by Gloria Dawson , illustrations by Justin Tran Nov 17, 2015, 2:00p

We’re constantly talking to our tech, and it’s finally listening. You can ask Alexa, the virtual “assistant” who lives in Amazon Echo’s smart wireless speaker, what the weather will be like on your walk to work and “she” tells you. Say, “play the latest episode of Bob’s Burgers” to the new Apple TV remote and there it is, without hopping from Netflix to Hulu. Facebook M, a “personal assistant” feature available for select Bay Area residents, uses a mix of artificial intelligence (AI) and humans to answer questions and field requests. Not surprisingly, a Facebook representative told TechCrunch the most popular type of request is for restaurant suggestions and reservations.

New technologies in the restaurant space, like Luka and TextRex, allow users to ask questions conversationally and obtain digestible, actionable answers. Rather than retrieving dozens of results after selecting from layers of filters on sites like Yelp, these new technologies offer significantly narrowed down results and personalization that only a real (or almost real) conversation can achieve, or so the creators believe. Looking for a restaurant with plenty of vegan options, great steak, and decent lighting for a business meeting? A place to break up with your significant other? Just ask.

While Luka uses artificial intelligence to gather users’ preferences and create restaurant suggestions, TextRex, a text-based restaurant recommendation service from the review site the Infatuation, employs real humans to solve restaurant dilemmas.
TextRex’s technology is pretty simple; it’s not an app, but a phone number. And the texts are not unlike the ones Simona Sudit, one of TextRex’s human recommenders, sends to her friends when they are looking for a place to go out. Sudit got the gig after using the product herself. “I sort of assumed that it was an automated service,” she says. But then she got a personalized response after looking for a Paleo-friendly restaurant in the theatre district, and inquired about a job.

“We went into this like, let’s just see what happens,” says Chris Stang, co-founder of the Infatuation. He went to the Apple Store, bought an iPhone, and connected it through iMessage to a few computers at the office, answering user questions himself. Today, “rexers,” as Stang calls his restaurant recommenders, use a platform to streamline service and answer questions faster. But the limitations of texting remain. “You have to hope that people go to their phones and remember to text you,” Stang says. TextRex has several thousand users, says Stang, and he currently employs 15 part-time “rexers.” Like Sudit, most employees came from using the product, and then asking about a job.

On the other end of the technology spectrum sits Luka, a chat app that utilizes proprietary artificial intelligence. “It’s all really automatic; we don’t have any humans involved,” says Eugenia Kuyda, the co-founder and CEO of Luka, available only in San Francisco (for now). Although it’s an app, the interface is designed to look like a text conversation. (The popularity of texting and chat apps like WhatsApp have inspired developers to push this feature to other apps, like the health app Lark.) Luka, which launched about five months ago, has plans for more cities and more verticals in the lifestyle realm, Kuyda says. “Think of it as of a messenger for your lifestyle. You have Slack for work and Luka to figure out where to eat or what to do on the weekends.”
“There’s a lot of research that says that consumers like choice, but they only like choice to a certain degree.”

The information Luka delivers is gathered from sites with open information. A snippet from Yelp or a Zagat review might pop up in the conversation to convince a user that the restaurant Luka suggests is a good one. Michelin and Foursquare are also Luka sources. In the future, Luka could create revenue by working with restaurants to offer discounts or sponsored recommendations. Depending on whom you ask, these sources might lend credibility to the app, an important part in its long-term success, says Ben Lawrence, an assistant professor of food and beverage management at Cornell’s School of Hotel Administration. “Consumers are smart they will move to the most credible source of information they can find.”

The interface requires Luka to make sense of casual conversation. And, for the most part, it does, although Kuyda stresses, “It’s very, very, early right now, so you’re looking at a super general product,” she says. Future iterations of Luka might include the ability to chat with friends on the app. “Luka will take into account what everyone’s preferences are to get the best option,” Kuyda says.

When users first login to the app, Luka introduces itself as a robot who can find and book restaurant and cafes. “I’m also just an AI who’s learning by talking to you, so don’t be harsh on me if I don’t know something.” The app serves up text prompts like “I’ll try it,” or “I’ve already been there” or suggests dining options like “ramen” or “breakfast.” But users can also type free-form questions. When Luka gets confused, it apologizes: “Sorry for being a little weird from time to time. I just gotta learn!”

The chat interface allows Luka to build a personal relationship while gathering information on your likes and dislikes by asking lots of questions. Luka remembers users’ preferences and learns from each conversation, which could be great for creatures of habit, but “choices are dependent on a number of different factors,” says Lawrence. “Whenever you make a recommendation based on past behavior, it’s very complicated. You might be eating Indian food because your girlfriend likes Indian food, but then you break up.”
And limiting the number of recommendations is a solid strategy in the app space. “There’s a lot of research that says that consumers like choice, but they only like choice to a certain degree,” says Lawrence. “We’ve all had that experience where we’ve spent all day scrolling on the computer [looking for a restaurant and thought], I just want to eat.” Luka users rarely get more than 10 options, says Kuyda. “Anything over five is too many, but it kind of depends… if you want something really quick, you don’t want to browse through 50 places,” she says. “If you ask, ‘What’s the best Afghan restaurant in San Francisco?,’ it’ll give you just one.”

When giving restaurant results, “three is a good number to have,” says TextRex’s recommender Sudit. It’s hard to imagine getting only three results on Yelp, where Sudit was an intern a few years back.
“MealPal didn’t work out because the economics didn’t work out.”

But TextRex and Luka aren’t the first text-based restaurant recommendation products. The short-lived MealPal, created by serial startup guys Richard Gong and Mike Lee, also texted users restaurant recommendations. Gong, Lee, and other texters developed personal relationships with users, Gong says. “It really didn’t feel like a typical app experience for users; it’s not even like Facebook [messaging] where it could be an acquaintance. It’s literally someone you’ve given your phone number to.”

Gong found restaurants for Bay Area residents when the startup was functioning for a few months earlier this year. But ultimately, “MealPal didn’t work out because the economics didn’t work out,” Gong says. Most people didn’t want to pay, and those who did used the service a lot. Some users’ requests required Gong and others to search for restaurants for 5-10 hours per month, an unsustainable commitment when users were paying only $9.99 a month.

Rexers on the other hand just comb one site, the Infatuation, for restaurants, and its business model isn’t reliant solely on TextRex. For the foreseeable future, humans will be manning TextRex. Stang doesn’t think the technology is there yet, but he could imagine eventually “using AI to help us grow and answer questions faster.”

Kuyda, meanwhile, is imagining using AI to help find a restaurant before a user even asks. “The ultimate product we’re moving toward is the idea that you can really get to know the user better than they know themselves,” Kuyda says. “[Luka would] proactively decide what you might need tonight, so if it’s Friday night, why not show you that a table just opened up in one of the restaurants you really like or one you’ve asked about before?” So in the future, we might have an answer to ‘Where should I eat tonight?’ — without even having to ask.