Searchandising: Do your autosuggestions help or harm?
With over 80% of ecommerce sites using autosuggest, the search feature is as mainstream as U2. But can it sell as well?
Also known as typeahead, autosuggest recommends search terms, categories, products and even content as a user types. When done right, it can guide your customer to more relevant results faster. When done wrong, it’s disastrous.
“During testing, poor autocomplete suggestions sent multiple subjects on detours, often with a completely wrecked search experience or site abandonment as the result.” Baymard Institute
How can you make the most out of autosuggest and avoid its deadly sins? Read on.
The 4 R’s of autosuggest
Responsiveness is how quickly your engine triggers the suggestion menu after a customer starts typing. Rarely are suggestions helpful after 1 or 2 characters -- especially if your catalog is large and diverse.
Suggestions are rarely useful after 1-2 characters, especially when matches appear anywhere in the string
Tuning autosuggest to trigger after a minimum of 3 characters is less interruptive, and ensures more precise suggestions. If your tool reports it, check average keystrokes in your reporting to learn how many characters your customers type before clicking a suggestion.
Recall refers to how many suggestions you show. With autosuggest, more is not more. Asking the customer to slow down, stop their task and review your suggestions adds to cognitive load and slows decision making.
Keep suggestions tight
If searchers wanted to browse menus, they wouldn’t be searching! Showing too many options adds noise to the search experience. Your sweet spot is anywhere between 2 and 8 highly relevant suggestions (I recommend limiting to 5, repopulating for relevance as your user types).
More than 8 suggestions adds noise
A well-designed widget that chunks options between departments, search terms and products can get away with a little more than 8, so long as each section is clearly labeled and limited to 5 items each.
Autosuggest typically populates by matching terms to product data and/or search logs, along with a mix of weighted ranking factors such as search frequency. To improve relevance, global or search-specific ranking factors can be adjusted by a developer who understands ElasticSearch, Solr or Lucene logic or through business-friendly controls within the application (if available).
Bonus Tip: Gauge how well autosuggest performs on your site by tracking acceptance rate (% of times a suggestion is clicked out of all autosuggest menus shown), if your application provides this report. A low acceptance rate suggests a tuning exercise following the tips in this post is necessary!
Kill substring matches
Most autosuggest tools only match prefixes, or the first characters of a word. Occasionally a tool is set to match substrings that occur anywhere within a word.
In ecommerce, substring matches rarely match user intent. The chances that someone looking for “antacids” will be enticed by dog food, a gazebo or dining set is slim to none.
Substring rarely matches ecommerce intent for ecommerce queries
If you spot partial matching on your site, enlist the help of a developer to disable it. For edge cases such as mobile searchers who miss the first letter of a term (e.g. “ntacids”), it’s better to let the customer self-correct or submit their search and let fuzzy matching and spell correction handle the error on the search results page than to impact relevance for all users.
Merge terms and categories
Maximize relevance by merging duplicate categories and similar terms (e.g. synonyms, plurals and misspellings) to canonical suggestions. For example, Men’s Sale/Shoes and Men’s New Arrivals/Shoes should merge with Men’s Shoes. Searches for “Boys tshirts,” “boys tops” and “boys tees” should merge to a canonical term or Boys’ T-Shirts category.
Improve relevance by merging duplicate categories, synonyms, misspellings and plural terms to a single canonical version
If you spot variants and misspellings in your autosuggestions, your application is likely pulling solely from search frequency and not applying relevance logic. You may be able to improve this by matching solely to catalog data. If your developer can’t adjust settings to filter out duplicates and near-matches, consider upgrading your search application.
Variants and misspellings suggest your tool is matching to search history without relevance factors
Advanced tools such as Bloomreach will merge terms and categories automatically, as well as filter misspellings, irrelevant terms and anything with less than n=X results (you can set this threshold). However, you may still discover cases you want to manually override.
Many autosuggest widgets go beyond keyword suggestions to include products. However, this feature is only helpful when a customer knows what product they’re looking for, such as a search for a specific product number or SKU ID, brand+keyword or specific product name.
Product results are only helpful when customers already know what they’re looking for
Showing a few product hits for anything but a specific search is rarely more helpful than a search results page. Rather than guiding searchers to the most useful set of results, the engine tries to predict preferences. The larger your catalog, the less likely you’ll get it right.
In the Sears example below, guiding searchers toward category options would be far more helpful than individual products, such as girls’ onesies, boys’ onesies and even newborn onesies, 3-6 months, 6-9 months and 9-12 mos.
