Why your search frustrates users and how to fix this as a Product Manager

Search is one of those features that quietly makes or breaks a product experience. When it works well, users find what they need effortlessly and often discover things they didn't even know they wanted. When it works poorly, frustration sets in quickly. If you're managing a search product, you're probably wrestling with three major challenges: finding the right balance with personalization, helping users navigate from vague ideas to specific results, and figuring out how to leverage AI effectively while actually measuring if it's working. Let me walk you through each of these.

Prefer talking through it with an ML Engineer? Get in touch!

Personalization in search feels like it should be straightforward. Show people more of what they like, right? And it does work. When you tailor results to someone's preferences and past behavior, they generally engage more and feel like the system understands them. But here's where it gets tricky, how personalized is too personalized? Think about what happens when personalization becomes too aggressive? Imagine you buy a few black t-shirts from a clothing site. Suddenly, every time you search for shirts, all you see are black t-shirts. You've been put in what people call a "filter bubble." The algorithm has decided you only care about one narrow slice of what's available. You miss out on discovering that perfect navy blue shirt or that interesting new brand. The system has optimized for immediate relevance but killed serendipity.

There's also the creepiness factor. When personalization feels invisible or invasive, people get uncomfortable. They start wondering what data you're collecting and how you're using it. This isn't just about privacy concerns. It's about whether users feel in control of their own experience.

So as a product manager, you're constantly balancing several tensions. You want results that feel relevant without being repetitive. You're trying to respect behavioral signals from users while also meeting business goals. Maybe your algorithm learns that someone always buys the cheapest option, but your business needs to surface higher-margin products too. You're weighing whether to show popular items versus truly query-specific matches.

The way forward involves a few key strategies. First, be transparent about what you're doing. When results are personalized, tell people. Use labels like "Recommended for you" or explain that you're showing items based on their history. Give people the ability to opt in or adjust their settings. Transparency builds trust and makes personalization feel helpful rather than mysterious.

Second, deliberately mix things up. Don't let personalization dominate everything. Blend personalized recommendations with trending items, new arrivals, or popular picks that everyone sees. You might have a "For You" section alongside a "Popular Now" section. Some products even let users toggle personalization on and off, which directly addresses the control issue.

Behind the scenes, tune your ranking algorithms thoughtfully. Start with solid text-matching and product relevance, then layer on behavioral signals carefully. You might set rules like ensuring at least some percentage of results are diverse. Different brands, different categories, items the user hasn't seen before. This prevents the black t-shirt trap.

Finally, use personalization to assist rather than override. Let it power smart autocomplete suggestions or filters, but don't hide results entirely. Give users sorting options and easy ways to refine what they see. When people feel they can easily adjust their view, they trust the personalized features more because they know they're not locked in. The real test is in your metrics. Are people clicking on a broader variety of items, or has their behavior become more narrow? Are they bouncing less and converting more? If you see diversity dropping, it might be time to dial back the intensity and reintroduce more variety.

Here's a common scenario: someone types "chocolate" into your grocery site's search box. Your system dutifully returns five thousand results. Now what? The user stares at pages and pages of products. Dark chocolate, milk chocolate, chocolate bars, chocolate chips, chocolate gift boxes. They're overwhelmed. Many will scan the first page, not see what they want, and either give up or struggle to figure out how to narrow things down.

This is the search journey challenge. Not everyone comes to search knowing exactly what they want. Many queries start broad or ambiguous, and your job is to help guide people toward specificity without forcing them down a particular path. Broad queries are underspecified. The search engine doesn't have enough information to know what subset of results you really care about. Ambiguous queries have multiple meanings. If someone searches for "apple," do they mean the fruit or the tech company? Both interpretations might be valid, and dumping everything related to either meaning into the results just creates confusion.

The key insight is knowing when and how to intervene. If a query is clearly broad or ambiguous, don't just show a flat list of thousands of items. Step in with helpful guidance.

For ambiguous queries, detect the ambiguity and ask for clarification. If someone searches "mixers," show them a quick prompt: "Did you mean kitchen mixers or audio mixers?" This saves them from scrolling through irrelevant results. Most users actually appreciate this kind of help because it shows the system is smart enough to recognize the confusion. For broad queries, proactively suggest ways to narrow down. When someone searches "shirts," show them clear options to focus their search: men's shirts, women's tops, dress shirts, t-shirts. Make important filters like size, color, and price range really visible. The idea is to make it effortless for users to "zig-zag" through their search. Starting broad, then clicking into a category, then applying a filter, progressively zeroing in on what they need.

Behind the scenes, you can also adjust how results are ranked for broad queries. Instead of showing the hundred most popular chocolate bars (which might all be similar), intentionally diversify the top results. Show a mix: milk chocolate, dark chocolate, different brands, bars and boxes. This gives users a sense of the range available in that first glance, which helps them decide what direction to go.

Watch for behavioral signals that indicate someone is stuck. Are they clicking back and forth a lot? Rapidly paging through results? Abandoning the search entirely? These patterns tell you the initial results weren't helpful. You might even gently prompt them in the moment. If they've scrolled through several pages without clicking, maybe show a friendly message: "Looking for something specific? Try filtering by type or brand."

One important thing to remember is that many users aren't search power users. They've often been trained by poor search experiences to keep queries simple and basic. They might avoid being specific because they expect the system won't understand anyway. Your job is to reverse this by making the search interface as helpful as possible.

