Inside chatbots’ year of growing pains: ‘We’re at an inflection point’
Chatbots in apps like Facebook's Messenger and Kik have struggled to live up to the hype, but that might actually help them succeed.
A year ago, chatbots reemerged atop a mountain of hype that messaging marked tech’s latest paradigm shift, after messaging apps Facebook’s Messenger and Kik officially opened up to chatbots. A year later, the trend has slid down the mountain into its own uncanny valley. However, that slide may provide the momentum that chatbot makers, from brands to media companies to startups, need to make it back up the next mountain and stay there.
Chatbots have been on a downward slide for several reasons that seem to have snowballed together, according to executives at brands, agencies and technology companies. Messaging apps like Facebook’s Messenger have not done enough to promote chatbots, so not enough people have been using chatbots, so chatbot makers don’t have enough feedback to improve their conversational capabilities.
Roughly 78 percent of adults in the US have not even heard of chatbots, according to an online survey that Publicis Groupe’s Digitas LBi hired Harris Poll to conduct in November 2016. As much as that lack of awareness would appear to be a curse, it may actually be more of a blessing.
“The waters haven’t been muddied when that many people don’t know what chatbots are or have an expectation of what they should be,” said DigitasLBi’s senior VP of social strategy Jill Sherman. “Platforms still have the opportunity to reset and figure things out and watch the early use cases to see where the fall-off was, where brands went wrong, where users’ expectations were maybe a little off.”
When Messenger welcomed chatbots into the Facebook-owned messaging app last April, Poncho was one of the first in the door. The weather-forecasting bot in the character of a cat was also among the first bots to greet Messenger’s roughly 1 billion users. The first impression was a rough one.
“The first thing we learned almost as soon as we launched [on Facebook Messenger] is that users didn’t really know what they wanted to do with the bots, but to the extent they did they had an expectation that they would work like Siri and they could ask [the bots] to do anything basically,” said Poncho CEO Sam Mandel in an interview earlier this month.
Poncho’s predicament presaged the problems that have curtailed chatbots’ popularity. I tried out Poncho on Messenger the day it became available. After a short, character-rounding introduction — “Zzzzzzz. Purrrrrrrrrrrrrrr. Huh? What time is it? Oh, hi. I’m Poncho. A weathercat.” — Poncho asked me where I was, so it could tell me the local weather forecast. I could have shared my location using Messenger’s built-in location-sharing feature, but that would take me through three in-app menu screens, and wasn’t the point of chatbots to chat?
“Venice, ca,” I typed to Poncho.
“Hmmmm, that doesn’t look right. Where do you live again?” Poncho replied.
“90291,” I answered.
“Hmmmm, that doesn’t look right. Where do you live again?” Poncho repeated.
“Hmmmm, that doesn’t look right. Where do you live again?”
On my seventh try, I finally said the magic words: “Los Angeles.”
“Oh Los Angeles, CA! Right now it’s clear there.”
I stopped talking to Poncho until a few weeks later when I wanted to see if the bot had become any more conversant. It had, sort of. It still couldn’t understand “venice, ca,” but this time it was able to prompt me to ask for help, which would prompt a response listing examples of what I could ask Poncho.
In the interim weeks between my chats with Poncho, Poncho’s team had worked with Wit.ai — a company Facebook had acquired in 2015 that provides bot makers with artificial intelligence technology to parse human language — to make the bot more capable of carrying on a conversation about the weather. The goal was for Poncho “to understand the vast majority of weather-related natural language queries. We met that goal pretty quickly with the help of Wit.ai,” said Mandel.
Poncho’s other immediate goal following its debut, Mandel said, was to do “a good job of educating users about what Poncho did and making sure Poncho was useful enough.”
Too utilitarian to be used
“Useful” is a word that came up a lot in the interviews I did for this story, with a dozen executives at brands, media companies, agencies and technology providers. It’s a simple word with a complicated interpretation that seems to have vexed some bot makers. If a bot serves a single purpose, is it too utilitarian to be used? If it serves multiple uses, will people become overwhelmed? Or will they overwhelm the bot, expecting it to serve even more uses and become dissatisfied when it doesn’t?
As Mandel mentioned, people met Messenger’s chatbots with the expectation that they be as general-purpose as Siri. But Facebook seemed to set developers’ expectations that their chatbots be more singular in purpose. When Facebook CEO Mark Zuckerberg announced that Messenger was opening up to chatbots last April, his first examples were a 1-800-Flowers bot that can be used to buy flowers and a CNN bot that can be used to check the news. Those examples seemed to have set a precedent for the more than 34,000 bots that followed.
