Skip to content
InsideUX Design
User Experience 16 min read

Site Search: Why It Often Fails and How to Improve It

Many site searches only find what is named exactly as in the system. Why search fails so often and how six levers and a look at your search log can fix it.

Picture an online shop for kitchen supplies. Someone needs a new spatula, types the word into the search field and gets: “No results found.” Yet the shop carries twelve models. They are just called something else there: “turner.” The next attempt, “spatual,” ends the same way. The search field is empty again, and instead of trying a third time, the person switches to a search engine and ends up buying from a competitor.

Moments like this happen on websites every day. People who know exactly what they want like to take the direct route through the search field. At the same time, site search is a feature nobody on the project really owns. Design adds a search box, development plugs in a standard search engine, and that is that. Whether it actually gets people to their goal is something hardly anyone checks.

This article explains why site search fails so often, what research says about the relationship between search and navigation, and which levers you can pull to improve it. The kitchen shop serves as a running example throughout. It is made up, but typical.

Search or click: what the research actually says

A claim that keeps popping up in talks and articles: half of all people go straight to search when they land on a website. It traces back to a 1997 column by Jakob Nielsen, in which he wrote that more than half of users are “search-dominant,” meaning they usually head straight for the search box, about a fifth are “link-dominant,” and the rest switch between the two. According to the text, the numbers come from his own usability studies, but the column gives no details on sample or method. And it is almost 30 years old.

Jared Spool put the idea to the test in 2001. His team observed 30 people performing 121 shopping tasks, each on three to six websites. Not a single person always used search first, even though some described themselves that way. About 20 percent, on the other hand, used links exclusively. According to Spool, the website mattered more than the person: some sites nudged people toward searching, others toward clicking. The type of product played a role too. Books and CDs tended to be searched for, while clothing was more often browsed through categories. People frequently turned to search when the links on the home page offered no clear trail to their goal, in other words, little information scent.

Later articles from the Nielsen Norman Group (NN/g) take a more nuanced view. In “Search Is Not Enough” (2014), Raluca Budiu explains why search does not replace navigation. To write a good query, people need to know what they are looking for and what it is called. They have to recall terms instead of recognizing them, and typing takes more effort than clicking. On top of that, many site search engines are weaker than the big web search engines people are used to. Navigation, by contrast, shows what is there in the first place. At the same time, the article is clear that for people with a specific goal, site search is vital.

In practice, this means there are no “searchers” and “clickers.” The same person searches one moment and clicks the next, depending on how well the website supports each. Good search is therefore not a replacement for well-designed information architecture but its counterpart. For how to build the structure itself, see Introduction to Information Architecture.

Why site search fails so often

The Baymard Institute, which has studied online shops in usability studies and benchmarks for years, shows how widespread the problems are. In its 2026 Search UX benchmark of more than 170 websites and apps, 56 percent fail to adequately support users’ search needs. On desktop sites, 46 percent perform “mediocre or worse,” on mobile sites 58 percent and in apps 64 percent.

The breakdown by query type is particularly revealing. Among others, Baymard distinguishes:

  • Exact searches for a product name or item number: 12 percent of sites have issues.
  • Product type searches such as “spatula”: 20 percent.
  • Symptom searches describing a problem, such as “pan sticking”: 37 percent.
  • Feature searches such as “induction pan”: 39 percent.
  • Use case searches such as “pan for crêpes”: 43 percent.
  • Compatibility searches such as “lid for 28 cm pan”: 44 percent.
  • Abbreviations and symbols such as “28 cm” instead of “28 centimeters”: 54 percent.
  • Non-product searches such as “returns” or “shipping costs”: 66 percent.

The examples in this list come from the kitchen shop, the percentages from the benchmark. The pattern is clear: as long as people type exactly the term the shop uses, search usually works. As soon as they search in their own words, for a problem, a purpose or a service topic, many searches fail. And that is exactly how people search.

So the core problem is rarely the technology alone. It is a gap between the organization’s vocabulary and the mental model of the people searching. The following levers address exactly that gap.

Six levers for better site search

1. A visible search box with room for the query

People look for the search box where most websites put it. Back in 2001, Jakob Nielsen recommended placing search at the top of the page, usually in the right-hand corner, and offering it on the home page as a type-in field rather than a link. That fits Jakob’s Law: people expect your site to work like the others they already know.

A common compromise is a magnifying-glass icon that only opens the search field once clicked. NN/g examined this pattern in 2014. The result: the icon alone makes search much harder to find, and it adds extra steps. For desktop websites, Katie Sherwin therefore recommends an open text field next to the icon, at least 27 characters wide, so people can see and correct their query. On smartphones, hiding the field behind the icon is acceptable, because space is tight and the icon is widely recognized.

A small but effective detail: the query should stay in the search field after it is submitted. According to Baymard, 33 percent of desktop sites and 42 percent of mobile sites clear the field after a search. Anyone who wants to tweak the query then has to retype it from scratch.

