At any given moment, you likely have dozens of marketing emails sitting in your inbox. “HUGE sale ends TODAY.” “Get yours now!” “SALE!” It’s as though your email is filled with dozens of desperate salespeople, all clamoring for your attention. But there isn’t necessarily a human behind them at all. There’s a good chance that some of these emails were generated by an algorithm that deploys individualized phrases based on what kinds of emotional pleas work best on you.
That’s what Persado does. The startup uses an algorithm to analyze a company’s audience down to the individual level, paying attention to what you’ve clicked on from that brand–data that’s already collected by the company and anonymized before it reaches Persado–and what emotional phrases are most likely to catch your attention. Are you attracted to words that indicate exclusivity? Or do urgent messages tend to catch your eye? Persado takes all that data and uses another machine learning algorithm to generate messages that may be more likely to make you click.
It’s a glimpse of the kind of personalization of language that could transform UX over the next few years, as AI becomes an integral part of research and design. And Persado is experimenting with real-world applications, applying what it’s learning to the real world, in use cases like subway PA announcements. The technology nods toward a sensor-filled future where individually targeted messages transcend the digital world and follow us into meatspace.
Though email marketing is its bread and butter, Persado’s AI isn’t just in your inbox. The company’s algorithms also write copy for text messages, advertisements across many platforms, landing pages, social media posts, and push notifications, which it says adds up to 2 billion impressions per month, for clients that range from Fortune 100 companies like Verizon, Microsoft, and American Express to household brands like Overstock.com, Kmart, Saks Fifth Avenue, Expedia, Sirius XM, and fantasy sports platform Draft Kings. In one campaign with the clothing retailer Lucky Brand, conversion rates increased by 127%. An anonymous case study with a Fortune 200 credit card company increased conversion rates by 410%.
Assaf Baciu, cofounder and SVP of product at Persado, says that the company is bringing the nuance of individual human communication back to mass marketing. “If we were face-to-face, I would strive to get signals to see if my message works, and I adjust the message so it hopefully inspires you to act,” he says. For companies trying to reach consumers, it can be tough to gauge the efficacy of its messages, or how they compare to subtle variations.
Persado’s technology plays on a fundamental truth of design with AI: That it should excel where humans tend to fail. “Writing messages day in, day out, and analyzing the signals of the feedback, is impossible to do for humans,” he says. “The machine can do that.”
Here’s how the system works. The algorithm was trained on the language of email campaigns, web pages, and search ads, each of which was broken down into variables: the product or offer description (“these shoes are on sale!”), the formatting (including capitalization, fonts, and emojis), the structure (paragraph, bullet points, and verb tense), the call to action (“buy this!” or “click here!”), and, most importantly, the emotional language. Using social psychology research around emotions, Baciu and his team identified five primary emotions that motivate people to click—joy, pride, trust, anticipation, and fear—and three emotional subcategories between each one. Each marketing phrase was tagged with the appropriate emotions, and the algorithm was trained to recognize the emotional intent of phrases using this data set.
By combining the trained algorithm with a client’s existing data about how their users have interacted with communications in the past, and testing different types of language on these users, Persado builds profiles that identify which emotions convince users most effectively. Then the company can use its second algorithm to piece together emotionally charged language that effectively targets messages to users based on their behavior in the past.
The key to all of this is data. The algorithm can’t simply generate “better” language for any old message, because it needs data about what a particular audience tends to engage with. “AI without context does not really work,” Baciu says. “There is no generic AI. We are still defining our knowledge with every campaign.”
This limitation has kept Persado squarely in the digital marketing industry, but Baciu says the company has aspirations that cross over into product design and user experience. Baciu posed two examples: What if your Fitbit knew exactly what to say on a particular day to motivate you to get off the couch and run a 5K? Or what if pharmaceutical companies or doctors could use an algorithm to individually target messages to users who haven’t taken their prescription drugs that day?