“There are apparently over 30,000 bots on Facebook Messenger, but most of them have a single utility. There’s no reason to go back or go back frequently enough to create a habit,” said Sherman. A Facebook Messenger executive was not available for an interview, and a spokesperson did not respond to emailed questions.
Bot makers had sufficient reason to play it safe early on with bots that stuck to a simple script. They didn’t know if people would intuitively understand how to talk with bots. Rather than overcomplicating the conversation, many opted for stripped-down, transactional exchanges. But even in simplicity there was nuance, which begot a need for artificial intelligence.
Poncho debuted without applying any artificial intelligence to parse what people were saying. After all how many ways are there to ask “how’s the weather?” The correct answer is more.
“You need to think about the fact that conversations can yield multiple responses. So if you think about everything as a happy path or flow, you’ll be missing out on that,” said Michael Zdepski, director of experience strategy and design at Dentsu Aegis Network’s Isobar.
How smart should a bot be?
In meetings with marketers to educate them about Messenger accounts for brands, Facebook has laid out three types of accounts: 1) live ones that require staffing a human to respond to inbound messages; 2) decision tree-based bots where there’s a fixed, automated flow for a conversation to take; and 3) artificially intelligent bots that are able to use natural language processing and other computational techniques to carry a somewhat broader conversation.
The chatbots that seem to have seen the most success are the ones that hover somewhere between #2 and #3.
When Lionsgate decided it wanted to create a chatbot in the persona of a “Power Rangers” character to promote the movie, the studio also decided that it didn’t want people to feel that talking to the bot was like filling out a multiple-choice questionnaire. So Lionsgate worked with Imperson, a technology company that offers tools to help bot makers script their chats.
“They can really script a conversation where there are 25 different ways that a normal person can react to a question,” said Lionsgate’s VP of digital marketing, Evan Fisk. “When they can write a script that accounts for all those different variations, it really brings a story to life and shows to our fans that they’re really able to test the limits of what [the bot’s protagonist] Alpha 5 can answer and what he can’t.”
Alpha 5 isn’t as conversationally agile as HAL 9000, let alone a human. But the bot is capable enough that people are spending four to five minutes at a time talking to it, according to Fisk. “And that’s across many, many users. Thousands upon thousands, hundreds of thousands, without getting into specifics,” he said.
No matter how many screenwriters companies may have hired, and how many paths they have plotted on their conversational flowcharts, their bots would never be able to map out all the places a chat may wind up. So bot makers have needed to develop ways to steer a conversation back on track without completely ignoring what a person might have said.
“You need to put boundaries or limits on the topic domain, otherwise the bot’s lack of intelligence or quirks shine[s] through pretty quickly,” said Zdepski.
In the case of the Power Rangers bot, when the bot doesn’t understand a message, it fesses up to the fact and maneuvers back to its script. Poncho makes a similar admission: “So, I’m good at talking about the weather. Other stuff, not so good. If you need help just enter ‘help.’”
“It’s better to have people fall in love with a bot and understand what it can provide you specifically rather than err on the side of not giving it enough personality,” said Pam Scheideler, chief digital officer at Interpublic Group’s Deutsch LA.
For Taco Bell’s TacoBot that Deutsch LA made for offices to place lunch orders over Slack, the brand and its agency had to decide whether to opt for “a more decision tree-driven experience” or one that uses natural language, which would be harder to pull off and offered more of an upside, said Scheideler. It opted for natural language, which meant “that even when the bot didn’t know the answer, we had to make it feel witty and on brand,” she said. (If you tell TacoBot that you’re drunk, for example, it’ll respond by adding a cup of water to your order).
“One of the reasons why the natural language piece is critical to every bot experience is because that’s what creates a connection between the user and the brand. We’re expressing the brand through that natural language,” said Scheideler.
Chat bots minus the chat
A year ago, chatbots promised a way for people to chat with brands, publishers and others like they do with their friends. But as documented earlier, that promise has been hard to put into practice. So over the past year, bot platforms have made it easier for chatbots to cut back on the chatting in hopes of lowering the bar for people to chat with them.
Within a few months of opening their messaging apps to chatbots, both Messenger and Kik had added options for bots to present people with preset reply buttons that they could tap to send a bot. Called “quick replies” on Messenger and “suggested responses” on Kik, the buttons served as training wheels for people and bots to learn each other’s language.“Assuming that people would pick up chatbot behavior just because it was a messaging app was maybe a little lofty,” said Sherman.