  • Before: In the kitchen shop, a magnifying glass in the top right that expands a narrow field. After submitting, the field is empty again.
  • After: On desktop, an open search field with a visible label, wide enough for “lid for 28 cm pan.” On the results page, the query remains in the field.

Keep accessibility in mind, too: the search field needs a label that screen readers can announce, even if it only appears visually as placeholder text.

2. Autocomplete that actually helps

Autocomplete, meaning suggestions that appear while people type, is standard in e-commerce. In 2022, Baymard found it on 80 percent of the sites studied, but only 19 percent got all the implementation details right. Recommendations include no more than about ten suggestions on desktop and about four to eight on mobile, no scrollbars in the suggestion list, a clearly highlighted active suggestion and full keyboard support.

NN/g measured how often people use suggestions in a study of online shops: in 23 percent of the cases where suggestions appeared, participants selected one. That may sound low, but it also means suggestions shape a meaningful share of searches. Which makes the rule Kate Moran sets out in her article all the more important: every suggested query must actually lead to good results. A suggestion that leads nowhere is worse than none.

NN/g therefore advises against pulling suggestions unfiltered from search logs and recommends curated terms that reliably return results. Visually, it should be clear which part of a suggestion was typed and which part was added.

In the kitchen shop, typing “spat” now suggests “spatula,” “spatula set” and “silicone spatula,” plus “spatulas in Utensils” as a category, styled differently.

3. Forgive typos

People make typos, especially on smartphones, and they do not always know how a brand name is spelled. According to Baymard, 69 percent of sites do not offer relevant autocomplete suggestions for slightly misspelled queries.

Writing for NN/g, Page Laubheimer describes several techniques that help:

  • “Did you mean …?”: A spelling suggestion on the results page. If the original query returns nothing, the search can show results for the corrected spelling directly, as long as it says so clearly.
  • Stemming: The search reduces words to their stem so that, for example, “pan” and “pans” or “grater” and “graters” match each other. This increases the number of results but can reduce precision.
  • Words that sound alike: Methods such as Soundex match terms that sound similar but are spelled differently. This is especially helpful for proper names and brands.
  • Before: “spatual” leads to “No results found.”
  • After: “Showing results for ‘spatula.’ Search instead for ‘spatual’?” with the twelve matching products above.

Transparency is key. A search that silently looks for something other than what was typed confuses people when the results do not match their query.

4. Speak your users’ language: synonyms

The spatula from the introduction is a classic synonym problem. The shop names the product the way its wholesaler does, while people search with their everyday word. Laubheimer recommends synonym tables that map the terms people use to the terms in your catalog. Abbreviations and spelling variants belong there as well. Baymard’s example is that “mm” and “millimeters” should return the same results.

An often overlooked case is searches that are not about a product at all. People type “returns,” “shipping costs,” “voucher” or “store locations” into the search box because they assume search covers the whole website. According to Baymard, 66 percent of sites struggle with such queries. The fix is often simple: add service pages to the search index and surface the right page at the top for the most common terms.

How do you know which synonyms are missing? From the search log, more on that shortly. Methods such as card sorting also reveal which words people themselves use for your content.

5. Zero results should not be a dead end

Even good search will not always find something. What matters then is what the no-results page offers. According to Baymard, nearly 50 percent of sites fail to provide effective ways to recover from a search that yields no results. Many show nothing but “No results” and perhaps a few generic search tips.

Writing for NN/g, Kathryn Whitenton sets out three ground rules: clearly state that nothing was found, offer starting points for moving forward and avoid jokes that people might take personally. Baymard adds five strategies for online shops: feature related categories, suggest alternative searches, show personalized recommendations, display contact options such as phone or chat, and promote popular products or categories.

  • Before: “No results found. Please check your spelling.”
  • After: “We couldn’t find anything for ‘cast iron crêpe pan.’” Below that, suggestions such as “crêpe pan” and “cast iron pans” with thumbnails, a link to the Pans category and a note: “Looking for something specific? Our team is happy to help by phone.”

The no-results page is also a good place for microcopy that helps rather than lectures. “Check your spelling” puts the burden on the person searching. A concrete suggestion takes work off their hands.

6. Filters and facets that fit the search

With large catalogs, even good search returns lots of results. This is where filters help. NN/g distinguishes between simple filters, which exclude content based on one criterion, and faceted navigation, which combines several filters for different attributes, such as material, diameter, stove type and price. Facets are more flexible and help people understand what is available. But they take more effort to maintain, because every product needs the right data, and they add options. A simple filter is often quicker to understand.

Budiu describes facets as a bridge between search and navigation: they turn search into navigation by making the structure of the catalog visible. Someone who searches for “pan” and then filters by “induction compatible” and “28 cm” does not need to know these attributes in advance.

