The Post-Click Problem

Something strange is happening in e-commerce.

We've spent the last few years building increasingly sophisticated ways to get people to click. Recommendation engines that predict what you want before you know you want it. Ad targeting that can find you anywhere on the internet. And now people research purchases by asking ChatGPT things like "I need something for my mom who just started hiking but has bad knees and hates anything that looks too sporty" — and getting genuinely helpful answers.

And then you click through.

Dropdown menus. Price sliders. A search box that only works if you type exactly the right words. "Sort by: Price Low to High." Pagination.

Welcome to 2015.

The expectation gap

Here's a scenario that happens millions of times a day:

Sarah needs a laptop for work. She's been doing her research the way people do research now — asking ChatGPT, reading reddit threads, maybe scrolling through some TikTok reviews. By the time she lands on Best Buy or Dell or Amazon, she has a remarkably specific mental model: something lightweight for travel, enough power for video calls and spreadsheets, ideally under $1200, definitely not HP because her last HP lasted 18 months before the hinge broke.

She types into the search box: "lightweight work laptop under 1200 no HP"

The results page shows her seventeen gaming laptops with RGB keyboards. One of them is an HP.

So she starts over. Clicks "Laptops." Finds the "Business Laptops" subcategory. Sets the price filter. Removes HP from the brand filter (if that option even exists). Sorts by weight. Starts scrolling through 340 results.

This is absurd.

Not because the technology to do better doesn't exist — it clearly does, because she just had a perfectly natural conversation with ChatGPT about her exact needs ten minutes ago. It's absurd because we've created a complete disconnect between how we bring people to the door and what happens when they walk through it.

Why this matters more than it used to

You could argue this has always been true. Catalogs were never great at understanding intent either. Sure.

But the expectation gap is new. And it's widening.

Three years ago, "AI-powered search" meant predictive text completion. Now it means actual understanding. When someone asks ChatGPT for laptop recommendations and gets a thoughtful response that accounts for their use case, budget, and stated preferences, they've just recalibrated what "good search" feels like.

Then they click over to your site and the recalibration snaps back. The mental model shifts from "I'm having a conversation with someone who understands me" to "I'm filling out forms and hoping for the best."

This isn't just a UX annoyance. It's conversion friction wearing a different mask.

The jacket problem

Let me give you another example, because this one's personal.

Last winter I needed a jacket for a trip to New York. Not a technical jacket, not a fashion piece — something that would look reasonable in meetings, keep me warm walking between buildings, and fit under my seatbelt without bunching up. Budget: whatever, but not ridiculous.

I knew exactly what I wanted. I just didn't know the words for it.

So I went to Nordstrom's website. Clicked "Men's." Clicked "Coats & Jackets." Got hit with subcategories: Bombers. Puffers. Peacoats. Parkas. Raincoats. Shirt Jackets. Vests.

Is what I want a puffer or a parka? I don't know. I'm not a jacket expert. I just want to not be cold while looking like a person who dresses himself.

Twenty minutes later I'd clicked through four subcategories, set a price filter, sorted by "Customer Rating," and still had no confidence that I was seeing the right products. Maybe the perfect jacket was in "Lightweight Jackets" instead of "Puffers." Maybe it was miscategorized. Maybe it didn't exist.

I ended up buying something from a targeted Instagram ad two days later. The ad understood me better than the site's search did.

The real cost isn't what you think

Most e-commerce teams measure conversion rate and average order value. Both matter. But neither captures what's actually happening when intent meets incomprehension.

The real cost is the silent abandonment.

Someone lands on your site with budget, intent, and urgency. They search. They get garbage. They don't complain. They don't fill out a feedback form. They just leave. They go back to Google. They try a competitor. They give up entirely.

You never know they were there. Your analytics show a bounce. They don't show that you lost a customer who was ready to buy.

Worse: the longer your search experience stays static while the rest of the internet gets smarter, the bigger this gap becomes. You're not competing with your competitors' search anymore. You're competing with the last AI interaction your customer had, wherever that was.

Why hasn't this been fixed?

Fair question. It's not like nobody's noticed.

