Leader Spotlight: Why every surface in ecommerce has a different job, with Vibhu Arora
Vibhu Arora is Director of Product Management, AI/ML at Walmart, where he leads AI Search and Personalization for Walmart eCommerce. He has spent over seven years at Walmart across roles spanning the full discovery funnel, from mobile search to ML-powered monetization, and previously held product roles at Facebook, where he led launch PM for Portal and AR/VR eCommerce. He holds a Master’s in Engineering (Systems Engineering) from MIT and dual degrees in Industrial Engineering from IIT Bombay.
In this conversation, Vibhu breaks down how he thinks about each surface in the ecommerce discovery journey and why getting them right requires treating autocomplete, progressive disclosure, and product presentation as distinct problems with distinct jobs to be done. He discusses the tension between helping customers express their intent and steering demand, the structural reasons why search results can explode rather than narrow as queries get more specific, and how Walmart overhauled its matching and ranking technology to fix it. He also shares a candid take on why the best monetization strategies follow great experience, not the other way around.
What autocomplete is actually for
How do you think about providing helpful guidance in autocomplete without overconstraining people’s ability to explore?
Autocomplete’s primary reason for existence is to help the user express their intent. That’s the primary goal.
We also have measurement in place and track metrics for this. One is coverage — do we offer autocomplete for a large set of keywords? The second, which is super crucial, is adoption: when we show the suggestions, do people actually like them and click on them? And the third is what we call MRR — it’s just a fancy way of saying the ranking of those suggestions. When we’re showing the suggestions, are people liking the top suggestions, or do they have to work really hard and find their way through suggestion four, five, six, seven? Having a system of evaluation and observation is very helpful.
Having said that, the nature of AI and machine learning systems has definitely evolved to an extent where secondary objectives can also be facilitated. Conversion or refinement are two examples of secondary objectives that could be layered on top of intent expression.
The objective can also change based on the moment in the journey. It’s a well-known industry fact that for the very first queries of a session — say someone wakes up, pulls out their phone, and it’s the very first query — previous session context is very helpful. As you go further in the session, what matters more and more is the activity you have performed within the session.
You could change the objective to be more focused on showing more diverse suggestions specifically when the intent is not very clear — let’s say someone has just started typing and has only typed one or two characters. In this scenario, it’s very fair game to show suggestions across different categories, different verticals, to maximize exploration. But as they start typing and the intent gets more solidified, the objective probably needs to adapt — moving more toward what people are trying to express, because the intent is now firming up.
Is there an example of an instance when an autocomplete suggestion shortened a journey prematurely?
Let’s say someone is searching for cucumbers. Should autocomplete’s role be to help them express this as quickly as possible and get out of the way? Or should it also shape the demand — “Hey, what about pickled cucumbers? What about cucumber spreads?” Different surfaces have different jobs to be done, and autocomplete is a surface where the primary job is to help people express what they want and get out of the way.
Once you take people to a listing page, it’s safer — for lack of a better word — to show more exploration. We can show different zones on the page related to directly matching intent, or alternate intent like pickled cucumbers or cucumber spreads. We try to index on letting people express what they want. The exploration piece we try to push down the journey a little bit, so that people can have an easy and quick takeoff with as little friction as possible.
Handling fuzzy and mission-based intent
How do you handle “fuzzy intent” categories — where the user isn’t searching for a product but for an outcome?
The example we love to use is drumsticks. Drumsticks is obviously the musical instrument. Drumsticks is also a green vegetable. There is also chicken and turkey drumstick. And Drumsticks is also a top-selling ice cream treat. Every person searching drumsticks has a different interpretation. It becomes challenging to put forward an experience when something vague like that comes in.
The game is, how much confidence do we have? In this scenario, it’s going to be very hard to have a high degree of confidence to lead with one of these categories. The way we handle this right now is to provide optionality at the moment. But it’s not an ideal experience. Ideally it should be a balance of optionality versus: do we have some data about the user? Can we look at some personalized signals and infer that person A is more about the chicken drumsticks, and person B is more about the ice cream Drumsticks, and then flip the optionality to a more guided experience that dynamically leads with a different ranking or ordering for person A versus something different for person B?
Back when LLMs were first introduced, we began thinking about how we could leverage this powerful technology and build meaningful experiences around it. One of the problems we always had was how to solve for customer missions versus sporadic purchases. The example here is, let’s say you’re throwing a birthday party for your kids — that’s the trigger for the shopping mission. People don’t usually go and search, “Hey, I’m throwing a birthday party with a Spider-Man theme, help me plan it.” People don’t usually search like that on shopping platforms.
We actually challenged that and said, “What if people could just enter their mission — how can I plan a Spider-Man theme party?” We built an experience around that where, just by inputting your mission, we’re able to deconstruct it and offer a solution so that you can solve for the entire mission: paper plates, balloons, T-shirts, party favors — all in one seamless interface. It’s not an ideal experience, but it’s an interesting take on how to handle some of these fuzzy intent scenarios.
Progressive disclosure and the too-many-results problem
As consumers narrow and refine, sometimes the problem isn’t a lack of options — it’s having too many. How do you think about progressive disclosure: what to show early versus later in the journey?
This is the crux of the search engine problem. The problem is not that there’s less data — the problem is there’s too much data. And the reasons are twofold, both systemic.
Compared to 20 years ago when assortment was very limited — maybe a few hundred thousand SKUs — for most retailers the assortment has now exploded. There are marketplace programs that have grown a lot, and by the nature of marketplace programs, they tend to bring a lot of tail assortment. We’re talking billions of items available in the catalog.
