist already in mind, and a set of expectations your site either confirms or fails to meet.

Gandharv put it plainly: 

"The more semantic clarity and structure that we have in our content, that's what makes the content discoverable, understandable, and portable by the AI models."

That's a different goal than ranking on page one. Of course, he’s referring to generative engine optimization (GEO), and the playbook looks nothing like traditional SEO. Long-form articles that once boosted organic rankings are being replaced by FAQ pages, because that's what AI models quote. Clear, structured, question-and-answer content gets picked up and surfaced in conversational responses. Vague or generic copy doesn’t.

The practical implication: if your content team is still writing for search crawlers, it's worth asking whether those pages are also written for AI models that are now effectively acting as product advisors for your customers. "Clarity converts…Nothing beats clarity. The more clear you are in your content, the easier it is," he confirms. 

For ecommerce teams measuring how shoppers arrive and what they do when they get there, this shift matters a lot. Knowing which traffic comes from AI-referred sources and how those visitors behave differently is the kind of signal worth tracking now, before the volume gets harder to interpret. 

Both Artin and Gandharv described a version of the same problem: broad segments are easy to define and mostly useless.

Gandharv walked through how Mejuri layers it. At the top, prospect versus existing customer. Below that, first-time visitor versus someone who's been to the site before. And within that, what they clicked on to get there. An existing customer who clicked a specific ring in an email isn't interested in brand-building content. They want information about that ring, fast. Otherwise, surfacing the wrong product or content at that moment is a miss.

Artin's work at SUN pushes this further. His platform integrates with note-taking tools and other context sources to understand user’s interests, but what's happening in their life right now. When his own late-night conversation with his wife showed up as context in the system, SUN served him a personalized podcast on maintaining a healthy relationship while building a startup. 

No one programmed that response. The system read the moment and matched content to it. The underlying principle is the same as Gandharv's segmentation work: the question isn't just who the customer is. Rather, it’s what moment they're experiencing.

A few years ago, experimentation for most ecommerce teams meant A/B testing button colors and optimizing click-through rates at each step of the funnel. That model assumed customers moved through your purchase flow in a predictable order. They mostly don't anymore.

Gandharv described the shift directly: when customers arrive already knowing the price, the reviews, and the material quality from an AI conversation, the funnel assumption breaks down. "We don't know what part of the funnel they are in," he said. "Earlier, it was very straightforward. You land on a page. From there, you need to optimize for click-through rate." Now, the question is whether you're surfacing the right information for where that specific person actually is.

Mejuri is currently testing landing pages that rearrange themselves based on where a visitor came from and which product caught their eye. Someone arriving from a Meta ad sees a different version than someone who clicked a product in an email campaign—same URL, different experience.

"Most of our experimentation is around how we surface the right information at the right place." 

The throughline from segmentation to experimentation is the same: stop optimizing for generic conversion steps and start building for the specific moment a customer is in. If you're combining product analytics with marketing data to run that kind of testing, the case for unifying those data streams is hard to ignore.

Analysts are projecting that as much as 25% of all shopping could be agentic by 2030. The panelists were more grounded than that, and more useful for it.

Mejuri was recently included in a Google pilot which allowed customers to check out within Gemini. The program started with 1% of users; now Google’s scaling it to 10%. There are no conclusive findings yet from the first phase, which Gandharv argues isn’t a failure but rather the realistic picture of where things actually are.

More concrete was Mejuri’s SMS clienteling example. 

The jewelry brand’s stores record which stylist a customer worked with, why they’re buying, and the items purchased. When the data suggests a milestone moment is approaching (e.g., an anniversary the customer mentioned at purchase), an agent assembles a personalized message, references the original purchase, lists new relevant products, and signs it with the name of the stylist the customer already trusts. 

The stylist does nothing until the customer responds. At that point, a human takes over.

Is that agentic commerce? Gandharv's point was that the definition is still being worked out. But it's a useful frame: the agent handles the data assembly and timing; the human handles the relationship when the relationship matters.

Artin's counterpoint was worth noting: agentic commerce will probably absorb commodity purchases like toilet paper, reorders, anything where the decision is low-stakes. High-consideration purchases, where the act of choosing is part of the experience, will stay human-mediated for longer. 

"I don't think the second part will be agentic…in terms of you would abstract the decision-making and the joy of selecting items that you enjoy." 

The session closed on a note that felt more urgent than the agentic commerce speculation. Both Artin and Gandharv landed on the same idea independently: every new AI layer is another potential intermediary between you and your customer.

Amazon already demonstrated what that looks like at scale. Brands that sold through Amazon traded margin for reach, and many lost their customers' email addresses in the process, making personalization, re-engagement, and direct relationships structurally harder. AI shopping interfaces could do the same thing, at a new layer of the stack.

Artin's advice was direct: invest in your brand now. "The one element that can protect you from it is, have 60 physical stores, or have the right positioning." The brands that survive the next round of intermediation are the ones customers seek out directly, bypassing the AI interface because the relationship is strong enough that they don't need it to make the decision.

Context—rich, behavioral, emotional—is what makes that kind of relationship possible. Knowing your customer well enough to reach them at the right moment, with the right product and content, without an AI acting as the go-between. That's the goal.

The context the panelists described: behavioral signals, segmentation triggers, real-time experimentation results, is only useful if you can actually query and trust the data. Mixpanel's MCP for ecommerce lets you do exactly that: pull behavioral data, combine it with order and acquisition data, and ask questions in plain language rather than SQL.

See what’s possible, including use cases and sample prompts, with Mixpanel MCP for ecommerce.