Tips for Tuning Site Search Relevance

Hey friends! This article was originally published on EcommerceIllustrated in 2016 - screenshots may be old but the theory stands! Will refresh with newer examples soon. Follow #ecomillustrated for updates

Unlike category browsing, visitors that use site search tell you exactly what they’re looking for (and, knowing what they want, are typically closer to purchase). How you handle their search queries can makes or break whether you win their business — and this all depends on how you’ve tuned your site search application.

Is your native site search enough?

Most commercial ecommerce platforms offer some degree of native site search out-of-the-box. While native search engines’ features and functionality vary, they?at a minimum?support indexing of product titles, descriptions and category associations, with the ability to autocorrect spelling, fuzzy-match (handle plural queries and stemming) and recognize synonyms (automatically or through a configurable dictionary).

Some native engines allow merchants to apply more fine-tuning, such as custom keyword tags, variant labels, autocomplete / autosuggest and relevance control through boost and bury rules.

Many enterprise ecommerce suites and third-party site search vendors offer advanced features like semantic matching, natural language processing, federated search,?searchandising?and personalization. The most advanced systems use machine learning to continuously optimize search relevance, and can incorporate Big Data sources into their algorithms.

Cadillac search tools are great if you can afford them, but fundamental site search problems are usually not symptoms of a?bad?search tool or lack of features, but of sub-optimal tuning of the core engine. Your site search?will suck, in some way, for some (or most) queries, if you “set it and forget it.” The good news is much, if not all of the following site search optimization tactics can likely be implemented with the technology you currently use today.

Tuning site search

First, check under the hood

How do steel-toed boots and baby clothes end up in search results for “white socks” on a site that has a full page of results when the Socks category is filtered by “white”?

This online retailer uses a robust site search vendor that is more than capable of returning killer results. The problem is the merchant hasn’t invested the time to understand the tool’s settings and apply the right logic to them.

When default settings are left unchecked, site search sucks across the board.

Exhibit B:

If you think your ecommerce platform’s native search sucks, it’s more likely that search logic requirements were not included in your implementation plan, and your systems integrator set you up with default functionality.?Upgrading your search tool to a new vendor will not “fix” anything unless you understand how search engines work.

Understanding recall vs precision

Recall?refers to the total number of search results returned for a given query. In the “orange long sleeved tee” search above, recall is over 2000!?Perfect recall?surfaces anything remotely relevant, sacrificing?precision?in the process. The example above appears to apply pure recall any product that partially matches any keyword in the query, such as “sleeve” and “long.”

Looking under the hood of your tool may reveal that default settings are returning results too high on recall or too tight on precision, and adjustments to search logic and index factors can improve your results.

Boolean logic

Precision can also be influenced by using “and” logic, rather than “or.” In the “orange long sleeved tee” example above, requiring product matches to include “orange” AND “long” AND “sleeved” AND “tee” would have produced a tighter, more relevant set of results.

Fuzzy logic

Most search tools include?fuzzy logic?to handle plurals, misspellings and other near-matches. This increases recall for a given search, and often improves results.

You may have edge case keywords for which you want to control the precision of results and override fuzzy match. For example, a search for the term “BLU” (cellphone manufacturer) surfaces two rows of BluRay players before showing a single BLU phone on Target.com.

Similarly, a search for “salt,” for example, shouldn’t return “salted” snack items or “bath salts.”

If your tool does not support keyword-level tuning, beware of optimizing for exceptions — think of your catalog as a whole and what the best matching strategy is. Fixing one keyword experience could break many more. Instead, you may be able to redirect searches for such keywords directly to their most relevant category pages — “Blu” to the BLU brand category, “salt” to the Salt category.

Mapping misspellings

Make sure to anticipate misspellings of any creative product or brand names that don’t exist in the typical English dictionary. For example, one of Levenger’s signature products is the “Briefolio.” But a simple (and easy to commit) transposition of letters returns zero product results.

Levenger can prevent this snafu by mapping anticipated misspellings to “briefolio” in its site search dictionary.