“With the chatbot there’s actually friction because people don’t really know how to talk to a bot, so there’s a lot of guesswork and that’s a bad user experience. That’s one thing quick replies solved for, to get people into a flow,” said Chao Liao, VP of partnerships at Snaps, a technology company that helps brands build chatbots.
“Facebook’s concern, which we totally understand, is that doing conversation well is really difficult,” said Mandel. “Everything for them is about big scale, and they’re really looking for a way to scale the bot platform and make it really easy for everyone to build a bot that people are going to use.”
“There’s this basic notion of conversational navigation, and where you see most of it going is this multiple choice type of thing, because as you start to expand to greater levels of depth, the complexity increases almost exponentially,” said Traction CEO Adam Kleinberg.
On Kik, “we definitely have more bots that use suggested responses than open text,” said Kik’s director of strategic partnerships, Jaclyn Ling. But just because more bots may opt for conversing through buttons doesn’t mean those bots automatically fare any better than those using normal chat interactions. “I couldn’t say which one generates more engagement,” said Ling. She cited a bot promoting the movie “Insidious” that used the standard free-text replies and averaged more than 60 messages sent per user.
“Open text works well if you have a strong AI back-end. Users get really tired if you’re constantly getting a response like ‘I’m sorry. I don’t understand.’ from the bot,” said Ling. “A lot of brands have a key end-goal in mind like [getting people to watch a movie trailer or visit a website]; that’s when suggested responses make a lot more sense to use because you can guide the users into the funnels that the brand is looking to get out of the bot.”
By providing preset replies, bot platforms and bot makers “are rightfully putting in other cues to help people navigate and understand the boundaries around conversations,” said Zdepski. He added, “In some cases, if you just need to make a simple selection, why should you force someone to type?”
Dick Clark Productions used these reply buttons for the chatbot it rolled out to Kik and Messenger to promote last year’s American Music Awards, which the company produced. And their ease of use likely contributed to the fact that more than 1.2 million messages were exchanged with the bot on Kik, including 11 to 13 messages sent per user on the day of and day after the show.
“Menu buttons help guide the conversation a little more. We wanted to make it as easy as possible. At this point in time it makes it more consumer-friendly to have different options available to them in addition to text. Just text can increase the friction of usage with a bot,” said Ariel Elazar, executive VP of brand, marketing and digital strategy at Dick Clark Productions.
These preset replies can not only make it easier for people to interact with chatbots, but they can also make them more accustomed to the conversational user interface, which is a departure from the graphical user interface that they’re familiar with on their computers, tablets and smartphones. But a move that Facebook made earlier this month risks killing the conversational interface in its crib.
In March, Facebook added an option for bot developers to disable people from sending text replies altogether and introduced a persistent menu that replaces the keyboard as the default mode of interaction. The persistent menu can feature a button for people to send a traditional text message to a bot — so long as those replies aren’t disabled — but they make it just as easy for someone to opt to wholly interact by tapping through a menu buttons instead of typing text.On the one hand, an entirely menu-based interface lowers the barrier to entry for people and brands almost to the floor.
“It may be one of those things where it’s a much easier sell to a brand to say we’re going to put training wheels on the training wheels. We’re just going to get you in a messaging environment and use your basic UX architecture and move your web experience into a messaging environment, just to get you into the environment,” said Sherman.
On the other hand, it raises the question of why someone would interact with a bot if it’s not so different from interacting with a website or mobile app.
“Using menu structures or more traditional UI structures, they’re almost admitting defeat in a way,” said Zdepski.
“I understand the motivation to simplify [the interaction] and make it graphical,” said Mandel. “But I think if you go to the extreme where the bot is basically a little webview on the Messenger platform, it’s really hard to make that entertaining or engaging to users. That means you have to win entirely on the utility front, and that’s challenging given the other options that users have.”
Why market through messaging?
Matt Tepper has been “saddened and disappointed” by the fact that chatbots haven’t gotten much traction. But he’s not willing to write them off just yet. As the chief strategy officer for North America at WPP’s Wunderman, he sees a need for bot makers to switch their strategies.
“One of the biggest misses that the bot industry has is that we haven’t been thinking about it as much as a CRM vehicle. If we start thinking about it as a CRM channel, we might be able to get a lot more impact out of it,” said Tepper.
By CRM vehicle, Tepper means that bots should serve as more of an intelligent intermediary between brands and their customers. When someone wants to check on an order, has an issue with a item or wants to find out different ways to use an item, the brand’s chat bot should be where they turn. And if a brand has a new product or feature that might appeal to subset of its current customers, its bot should be able to communicate it to them.
