EMAIL SEARCH IS A COMEDY OF ERRORS

EMAIL SEARCH IS A COMEDY OF ERRORS
My flight receipt is in there somewhere, right?

DECEMBER 13, 2023

Before a flight, I get a Pavlovian stress reaction—not from the prospect of hurtling at 30,000 feet in a concrete tube, but from my email inbox. Airline tickets sometimes get lost in an email abyss, requiring a few stressful minutes of frenzied searching when it’s time to check in. “I can’t find my confirmation for this flight in my email but I know I bought it 😭,” goes one tweet. “Gmail search is amazing,” another user posted just last week. “You can search something like ‘flight sacramento receipt 2023’ and it will somehow manage to serve up literally every email in your inbox that isn’t the receipt for the flight you just took to Sacramento.”

Searching your email can sometimes feel basically impossible. Typing in a mix of search terms goes only so far. At some point, it almost feels personal, like the software is purposefully not showing you a conversation you absolutely remember having. One Beyoncé fan posted that he couldn’t find a ticket to the Renaissance movie in his email and ended up buying another. Over Zoom, I asked Hamed Zamani, a search expert at the University of Massachusetts at Amherst, if he ever has trouble finding an email. He smiled and said, “Who doesn’t?”

When it works, search is magical. Google Search, for all of its problems, can pull information from an infinitely growing web in just a second or two, maybe before you even finish typing in your query. But searching a comparatively tiny inbox of your own emails is functionally hit or miss. The frustration of searching your messages feels particularly silly considering that Google makes both the most popular search engine and one of the most popular email clients in the world. Its competitors, Yahoo and Microsoft, also have both search and email products.
Our search woes do make perfect technical sense. Typing LAX to DCA ticket into Google Search and your email seems like the same thing, but it isn’t at all. The latter is called “personal search,” and it is an entirely different breed of information sorting from web search (anything you type into Google, or Bing, or whatever). It is genuinely harder to sort and rank 10 gigabytes of your emails than it is to sort all of the websites in the world.

Comparatively speaking, web search has it easy. It receives billions of search queries every single day, creating data that companies can use to hone their results. Even if you’re searching for something very weird—say, Teenage Mutant Ninja Turtles pickleball paddle—“probably dozens of people before you have already done that search,” Thorsten Joachims, a search expert at Cornell University, told me. Google or any other search company can track our behavior, seeing which search results users click on as they try to find the answer, and what websites they eventually end up on. It can track the different queries a user tries as they attempt to zero in on exactly what information they are seeking.

This information can help search companies rank the pages that show up the next time someone makes a similar search. “A search engine can harvest the work that they have put into it to make the search better for you,” Joachims said. Even if you horribly misspell your query, someone’s probably already made that misspelling before in the past. Google might show you results for the correct spelling instead. For instance, it knows that people who search for Napokean movie are actually asking for the Napoleon movie, and dutifully offers local showtimes. Search algorithms have so much data to work with that they get very, very smart at knowing how to find exactly what you need.

An email inbox can’t quite be searched in the same way, because algorithms are working off of far less data. “Your emails are yours and yours alone. My emails are mine and mine alone. And they are different,” said Zamani, who previously worked on search for Microsoft and has done research for Google.  The machines can learn only so much about how to rank the important emails in my inbox and yours, because they can see how you interact with these messages—which ones you open, which ones you ignore—but that’s not a lot of data. Sure, if you get the search term right, it can pull exactly what you’re looking for, and you can filter by date, sender, and a few other things. But you might have tons of emails from the same sender using the same few terms. Web search, by contrast, can rank results by leveraging a history of billions of clicks.

This is the paradox of search: Scale makes locating things much, much easier. Finding an email is akin to, say, searching for a pair of scissors in a stranger’s house. Finding Taylor Swift’s online merchandise store is like following the exit signs on Interstate 5 to Disneyland: Millions have made this journey before you, and the path is obvious.

Another major challenge for those hoping to build personal search tools is privacy. A lot of the tools that make web search so useful—auto-complete, spell-correction, follow-up questions (“Did you mean to search for Napoleon?”)—rely on machine learning. Machine learning from private data is risky. Emails contain all kinds of exceptionally private information—information that would be at risk of exposure were algorithms to try to train on it. Companies need to proceed with caution lest they accidentally spill private data from one person’s email to another’s. As Joachims put it, “if it learns from my private queries that for the string ‘thorsten joachims 123-’ the best autocomplete is ‘45-6789,’ then this may leak my SSN.” (Before you try anything, his Social Security number is not actually 123-45-6789.)
All of these challenges are real, but email search is improving. “Gmail is constantly working to improve the quality of search results, and AI advancements in Gmail are powering an even more accurate and personalized search experience, from contextual contact suggestions to relevancy-based results,” a spokesperson for Google told me. I heard similar things from representatives of Microsoft and Yahoo. Half of all personal-email searches are for a sender’s name, “because people tend to remember the sender’s name but not their exact words,” Josh Jacobson, the senior vice president and general manager of Yahoo Mail, said in an email statement. Because of this, the company highlights frequent senders.

For all the attention on how chatbots might change Google Search, Bing, and other search engines, recent advances in AI might be more promising when it comes to personal search. Yahoo recently launched an AI-powered “shopping saver” that finds discount codes and lingering gift cards in the inbox. Gmail users can now add a Bard extension to use Google’s chatbot to help find a lost email. Outlook has something similar with its Microsoft 365 Chat. Most email search is what’s called “known-item retrieval”—you’re looking for something you know exists. Sometimes it sits on the tip of your tongue, but you can’t remember the precise words that would pull up the email. Large language models, which pick up on language patterns, can perhaps better understand our queries, analyzing the content of our inboxes.
AI alone may not fix everything. In its 25 years, Google Search has become so popular and so instantaneous that it now defines our use of the internet perhaps more than any other website. Along the way, it has warped our expectations of what search can and even should be. Basically every search bar looks the same, whether they are filtering our chat history, a shopping website,  a music library, or the entire internet. We expect them to all work in the same way—without fully contemplating what is happening on the backend. The ultimate triumph of the world Google Search has built is that we now expect everything else to work just as simply.

Maybe us frustrated searchers have gotten greedy, expecting correct answers all the time no matter how complex the challenge. That’s not necessarily a bad thing. It’s how we ended up with Google Search in the first place, Zamani said, by asking for more than the earlier search engines that worked so poorly. “Our expectations are justified, and we should have those expectations in order to push this even further,” he said. Perhaps the road to technological progress is paved with a bunch of really annoyed posts about struggling to find airline tickets.
Caroline Mimbs Nyce is a staff writer at The Atlantic.