Search engines are now recommendation engines


Categories: Search Tags: Search

Search engines increasingly blend recommendations with search results. I differentiate between a search result and recommendation: A search result is a “match”, a document that contains the search string. A recommendation is a suggestion, and may not be a “match”.

I’m mostly concerned about search engines showing related content when there is no information on a query, and situations where recommended results are interlaced with search results. A recommended result might be an ad, but that’s not necessary.

I wish there was a switch on Google to disable machine learning in Google search or to choose the search algorithm.

I think content recommendation engines (like Netflix search, Spotify search, also called recommender systems) have had an unfortunately outsized impact on traditional search engines, such as Google.

At this point it is hard to see a difference between the two types of “search”: in both cases1 the user enters a search string and receives a list of results from the engine. In both cases, the search or recommendation provider is trying to deliver the most relevant results to the user.

But the goals of search and recommendation differ. A recommendation engine is often trying to keep the user engaged. For example, If a show is not available, maybe the user can watch something else? Engagement is not a good goal for a search engine. If there is no information on a topic, then the user benefits most from knowing that there is a lack of information.2 How to act based on that lack of information is best left to the user, and not the search engine.

The results also differ. A search engine should return results that are closer to ‘matches’3, while a recommendation engine is expected to return search matches and items that are related to that search string (or unrelated items that the recommendation provider believes the user might be interested in).

Finally, the methods differ. A recommendation is usually the result of a machine learning system, a search is usually the result of a simple or explainable algorithm.

Further, because it is hard to measure the effectiveness of a search engine, search companies likely don’t know how to dial in the right amount of recommendations in a search engine. What is the right proportion of similar results that we should sprinkle into a results listing? The metrics we use for user engagement don’t point to search effectiveness.

You have certainly had the frustrating experience of searching for a movie on Netflix that is not in the Netflix catalog. If you search for “Simpsons”4 Netflix won’t say “Sorry we don’t have the Simpsons”, which is my preferred result. Instead the engine will display some other cartoon shows that they think teens and adults will like.5 This is reasonable behavior for a recommendation engine, and less reasonable for a search engine.

I feel like this experience is happening more often with traditional search engines, which is frustrating. If you search for a specific product on Google, you will be recommended the product’s competitors. This could be good because a more informed consumer is better equipped to make decisions. But more often it is annoying. Further, it’s impossible to know why a certain result ended up in a search results listing. Is the result there because it “matches” the query? Or because it is similar to the results from the query?6

When the stakes are as small as Netflix’s, it’s okay to return spurious data to the user, because the worst that could happen is that the user watches a movie they don’t like. But users are making much higher stakes decisions involving search engine results, such as where to work and who to date.

Search is not recommendation, it’s search. Obviously, search engines want to be as helpful as possible, and if introducing recommendations to search makes it more helpful for users, that could be good, but when search engines return results which are purely recommendations, not related to the search query, then that search engine is reducing it’s utility.

I think most users unfortunately want to see some recommendations in their search results. If one is searching for hotels in a city, we want to see competitors. But the line between helpful recommendation and spurious data is thin, and we don’t know where that line is. It’s probably in a different place for every user. Defining the difference between a search engine and a recommendation engine is now one of those “I’ll know it when I see it” things.

I don’t have the answers, I just have a low tolerance for recommendations in my search results7. I just wish that search was more like grep or PageRank and less like recommender systems. I wish search engine algorithms were more transparent about how results are selected and presented to users.

It does look like Google is trying to improve some aspects of search over recommendation.

  1. Sometimes a search string is not required for a recommendation engine.↩︎

  2. One can usually infer the absence of information, but it’s better if its explicit.↩︎

  3. like the results from grep, which is the most transparent and auditable search of all.↩︎

  4. Not sure if Netflix has streaming rights or not. Just an example.↩︎

  5. Of course, those results will be prioritized according to the shows that Netflix wants to promote given the current release schedule and business environment.↩︎

  6. And what does “similarity” mean when it comes to search results?↩︎

  7. Even those that aren’t ads!↩︎