Latest Hype: Search Relevance
Search is having a moment. Again.
It’s not as if we haven’t always wanted search engines that bring us the most relevant results-read our minds and show us exactly what we want. It’s more that search practices tend to focus on what the content is about, not necessarily what is relevant to the search we’ve performed.
That’s because search technologies break down the words used in documents to determine a relevance score. But those methods don’t take into account the searcher’s intent -and true relevance is all about how well a search result meets a specific need.
For example, a search query for “alien movie” returns results on the Alien movie and franchise, movies about aliens, articles about the actors, producers, and writers of said movies, where to stream those movies, and so on.
In the past, search relevance was calculated by the frequency of keywords within a given document. But that ranking model doesn’t distinguish between the movie named Alien and movies with aliens in them. Nor do they show you what you may want to know about aliens in movies. In other words, we don’t know much about the searcher’s intent from the search phrase alone.
AI-powered search relevance aims to give us better, more personalized results-and we’re seeing a lot of hype around its promise. But how much of this is reality and how much is marketing? Vendors love coming up with the latest way to set their products apart from the rest of the field. But there are claims-and then there are claims that can be backed up.
Why Search Needed to Evolve
The relevance issue has been around since humans started archiving information, and it’s been a matter of research since computers could store information.
Information retrieval methods use natural language techniques of vectorization (a way of turning textual data into configurations of numbers that machine learning algorithms can then process and ultimately find patterns in), indexing (making a list of where every word or phrase can be found in a set of text), and term frequency (how often a word appears in a document helps determine what that document is about) to break down text.
Ranking algorithms on the web help determine what documents are more closely related to others and tackles the problem of relevancy by using web links between pages as a proxy for relevance. Enterprise search has relied on manual processes for ranking that involve:
- Search schemas that control how and what people use to search. This, of course, limits what searchers can find and relies on everyone to think along similar lines.
- Query rules return specific results based on a query. For example, sites can use query rules to promote sales products to the top of search results for a specific promotion.
- Synonym matching that allows users to type one term, and get results that use a different term that means the same thing. Ecommerce sites use this to make sure there are no null results for queries. To ward against this, merchandisers often manually pore over log files looking for misspellings and synonyms. These words and typos are carefully added to a thesaurus or dictionary.
- Persona-based rules create actions based on the type of person the site thinks is visiting. Writing enough rules to be useful takes a lot of time and requires constant maintenance, and everyone matching that persona gets the same experience.
There are more manual ways to tune search engines that involve poring over search logs for long-tail results, adding wild card search to external and internal sites, auto-completion, and more. These things do boost results, but they are onerous, and it’s impossible for humans to keep up with changing language and fashions.
Enter artificial intelligence-the latest innovation in search, triggered by the weariness of both searchers and search managers.
AI-powered search promises to improve relevance by automatic synonym detection, ranking algorithms that learn from feedback, and geolocation information. Perhaps one of the most interesting advances in search relevance is tracking user activity on-page and throughout the site during a session to understand what the user is searching for. This goes far beyond providing good search results to instead anticipate what people want and show that to them before they ask for it.
Enterprise search customers get excited when search vendors promise the ability to do this. Then, their expectations are that manual work will disappear while search gets better. Disillusionment invariably follows when they have to continue to manually tune, or when they can’t scale with more product or more documents, or when they do A/B testing and find there is nothing relevant at all!
What It Takes to Factor in Relevance
To be successful in determining intent, you have historically needed hordes of data. AI and machine learning draw conclusions from data, which can make relevant results for new searches, especially on ecommerce sites, complicated. By some estimates, nearly 70% of retail shoppers are “new”-either because it is their first time on a site, first time visiting on that device, or first time since a cookie expired (and soon to be worse, as people can avoid cookies altogether).
Having big data hasn’t been a big challenge for titans like Amazon, Walmart, and Wayfair — but it has been for virtually all other-sized companies.
Coveo’s research team has been studying this ‘sparse data resolves sparse data challenge,’ or what they refer to as the cold start problem. To counteract this problem, they look at what the user is doing within a session, akin to how a store clerk might observe a person in a store-picking up one thing, putting it down, choosing another. In this scenario, it’s clicks, page visits, and other relevant, timely information.
By plotting products in the aforementioned vector space, Coveo has been able to offer highly personalized suggestions-despite a retailer never having seen the user before. Even with no historical data to draw from, Coveo can start customizing and personalizing the experience starting with the user’s first click.
This is responsive search: following the user, collecting data dust along the way, and feeding that to the machine learning models so that it adapts as the user does. It studies all previous searches for a specific term and analyzes whether different pieces of content were more or less satisfying to the people who did the search.
AI doesn’t need rules to be tweaked-it adapts as it consumes more data, adjusting for who you are and what you like, and what has worked for other people like you in the past.
But there’s more to relevance than search. With search, you can’t be helpful until someone experiences a need or pain and asks for help. With recommendations, you’re not waiting for someone to experience a pain point or have a question. You’re getting ahead of it, predicting what they might need even if the person doesn’t know yet that they need it.
Machine learning again uses everything that a person does online to infer intent. This signal data includes what a person types in a query, what they retype, and what they eventually click.
Signal data can be historic or real-time. Historic data trains the machine on what others have done. Real-time data informs the machine of what the person is doing within a session. Both give signals about what a person wants, as well as what they may need in the future.
Search Relevance Is Continuous
Giving people what they’ve actually asked for instead of the words they’ve used to search with is the first step on the path to relevance maturity-the ability to give people the right content at the right time. The prescriptive phase of the relevancy maturity cycle is recommendations. And a fully mature relevance platform allows users to enjoy search experiences across platforms.
The goal for any company should be to deliver relevance in all its digital interactions-360° of relevance and the relevance maturity cycle is key to achieving that goal.
Unlike a point solution — e.g., a product that only fixes search-a relevance platform joins up digital experiences and delivers relevance wherever you need it.
It gathers interaction data from every digital experience a customer has, and converts that into relevant, intelligent experiences wherever the customer goes next. A relevance platform has the ability to join silos together and unify user experiences along an entire journey or lifecycle.
As Max Irwin said in the Haystack 2019 keynote, “Search is subjective. People are unique. Everyone is right.”
Originally published at https://blog.coveo.com on November 23, 2021.