The secondary reason is more technical: the nature of a primitive search engine is that the more keywords you stuff into the query, the more results get generated — and that’s so counterintuitive. In an ideal world, the more specificity you add, the data should get more and more restricted. However, primitive search engines work on the concept of matching keywords in a relaxed way — either/or, if word A is matching or word B is matching or word C is matching. Because of the OR condition, the more keywords you add, the dataset keeps increasing.
To address the second problem, a couple of years ago we went on a journey we called “exact matches” — fixing the problem of specific queries. We up-leveled and completely overhauled our matching and ranking technology to adopt Google’s BERT algorithm, which is a state-of-the-art neural algorithm. We put this in our reranker — the second stage of sorting and ranking items — and it completely changed the game for us. We had massive, massive relevance gains unlike anything we’d seen for the longest time. This was a game changer.
Progressive disclosure — we also call it contextual refinement or contextual nudges — means providing a more guided and assistive experience that stays in context to where the customer is in the journey. We are investing very heavily in this space, using a lot of LLMs and query graph data to create these nudges, moving toward more progressive data sources compared to the primitive data sources we had in the past.
Do you find that users lose some control with these new progressive methods?
Yes. There is a trade-off. The more guided and assistive experiences we build, by definition, the user will lose some control. The thing is to do it tastefully and with intention.
It’s well known in the industry that most general merchandise traffic starts with a broad intent. Complex purchases especially — a coffee maker, air fryer, vacuum cleaner, bunk beds — for most people, it’s very hard to start anywhere more refined than that level of intent. It becomes very important to help someone navigate and refine because it’s going to be very hard to make a decision if they can’t think beyond “coffee maker.”
We would not want to disrupt the flow or provide a lot of navigational guidance when the user is already very specific in the journey. It needs to be contextual — showing up at the right moment, at the right point in the journey.
What you show on the product page
How do you think about highlighting product features without overloading people with too much information?
This is fundamentally rooted in the same root cause: too much data, too much to consume. Cognitive overload is a huge opportunity and problem, and endless scroll is not helping either — it’s a double-edged sword, because the list of items is just endless.
There’s no textbook answer to what the right balance is, and every retailer probably needs to run a continuous program of testing and learning. We continuously run a series of tests to keep finding the balance.
The layout you show information in — single column versus double column — can have profound implications in terms of people being able to focus versus being able to scan quickly. One of my favorite examples: a lot of the industry was following a pricing format that put the dollar price and the cent price at the same font size. But really, people are looking at the dollar amount. We built a hypothesis: if that’s how people are actually thinking, why don’t we present information in the same format — enlarge the dollar amount, reduce the cent amount? This is a standard practice in offline physical retail as well. And it did test positive.
Images are very, very crucial. Another powerful concept is badges — bestseller, or trust signals. Trust comes from social proof: if more people are doing this, then surely they must be right, and I can rely on the judgment of the crowd.
But even showing badges, it’s very easy to overdo and over-optimize to the point that it becomes detrimental. Imagine a page where every item has a badge — anything can be overdone and over-optimized, and then it stops being helpful.
Monetization and the experience cycle
Across the full discovery journey, where do you see the biggest tension between maximizing monetization and minimizing user effort?
From a customer perspective, they will probably be fine not seeing any ads. But if you’re running a business, your incentive is toward monetization, and one of the forms is advertisement.
There will be a tension between ad revenue and commerce revenue — what is known as GMV. However, there’s also going to be a right balance — a point where you can maximize both. If both ad revenue and e-commerce revenue can be maximized, that’s our ideal state. Everybody wins, the customer wins and the business wins.
Monetization and ads in themselves are not inherently bad. They actually help a lot in discovery, in providing diversity, and in keeping the ecosystem healthy. As more products and sellers onboard into the system — there are millions of items for any given intent — monetization tools like ads give an opportunity to brands that are onboarding to surface a genuinely amazing product to the customer who has a relevant intent.
One of the ways to control for this is to view the surface from the customer lens in totality. The whole page is composed of monetized capabilities and organic capabilities, and even if different teams have their own independent metrics, what can help is to think of something like whole page relevance. Is the page in totality still relevant? Yes, we’re showing ads at the top, in the middle, in the bottom — but together with the organic, can we measure that whole page relevance and trend it over time?
Is there a broader principle you hold onto when navigating that tension?
What really happens is that experience paradigm shifts occur. A new experience paradigm is created — think of the feed, or endless scroll, or the video module. The experience paradigm changes, and then monetization engines start over-optimizing it to the point that it gets over-leveraged. A few years later, the experience paradigm changes again — and it’s a never-ending cycle.
Take Google AI Mode as an example. Google for the longest time was built on its search results listing, but recently it moved fast and overhauled its entire experience toward AI mode and conversation. Now the monetization is following the ideal experience they think is the future. Similarly, Meta stories — the experience was evolved first, the question was asked: “What is the best experience for the customer?” The answer was, “Let’s try stories.” Once the experience started resonating, monetization followed, and it became one of the most successful monetized products in history.
The principle to keep in mind is to always focus on what is the best customer experience and keep evolving that — keep pushing the boundaries, thinking big, thinking bold. As it keeps evolving, the monetization will continuously keep following it. That’s how you keep a sustained winning program: you’re not getting old, you’re not getting stale, you are continuously innovating and then monetizing it as well.
What does LogRocket do?
LogRocket’s Galileo AI watches user sessions for you and surfaces the technical and usability issues holding back your web and mobile apps. Understand where your users are struggling by trying it for free at LogRocket.com.