Two details make or break usability. First, active filters must be visible. According to Baymard, 28 percent of sites do not show an overview of applied filters. People then wonder why a product they saw yesterday is missing. Second, be careful with scoped search, such as a search that only covers the current category. Nielsen called this “dangerous” back in 2001, because people often overlook the scope. His recommendation: search the entire site by default and clearly state the scope on the results page.

The search log: an underrated research source

Every website with a search function collects, almost as a by-product, a dataset other research methods can only dream of: thousands of queries in the words of the people looking for something on your site. Writing for NN/g, Susan Farrell calls site search log analysis “one of the biggest missed opportunities in UX research.”

Her approach is manageable:

  1. Analyze the top queries: Export the 1,000 most frequent search terms from three to six months. Go through roughly the top 300 and group them by topic and synonym.
  2. Check zero-result queries: They reveal typos, missing synonyms and content people expect but cannot find.
  3. Read the longest queries: Full sentences reveal the needs behind short search terms.
  4. Look for numbers: Many numeric queries are item numbers. If search does not find them, a product number search is worth adding.
  5. Look at sessions: Several similar searches by the same person in quick succession suggest the first attempts did not return anything useful.
  6. Test the top queries yourself: Enter the 100 most frequent queries and check the first page of results. Where results do not fit, synonyms, manually promoted results or tuning can help.

Farrell recommends repeating this about every six months and reading outliers in context: an advertising campaign or a news article can push a query to the top for a while.

In the kitchen shop, the very first analysis surfaced three things. “Spatula” ranked high among zero-result queries, as did “return label.” And many of the most frequent queries were item numbers from the printed catalog that the search engine did not recognize. All three problems were solved with a synonym table, an indexed service page and a number search, without any new design.

One limitation applies here, too: the search log shows what people search for, but not why they fail or what they do next. For the why, you need qualitative methods. A usability test with realistic search tasks, based on the most frequent queries in the log, is a good starting point. Also keep in mind that search queries can contain personal data such as names or email addresses. Check with whoever is responsible for data protection what may be stored and analyzed.

Pitfalls: when search becomes the problem

Search as a crutch. Strong search makes it tempting to neglect navigation. But as described above, people need both. If you do not know the domain, you cannot write a good query, and if you want to browse, you do not search. If certain content is found almost exclusively through search, that points to a structural problem you can investigate with methods such as tree testing.

Overly aggressive corrections. Typo tolerance and synonyms can overshoot. Someone searching for an exact item number who gets similar numbers ends up with the wrong product. Test new rules against real queries from the log.

Paid placements in the results. Manually promoted results make sense when they help people. Turn them into ad space, and you risk people’s trust in your search. Laubheimer also advises against showing promoted results in a separate box, because people tend to ignore such boxes. It is better to integrate them into the regular results list.

Benchmarks as the measure of everything. Baymard’s numbers come from e-commerce. A government website, an intranet or a knowledge base has different content and different search patterns. The principles carry over, the percentages do not.

Conclusion

Whether people search or click depends less on their type than on the situation and the website. Good search is therefore not a replacement for good navigation but its counterpart for everyone who knows what they want. It rarely fails because of technology alone. Usually, it fails because of the gap between the organization’s terms and the words people use.

The kitchen shop in our example did not need a new search engine. A visible search field that keeps the query, curated suggestions, tolerance for typos, a synonym table, a no-results page that helps and visible filters were enough. And the most important foundation was in the search log all along: thousands of queries showing how people really search for spatulas.

Checklist

  • The search box sits where people expect it and is visible as an open text field on desktop.
  • The field is wide enough for typical queries and has a label that screen readers announce.
  • The query stays in the search field after submission.
  • Autocomplete shows a few curated suggestions that all lead to relevant results and works with the keyboard.
  • Slight typos still produce results, and corrections are clearly labeled.
  • A synonym table maps everyday terms, abbreviations and spelling variants to the catalog.
  • Service topics such as shipping, returns or contact can be found through search.
  • The no-results page clearly says nothing was found and offers concrete ways forward.
  • Filters fit the catalog, and active filters are visible and easy to remove.
  • Search covers the entire site by default, and any scope is clearly labeled.
  • The search log is analyzed regularly: top queries, zero-result queries, long queries and numbers.
  • The most frequent queries are tested regularly and used as tasks in usability tests.

Sources

Back to “All articles”

You might also like

Synthetic Users: Can AI Replace Real User Research?

User Experience

AI-generated users deliver interviews in seconds. But how do they compare to real people? A look at recent studies and at where synthetic users can still help.

Measuring UX: How to Use Success Rate, Time on Task and SUS Properly

User Experience

More page views are not a success in themselves. Which metrics really measure the quality of a user experience, how to calculate them properly with small samples and which thinking errors to avoid.

Usability Testing with Five Users: What’s Behind the Rule of Thumb

User Experience

Five users find 85 percent of usability problems, or so the saying goes. Where the number comes from, what follow-up studies have shown and how to plan test rounds that actually improve your product.