“How do you actually make the bots more useful to the end user by powering them with data of things we already know about them,” said Tepper. “It’s kind of like dumb social versus data-informed social. Dumb social you’re just saying stuff to people because they’re on your channel, whereas data-informed social you’re using the fact that you know something about them to tailor the message and approach.”
Tepper isn’t alone in seeing this potential for chatbots. Some brands are already jumping on the opportunity. For example, Nike’s Jordan brand worked with Snaps to roll out a Messenger bot that serves as a personal trainer by asking people to link the bot to their Nike+ account, which creates a way for the brand to use customers’ interactions with the bot to extend their relationship with the brand and vice versa.
Matt Gierl, head of growth at tuxedo rental brand The Black Tux, hopes that chatbots can serve as a safety net of sorts for potential customers who don’t immediately convert. Right now, The Black Tux can run ads on Facebook or Instagram to get people to visit The Black Tux’s site. Say, some will. But a smaller number will go to the next step of actually renting a tux, and the rest will be coulda-been customers that the brand will have to try to attract all over again from scratch.
But what if Gierl ran ads that point to the brand’s Messenger bot instead of its site? By doing that, The Black Tux can stay in contact with everyone who clicked on the ad to open a conversation with the bot, even if they aren’t ready yet to rent a tux. “Now we have the ability to message that person [using Messenger’s push notification] down the line with another marketing-related message,” said Gierl.
Lionsgate has a similar strategy in place with its Power Rangers bot. While it was created to push people to see the movie in theaters, “the bot will likely outlive the theatrical release,” said Fisk. “And that’s a good thing. We expect to keep it running once it’s out of theaters. We’ve already had chats with other departments within the studio about how to convert from pre-theatrical to the theatrical window and beyond.”
Chatbots “are a tremendous vehicle for any brand to develop a long-term relationship and then guide that relationship where they want to see it go, whether that’s pushing me to a sale, pushing me to a conversion event, giving me content options. All of those become available once you have that relationship with me in this environment,” said Doug Robinson, CEO of Fresh Digital Group.
If a bot talks in the forest…
In order for bot marketers’ dreams to become reality, people will have to actually use their bots, which means people will have to find out about their bots, which means messaging apps will need to do a better job making more people aware of their bots.
“Discovery is the biggest challenge. Period,” said Robinson.
Perhaps the biggest challenge within bot discoverability is that bot makers are in something of a catch-22. The messaging app that appears to offer the most ways for people to find out about bots may not have enough people, and the one that has tons of people lacks enough ways for those people to find the bots.
Kik has done a lot to promote the 20,000 bots on its platform. For starters, Kik features more than 200 of them in its in-app bot store that resembles Apple’s or Google’s app stores and makes it easy for people to browse bots they might like to try out.
Kik also has a feature called Mentions so that bot makers could allow people to invite bots into their conversations with friends. The feature not only reinforces a bot’s utility but also serves to make people more aware that a bot exists and how they can use it. For example, if I’m messaging with my friends about whether we should do a bonfire at the beach this weekend, I can type “@poncho los angeles weekend forecast,” and Poncho will send a message to the entire thread showing the forecast.
Bots that enable these inline conversations on Kik “are ones that perform really well and get viral growth at a rate much higher than those that do not enable inline experiences,” said Ling from Kik.
Kik has also worked closely with marketers like Dick Clark Productions, Lionsgate and Target to promote their bots on its platform. “One of the greatest things that Kik was able to provide as platform to our Target efforts was that discoverability. The Kik promotion itself in the app was a huge driver,” said Scheideler of Deutsch LA, which worked on the retailer’s bot.
The American Music Awards bot was available on both Facebook Messenger and Kik, but some content from the show could only be accessed through Kik because Dick Clark Productions “had a really strong partnership with Kik. They featured our bot across the platform. They were really supportive in the building of the bot and hand-held us through the experience of building the bot,” said Elazar of Dick Clark Productions.
“Kik was smart to make bots very central to their user interface and the way people use the app. They merchandised chatbots front and center so that their user base was exposed early and used them often,” said Sherman from DigitasLBi.
“We actually see a lot of people spending a lot of time chatting with brands in Kik,” said Liao from Snaps.
But it’s unclear to what extent those brands may be exceptions or the rule.
“Last time I caught up with [Kik] about this, they said they had a lot of bots but a shockingly small number of bots that had over 30 subscribers,” said Wunderman’s Tepper. [Ed. note: See postscript at the end of this story.]