In most cases, guiding searchers to categories and attribute-based terms is more useful than individual products
What’s more, product suggestions are typically tuned to products trending by clicks or sales, which often doesn’t match trending search terms!
Product suggestions rarely match search term suggestions
Products can distract
Pictures are more eye-catching, and can distract users from more useful suggestions thanks to the “banner blindness” effect.
Product results can distract from more useful terms and category suggestions
Product results can misrepresent your offering
When only a small selection of results appear in your widget, your customer may conclude that’s all you have!
Customers may assume what you show in your widget is all you offer
If you do choose to show products in your autosuggest menu, make sure you provide context such as an “Our picks” or “Best sellers” label, and ensure they’re always in sync with search term suggestions.
Autosuggest usability tips
Widget design can help or hinder usability. Make sure you support scannability and quick comprehension of your list items.
Visually separate scope
If your catalog is large and diverse, department scoping is essential. Ensure scoped suggestions appear first and are visually distinct from unscoped items.
Visual scoping improves menu scannability
Resist the temptation to scope everything -- if your widget looks like this, you’re triggering suggestions too soon, or you don’t need scoping as product tend to live in discrete categories.
Don’t scope every suggestion
Highlighting matching terms in a different color makes it easier to scan what’s different about each suggestion.
Highlight matching terms helps menu scannability
For blended menus (e.g. with scoped results), apply a secondary style to highlight differences.
Apply a secondary style to highlight terms when results are scoped
Know your issues
The first step towards bulletproof autosuggestions is to audit your current experience, following the same process as site search auditing described in our last post. Find your top 20-50 keywords and test them on your site.
As you test your suggestions, look for recurring issues like irrelevant results for a given search term:
Audit your results for irrelevant matches to a given search term
And suggested products that mismatch the input and suggested terms:
Audit your results for irrelevant product matches
Watch for suggestions that don’t match products or categories. Unless your search engine handles functional and semantic queries well, you’re likely to have zero or very few matching results.
Terms that don’t match products or categories may lead to few or no results
Depending on your tool, you may be able to globally filter suggestions that point to less than n=X results (your desired threshold of matching products). Otherwise, consider blacklisting such suggestions from your top searches by volume with exclusion rules.
Throughout your audit, note adjustments you want to apply both globally and to specific searches.
Know your context
Your tuning strategy should be informed by your business strategy, taking into account your catalog, customers and merchandising goals.
- Is your catalog large and diverse or tight and focused? Which categories are impacted by turnover, and how frequently?
- Do your search reports follow an 80/20 rule or lean towards a longer tail?
- How sophisticated are on-site searchers? Do they search for single terms or more complex queries?
- Do searchers favor brand, attribute or category search? Part or SKU numbers?
- Do visitors often sort search results by star rating, price, best sellers or new arrivals?
- What is your autosuggestion acceptance rate? Are visitors who click autosuggestions more likely to convert or exit your site?
- What percentage of search is conducted from smartphones and tablets? How is their search behavior different?
- How can you use autosuggestion as “guided selling”? What categories and products are most desirable to boost given your merchandising strategy?
Know your tool
Determine what you can adjust yourself as a business user, and what you may need a developer’s help for in your respective site search application. Most search tools are built with Solr or ElasticSearch and share their respective back end logic, however “instant search” tools like Algolia may be limited in what you can tune.
Basic tuning factors
- Does your tool rely solely on search logs or does it include relevance ranking?
- What index factors (e.g. title, description, keyword) and attributes (e.g. color, brand, star rating) does autosuggest use to determine relevance? Can you adjust factor weighting or boost-and-bury attributes?
- Can you apply inclusion and exclusion rules globally and to individual keywords, categories and attributes?
Can you set min/max suggestions to display?
Advanced tuning factors*
- Does your engine use semantic or natural language processing to identify whether a query is navigational (find a category), usage (find by attribute) or specific-product related?
- Does your tool automatically merge terms and scrub irrelevant matches or misspellings? Does it provide a manual override?
- Does your engine use machine learning to dynamically adjust relevance based on seasonality, user behavior or user context? Does it provide manual override?
- Does your tool support advanced boost-and-bury for quality signals such as add-to-cart actions, sell-through rate, revenue per visitor, return rate or star rating?
- Does your tool allow you to set device-specific rules?
*Available in advanced, personalized search solutions such as Bloomreach
If your audit uncovered a significant amount of issues that you can’t correct within your current solution, it may be time for an upgrade.
Once you’ve nailed search tuning and autosuggestion optimization, you’re ready for the next level of searchandising: merchandising search pages with targeted content. We share our tips in the next and final post of this series.
Are you subscribed?