Some of it is architectural. Most e-commerce search is built on top of systems designed for keyword matching and faceted filtering. Elastic, Algolia, Bloomreach — they're good at what they do, but what they do is fundamentally different from understanding natural language intent.

You can bolt NLP on top. Companies have tried. But retrofitting understanding onto a system built for matching is like teaching a calculator to have a conversation. You can make it sort of work. It's never going to feel natural.

Some of it is organizational. Search teams optimize for search metrics. Marketing teams optimize for acquisition. Nobody owns the gap between them. The moment of handoff — where a carefully nurtured lead becomes a self-service browser — belongs to no one.

And some of it is just inertia. Filters work well enough. People are used to them. The marginal improvement doesn't seem worth the engineering effort.

Except the marginal improvement is growing. Fast.

What good actually looks like

Here's what should happen when Sarah searches for "lightweight work laptop under 1200 no HP":

First, the system should understand the query. Not parse it for keywords — understand it. "Lightweight" is a preference for portability. "Work" suggests use case (productivity, video calls, maybe some light travel). "Under 1200" is a hard constraint. "No HP" is a negative brand filter.

Second, the system should ask smart questions only when it needs to. Not an interrogation. If Sarah's query is specific enough, just confirm: "Looking for a portable work laptop under $1200, no HP. Sound right?" If it's ambiguous, one question: "Are you looking for something with long battery life, or is power more important?"

Third, the results should be ranked by how well they match her intent, not just how well they match keywords. A $1250 laptop that's perfect in every other way should probably show up — with an explanation. "Slightly over budget, but 2 pounds lighter than anything else in this range."

Fourth, the explanations should be specific. Not "4.5 stars." Tell her why this laptop fits her need. "13-inch display keeps it portable. All-day battery for travel. Zoom-certified webcam. Users mention it runs quiet."

This isn't fantasy. This is how a good salesperson would handle the interaction. We've just never expected software to do it.

The two-sided problem

Here's what's interesting: the technology to do this exists. LLMs can understand natural language. They can maintain context. They can explain tradeoffs.

But most companies that try to add "AI search" end up with a chatbot bolted onto the side of their existing experience. "Ask our AI assistant!" And the assistant is... fine. It can answer FAQs. It can help you find store hours. But it can't actually search the catalog intelligently, because it's not integrated with the search system. It's a separate thing.

The real solution isn't a chatbot. It's rebuilding the understanding layer.

This means starting from intent. When someone describes what they want, you extract the structure: constraints, preferences, use case, context. You translate that into something your existing search infrastructure can use. You retrieve candidates. You rerank them based on semantic fit, not just keyword match. You explain why each result made the cut.

The retrieval system stays the same. Elastic still does what Elastic does. But everything before and after the retrieval is different. Intent in, explained results out.

Why this is Kinect's bet

We've been building something we call an "intent layer." The idea is simple: sit between the user and the search system, handle the translation, and make sure the results actually match what the person wanted.

We separate it into two agents. One that focuses entirely on understanding the user — extracting constraints, making reasonable assumptions, asking clarifying questions only when necessary. And one that handles retrieval and doesn't try to be smart about intent.

Most agent failures come from mixing these up. You ask a system to both understand the user and fetch products and explain results, and it does all three badly. By separating them, each can be good at its job.

The gap is the opportunity

Right now, every major e-commerce company is investing heavily in AI. They're building recommendation engines. They're personalizing ads. They're training models on purchase history.

Almost none of that investment is going into the search box.

The result is a strange asymmetry. The path to your site has never been smarter. The experience when you get there has barely changed in a decade. The post-click problem isn't getting better — it's getting worse, because everything around it is getting better while it stays still.

This is fixable. Not by replacing your existing infrastructure, but by adding a layer that translates between how people think and how your systems work.

The companies that figure this out first are going to win customers from the ones that don't. Not because their products are better. Because their search is.

And the customers won't even know why. They'll just remember that shopping there felt easier. Like someone actually understood what they wanted.

That's the post-click problem. And that's what we're working on.

Kinect is building intent infrastructure for commerce. If you're thinking about this problem, we'd like to hear from you.