That may owe to the fact that Kik’s user base skews young, so not all bot types may fare as well with the high school crowd. “We’ve seen a lot of game or lightweight entertainment experiences perform really well on Kik. I’d say this is mainly due to the fact that our demographic is a little bit younger and they’re looking for experiences to pass time with,” said Ling. And while Messenger has enabled people to buy products through bots, Kik has only tested bot commerce, she said, because it would need to “provide a payment system where a majority of our users can transact,” including the ones who aren’t old enough to have their own credit cards.
Another problem for Kik is that doesn’t have as many users as some bot makers would like. Kik’s “ability to grow the audience no longer exists. When you look at the numbers, it’s not scaling anymore. It stopped scaling six or seven months ago,” said Robinson.
A Kik spokesperson declined to say how many users the app has but seemed to wink at a number in a statement that described the app’s business as “strong and healthy.”
“We haven’t seen the same level of growth as we did in the early days — it’s much easier to claim 100 percent growth when you only have a couple hundred thousand users than it is to claim that same percentage when you have 300 million users — but we’re certainly continuing to grow,” according to the Kik spokesperson.
Meanwhile, Messenger is used by more than 1 billion people each month, but it hasn’t done much to actively market bots to those users. “Even very sophisticated people didn’t know how to find chatbots [on Messenger]. They didn’t make the bots front and center,” said Sherman. Instead of a bot store, to find bots on Messenger, people have to tap on the search bar, so that Messenger will preempt their search with a list of people they might be searching for, as well as a list of bots. But it’s unclear whether Messenger personalizes that bot list based on a person’s Facebook profile.
Given how much Facebook knows about its users, she said Facebook could have used that information to promote bots to individual users, but it has opted not to. “They’ve had a merchandising problem.”Bot makers can do their own merchandising through Messenger. They can create QR-style codes to paste as images outside of Messenger that people can scan to start a conversation with a bot. But they can also do that through Kik. And they can run ads on Facebook that link to a bot on Messenger. But there’s nothing stopping a brand from similarly linking a Facebook ad to their Kik bot.
Of course, there’s nothing stopping Messenger from copying Kik’s bot marketing strategy, using the smaller app as an R&D department in the same way that Facebook Inc. has used Snapchat.
For example, The Black Tux’s Gierl was intrigued by Kik’s Mentions feature as a way to serve brides and grooms trying to organize their wedding party’s attire, but he was wary of its audience size. “That might be a limiting factor. I’m also hesitant to build on something because it has a feature given what we’ve seen in the landscape with everybody just copying each other,” he said.
It’s unclear if Messenger plans to open a bot store a la Kik. But Facebook’s messaging app appears poised to roll out its own version of Kik’s Mentions feature. A few weeks before Facebook’s annual developer conference in April, where it’s expected to announce new features for bot developers, Messenger added the ability for people to bring their friends into conversations with other friends by @-ing them, like you might add someone to a conversation on Twitter or on Kik. The feature is called Mentions.
Messenger had already quietly introduced a Mentions-style feature for what appeared to be in-house bots, like @dailycute that sends an adorable photo when pulled into a conversation. It wouldn’t be that much of a leap for Messenger to extend those invocations to external bot accounts as well.
How long can the conversation carry on like this?
Poncho may have made a rough first impression when it initially debuted on Messenger, but the weather bot has been able to recover. Now available on Kik and Viber in addition to Messenger, on average around 60 percent of the people who start a conversation with the bot pick up the conversation again within a week. “We’re very proud of that. That’s comparable to app engagement numbers,” said Mandel.
While Mandel admits that “the biggest disappointment of the past year is that more people are not using bots,” he’s quick to say that it’s too early to call the conversation over. And he’s not alone. “I actually think bots have a good future for us in marketing,” said Tepper. And Scheideler said, “Clients are very interested in it, and their interest continues to grow.”
Businesses may be interested in bots, but the issue remains how interested people are when so many remain unaware that chatbots even exist.
“When you think of chatbots, we’re at an inflection point,” said Fresh Digital Group’s Robinson. “We have to figure a couple things out or we risk losing momentum. And that’s bad for everybody.”
Postscript: This story has been updated to reflect new information from Kik, which says it has more than 200 bots in its bot store. (Our article originally said “more than 100.”) Kik also disputes the quote from Matt Tepper saying that a “small number” of bots on its platform have more than 30 subscribers. A Kik representative says it currently has 10,000 bots with more than 30 subscribers.
Opinions expressed in this article are those of the guest author and not necessarily Marketing Land. Staff authors are listed here.