Also watch out for alternate spellings of English words, like “gray” and “grey.” Create a process where “gray” items are tagged as “grey” and vice versa to ensure consistent recall across variants.

Don’t let “zero results” be found!

Map numeric characters to their textual equivalents, such as “2 piece” and “two piece.”

Another gotcha to watch for is search terms that may be separated by a space or compounded, like “iPhone 6” vs “iPhone6” or “Goretex” vs “Gore-Tex.” If your engine is properly correcting for these variants, the number of returned results for each should match.

Index factors

You don’t have to trade-off relevance and recall if your tool supports weighting of?index factors?like product name, description, synonym tags, category associations and other metadata.

For example, you might weight {product name} at 200%, {product description} at 150%, {summary} at 150%, tags at 100%, specs at 100%, category associations at 75% and review content at 25%. In this configuration, products with {keyword} in the product name will be deemed 2 times as relevant as products only matching by specs or tags. You may also include custom attributes like “full price,” “house brand,” “new item” or “available quantity” in your strategy.

Semantic relevance

Semantic search aims to match queries with relevant results absent of an?exact?keyword match.

The more brands and manufacturers represented in your catalog, the more useful semantic search, as brands’ official product names may use variations of commonly used terms.

For example, most USB memory stick manufacturers dub their products “flash drive,” but customers commonly search for “memory stick,” “thumb drive,” “jump drive” as well.

On Best Buy, if you search for “flash drive” you get the best results.

Search for “jump drive” and…

A search for “thumb drive” is a bit “screwed”…

And “usb memory stick” just stinks.

The old school way to handle semantic relevance is to manually map synonyms to search queries in a thesaurus, and in some cases create custom rules. The larger your catalog, the more tedious this task.

More advanced search tools offer semantic matching out-of-the-box (including natural language processing) which can get you ~90% of the way there without manual configuration — but as with fuzzy logic, the remaining ~10% requires your domain knowledge of your own catalog’s brands and products to ensure semantic search doesn’t reach too far on recall for your most important keywords.

For ninja-level semantic matching, consider popular product names that you?don’t?carry, and map them to similar options you do.

For example, rather than return zero results for “fitband,” REI’s activity trackers are front-and-center.

Assign a resource to regularly audit your site search keyword reports to identify semantically related terms that may not appear in product names, descriptions, tags or other meta-data, and add them to your dictionaries and thesauri. For example, “vegan leather,” “hipster glasses” and “puffer jackets.”

Tuning autocomplete and autosuggest

Autocomplete and autosuggest are often used interchangeably, but they’re not exactly the same.

Autocomplete?aims to predict what the searcher is typing and quickly surface relevant matching results. For example, as a customer types “apple,” Zulily pulls up matching brands that begin with Apple.

A customer looking for Apple Bottoms may have, without autocomplete, typed “Apple Bottom” and received a less complete set of matching items. Autocomplete steers customers to the canonical set of search results.

Amazon uses?autosuggest?to guide the customer to a more focused set of search results. Showing relevant departments and related search terms helps the customer bypass filters, facets and page loads and find what they want quicker. (Related search terms are also a form of?autocomplete, as the searcher may have had the initial intent to submit a more detailed query).

In the example above, the engine only returns results that match the prefix of the terms being queried. As soon as the user types “apple bottoms p,” all other suggestions aside from “apple bottoms perfume” will drop (and more suggestions that match the new prefix string may appear in their place).

Contextual autosuggest

Contextual autosuggest?doesn’t complete the query by appending what the customer is typing, rather it returns items regardless of the position of the query prefix within suggested terms.

Autosuggest precision vs recall

As a sub-set of site search, autosuggest can also skew towards too much precision or recall, depending on how it’s tuned.

Consider the optimal number of suggestions in a dropdown menu. More recall isn’t better, especially if suggestions are more loosely related to each other and require the user to “think.” Showing irrelevant suggestions is as disruptive as returning irrelevant search results, and there are several tactics you can use to dial up relevance and keep autosuggestions tightly focused.