Autocomplete is incredibly valuable here. As someone types "choco..." suggest "chocolate bar," "chocolate gift box," "chocolate cake mix." This not only speeds them up but shows them the system can handle specific intents. Make sure your search is forgiving. Handle typos, recognize synonyms, be flexible with language. Nothing destroys trust faster than getting zero results for a query that obviously should match something in your catalog. Over time, as you improve these aspects, users will start trusting your search more. They'll realize they can be specific and get good results. This creates a positive feedback loop where users provide better inputs, which makes your job easier.

The third major challenge is on the technical side, and it's where a lot of excitement and confusion lives right now. Search technology is evolving rapidly with vector embeddings, semantic search, and large language models. These promise more intelligent results, but they come with real pitfalls. At the same time, figuring out if any of this is actually working (measuring search quality) remains genuinely difficult.

Traditional search is based on keyword matching. It's straightforward but limited. Vector search uses machine learning to represent queries and documents as arrays of numbers in a semantic space. This means you can find things that are conceptually similar even if the exact words don't match. Search for "sofa" and get results for "couch." It sounds magical, and it can genuinely improve things, especially for natural language queries.

But it's not a magic bullet. You need to pick the right embedding model for your domain. A generic model trained on internet text might not understand your specific vocabulary or industry nuances. You'll likely need an experimentation phase to see what actually works for your content and queries.

Pure semantic search can also produce weird results. The embedding might consider two items similar because they're often used together, but users don't see the connection. You might search for outdoor cooking equipment and get patio furniture because the model learned they co-occur. To handle this, many teams use hybrid search. They combine vector similarity with traditional keyword matching. You ensure truly literal matches are still respected while using semantics to catch fuzzier connections. This usually produces more predictable, explainable results.

There are also practical concerns. Vector search can be computationally heavy. You need proper infrastructure with optimized indexes, maybe a specialized vector database. You might use vectors to retrieve a broad set of candidates, then re-rank that smaller set with more sophisticated logic. Latency matters in search. Users expect results instantly, so you have to be clever with caching, pre-computation, and optimization.

And you need to monitor for failure modes. Sometimes, embeddings learn strange associations from your data. Maybe users who search "bread" also click on bread-making machines, so the model starts blurring the two concepts. If weird results start appearing, you might need guardrails. Additional classification steps or context cues before search runs. Now let's talk about large language models. There's enormous hype about using LLMs in search. Having them parse conversational queries or generate answers from your content. LLMs are powerful at understanding language, but they have well-known failure modes. The big one is hallucination: they can confidently make up information that sounds plausible but isn't real.

Imagine asking an LLM-powered search: "Which running shoes have five-star ratings under a hundred dollars?" The model might generate a nicely formatted answer listing shoes that don't actually exist in your catalog. It's predicting based on training patterns, not querying your real database. This is dangerous because the answer looks authoritative.

If you do integrate LLMs, use them to enhance your search, not replace factual retrieval. Maybe use an LLM to rephrase queries or understand intent, but then use a reliable search index to get actual results. Or use LLMs to generate explanations or summaries of results that your core search has already retrieved. Keep guardrails in place. Techniques like Retrieval-Augmented Generation help by feeding the model real documents and asking it to answer based only on those.

Test heavily for failure cases. It's tempting to advertise an AI chatbot for search, but if it starts confidently giving wrong answers, you'll lose user trust fast. Be skeptical of hype. Use these tools thoughtfully, but keep fundamental search quality principles in place.

Finally, let's address measurement, which many product managers find surprisingly hard. Unlike a simple button where you track clicks, search is a complex interactive process. What even counts as success? Someone clicking a result? Making a purchase? Browsing for a while? There's no single perfect metric.

Start by tracking a mix of fundamentals. Search volume tells you how much the feature is used. Click-through rate on results indicates whether they seemed relevant. Zero-result queries show gaps in your coverage. Refinement rate (how often people search again or apply filters) suggests initial results weren't satisfying. Time-to-click can indicate efficiency.

When it comes to measuring discovery specifically, think about engagement depth. How many products do users view in a session that started with search? Are they interacting with a broader range of categories after you launch a new feature? Are they returning to use search multiple times, suggesting they trust it as a discovery tool?

There's an interesting tension between friction and exploration. In an ideal frictionless experience, users find what they want in one query and one click. But that might mean they never see the variety of your offerings. Sometimes a little friction actually encourages exploration. Showing alternatives or having users spend a moment with filters exposes them to more. What matters for your product depends on your goals. Is it a quick commodity purchase or a browsing experience where discovery adds value? You might measure this trade-off by looking at conversion rate versus average order value or items viewed per session.

Don't rely solely on quantitative data. Gather qualitative feedback too. Simple post-search surveys asking "Did you find what you were looking for?" can be illuminating. Usability testing reveals where people get confused in ways metrics might miss.

Invest in internal analytics tools. Build or buy a system that captures events like query submitted, which result was clicked, what rank it was at, whether the query was abandoned. Tie these to outcomes like purchases. Even basic logging of top queries and their click-through rates can highlight problems. Why does one popular query have terrible engagement while another performs well?

Whenever you make significant changes, try to A/B test them. Search performance is sensitive. A change might help some queries but hurt others. Compare user behavior side-by-side to know if something is truly an improvement. You might find mixed results where one metric goes up and another goes down, forcing you to make trade-offs and iterate further. Having an experiment-driven culture around search will help you mature the product over time.

Bringing this all together, remember that improving search is continuous work. You're always balancing user needs with business goals, relevance with discovery, innovation with reliability. The most important thing is keeping the end user at the center: are we helping them find what they want easily? Combine that empathy with solid data from your analytics, and you'll turn search from a source of frustration into something that genuinely delights users. It's a journey (appropriate, given that search itself is often called a journey), but it's one of the most impactful things you can optimize in a product.