Persado is actively experimenting outside of marketing. The company recently completed a similar internal experiment on New York’s MTA transit system, rewriting the audio messages that notify riders that their train has been stopped by traffic ahead, or reminding them to not lean against or block the subway doors, and applying what it has learned about effective messaging to make these often annoying notifications a little more engaging–even pleasant. According to CityLab, the company changed the classic “Stand clear of the closing doors, please” to “Please be careful of the closing doors,” because adding politeness to the front of the phrase is nicer for listeners. “Stand clear,” which is apparently “technically worded,” is replaced with “be careful,” which is clearer, conveys importance, and is more emotionally resonant.
It was purely an internal experiment, and while Persado says the MTA does know about its existence, they’re unaware if the MTA will use the new messages or not. The company has no way of testing whether these changes are actually more effective, since it can’t carry out controlled A/B testing on one of the busiest subway systems in the world. But it hints at how optimizing language itself, based on troves of existing data, could manipulate listeners to behave differently.
To demonstrate, Co.Design asked Persado to try rewriting two possible headlines for this article using its AI. But again, since the copy doesn’t fit the normal use-case for the algorithm and it couldn’t test it using any data about Co.Design‘s audience, we had to settle for headlines that were informed by the company’s copywriting experience. One potential headline, “This Algorithm Tailors The Web To Your Personality,” became “Whoa . . . This Algorithm Knows Exactly What Makes You Click.” Persado told Co.Design in an email that this language taps into the emotion of “exclusivity.” And as for the “whoa”? Persado says that “our data shows that adding brief, introductory language—in this case conveying excitement—can set a more emotional tone and draw attention to what comes after.”
The company transformed another potential headline, “Made You Click: The AI At Work In Your Inbox,” into “Made You Click 😉 We’re Letting You In On The Secret AI Behind Your Inbox.” Yes, that’s a smiley face. Persado claims its data shows that the 😉 symbol “outperform[s] other variants 79% of the time in editorial campaigns.”
Both the MTA and Co.Design headline experiments demonstrate some of the hurdles Persado needs to clear before it can use its technology in broader applications. In terms of MTA, there’s no way to know if Persado’s language would actually make frustrated New York subway riders less angsty–especially when the MTA lacks the infrastructure to deliver personal messages to each rider. And per the headline experiment, it’s unclear if Persado’s emoji-fied, clickbait-y headlines would drive readers away in the long term, making them unsuitable for use in media organizations without a human editor to proactively make that decision.
More broadly, these factors are what’s stopping the company from moving from digital promotional messaging to language in offline user experience, whether that’s in the subway or in a physical store. First, there’s a lack of physical infrastructure that would permit the company to truly individualize its messages. But beyond that, there’s a lack of clean, objective data about how users are reacting to stimuli in the real world, making it difficult to train an algorithm to generate language or conduct experiments to see what kind of messaging is most effective. Persado’s algorithms need data to learn, and a control audience on which to test its ideas.
Yet more and more “smart” objects are colonizing our world, tracking their owners and harvesting data about their behavior. That includes cities, which are using increasingly connected systems to test and manipulate citizen behavior. “If the purpose was to get the trash in the trash can, we could probably work on that, assuming we can measure how many people actually put the trash in the can,” Baciu says. “Connectedness allows AI to emerge across many industries.”
Persado is firm that it wants to use its algorithm to generate promotional content that’s simply more in tune with human emotions, but its business points to the reemergence of language as a vitally important part of UX design, through AI-generated messages that are personalized to each person who looks at them. While chatbot-style applications promised to tailor communication to each user, this technology could be embedded seamlessly across platforms, whether through an app or a verbal interface like Google Home.
The potential of similar technology down the road could be powerful, and even a bit unsettling. It’s easy to imagine a more ominous vision of the future here–one where every piece of language you see, whether it’s on a store sign or an app, is tailored to your personality to convince you to buy, a la Minority Report. If your phone knows how you’re feeling at all times, every bit of language it broadcasts to you could be tweaked to suit your mood and capitalize on your emotions. It would mean mass manipulation on an unprecedented level—especially if these tactics aren’t disclosed to the consumer.
So next time you take a stroll through the torrents of promotional emails sitting in your inbox and you find yourself drawn to certain ones over others, remember: An algorithm may have made you click.