Trigger autosuggest?after?three characters

Many retailers return autosuggestions prematurely, before the searcher has typed enough characters to accurately recognize search intent. They may even attempt to autocorrect a two-character string, which can result in very random suggestions that are more distracting than helpful.

For example, Best Buy’s autosuggest attempts to autocorrect “te” to “tv” whilst matching products containing “te” in the middle of their keywords.

Tuning autosuggest to trigger after a minimum number of characters, with a certain typing speed, after a significant pause in typing or after a space is entered can make suggestions more relevant and precise.

Keep in mind, customers typing quickly on desktop keyboards may be so focused on their entry, they pay no attention to the dropdown menu. But mobile users who type much slower and work with limited screen space may find premature suggestions even more interruptive. Consider testing autosuggest separately on desktop and mobile.

Only match to prefixes

Autosuggest is only helpful if it matches searcher’s intent, and searchers don’t look for the middle of words. Someone looking for “ZZ Top” or “laptop” is unlikely to start his search with “top.” Rather, he’s more likely to be looking for a “top load washer” or “top freezer refrigerator.”

This is not to be confused with?contextual autosuggest. It’s okay to suggest attributes and modifiers before or after the input query. For example, a search for “sun dre” may suggest “Roxy sun dresses” or “strapless sun dresses” because “dre” matches the?prefixes?of words in the suggestions. But matching “Samsung dresses” (were there such a thing) would be far less likely to match the searcher’s intent.

Stick to head terms

Another way to return tighter, more focused results is to only suggest?head terms?(most frequently searched), and leave the long-tail, long-stringed results out.

Head terms aren’t always short in length. Longer queries may qualify as head terms if they have relatively high search volume. For example, “Harry Potter and the Order of the Phoenix” may receive far more search volume than “Harry Potter series” (especially immediately after release).

Filter duplicate matches

Avoid showing multiple variations of terms, like “omega 3 6 9,” “omega 3-6-9” and “omega 369.” Choose a canonical version and map alternatives to the preferred term.

Leverage index factors

Many tools support weighting of suggestions, which can improve the precision of items returned and show the best terms first. For your industry and catalog size, the optimal solution may include a mix of rules.

Optimizing site search

Tuning site search can pay dividends, and it’s worthwhile to adopt an ongoing, iterative approach to site search optimization.

Manually audit your top search queries

Begin with a manual audit your top 50-100 site search queries. Think like a customer – how relevant are results? And think like a merchandiser – are these the best results based on the business’ merchandising strategy? Is precision sacrificed for recall? Is fuzzy matching too fuzzy? Are spelling variants returning a consistent number of results?

Keyword research

Just as keyword research is never finished in SEO or PPC, mining for synonyms, slang, abbreviations and misspellings to add to your dictionaries and thesauri should be a continuous activity within your organization. Don’t stop with keyword research tools and your own site search reports — read product reviews, research how competitors name and tag products, and look for product names that you?don’t?carry that are similar to products you do.

Barney’s knows popular skincare brand La Prairie is similar in target market, price point and?name?to La Mer. Well played, Barney’s.

Develop site search tuning requirements

Search tuning strategies are rarely mapped out and considered in the requirements stage of an implementation, and systems integrators typically won’t ask for them. This is why so many retailers scratch their heads about why site search sucks.

Business leaders need to develop prioritized use cases and work closely with IT to ensure requirements can be supported by the current search tool, and that they won’t cause collateral issues (tuning for one thing, and breaking the experience in another way).

Test, test, test

After site search tuning has been implemented, it’s important to manually check results and validate that tuning tweaks are working. A/B testing may be involved if your technology supports it.

If you use autocomplete and autosuggest, ensure you’re tuning and testing these too.

Site search optimization is an ongoing process, after you’ve tackled your top 50-100 search terms, attack the next set. Your customers will thank you :)

Looking for more ecommerce design and usability tips? Follow the #ecomillustrated hashtag!