Marketplace apps and websites are increasingly integrating artificial intelligence (AI) into their consumer and backend processes. It’s a trend that shows no signs of slowing down – and why should it? AI in marketplaces holds the potential to “radically simplify the selling experience,” said OLX co-founder and former CEO Fabrice Grinda. The Harvard Business review, in a 2018 cover story, called artificial intelligence “the most important general purpose technology of our era.”
Google and Facebook are both pushing hard to build software that learns on its own to recognize patterns within data. Google uses AI in its Hire product for recruiters and across its Cloud Talent Solution. Facebook in May updated PyTorch, the company’s neural network construction kit.
AI can more fluidly translate between languages, broadening the scope of a location-specific marketplace. A 2018 study by researchers at Washington University and MIT found that EBay saw sales jump by 10.9 percent after it added AI-powered translation for product listings in 2014. Better translation between countries is equivalent to reducing their distance from another by 26 percent, the researchers added.
Facebook is also heavily invested in AI-powered translation technology. The company said it is “now serving translations for a total of 4,504 language pairs” – for example, English to Spanish, Spanish to Russian – for a total of “nearly 6 billion translations per day.”
Broadly speaking, AI can make marketplaces smarter and more effective in four key areas:
- It can suggest prices to marketplace sellers by comparing uploaded images to those already in the marketplace’s database.
- Automatically categorize new listings, saving time and hassle for sellers.
- Optimize which listings get shown to which buyers in which order. The same goes for matching open job positions with potential candidates.
- Improve chat functionality by inserting bots into a conversation that becomes partially human and partially robotic.
Facebook shared details in 2018 on how it is implementing AI across its Marketplace app and website. One key area: providing sellers with price range suggestions. “If you wanted to sell your home office chair,” Facebook Marketplace VP Deborah Liu wrote last year on the service’s two-year anniversary. “AI might suggest ‘you price it between $50- $75’ based on what similar chairs recently sold for.”
Marketplace competitor LetGo has its own price-suggestion feature. Dubbed Reveal, it lets you point your phone at an item so that LetGo’s real-time image-recognition algorithms can suggest a ball-park price based on its database of hundreds of millions of listings. “It takes the guesswork out of selling secondhand,” a LetGo spokesperson told the AIM Group, by “automatically suggest[ing] a title, price and category.”
Singapore-based Carousell called its AI price suggestion functionality “Classified 4.0,” with an emphasis on reducing how much users need to type and search.
EBay-owned Marktplaats in the Netherlands does much the same, generating automatic price estimates based on user uploaded photos or videos.
Real estate classified giant Zillow has long used AI to generate its home valuations, which can been seen as a form of price suggestions. In January, Zillow awarded $1 million U.S. to the winner of a competition designed to improve the accuracy of Zillow’s Zestimates tool, The winning team’s new algorithm is 13 percent more accurate than Zillow’s current model.
AI price suggestions haven’t caught on everywhere. Classified app Shpock in Austria told the AIM Group that it had wanted to use AI, but found “no significant impact on lead generation [so] we withdrew the feature.”
Automatic categorization of listings
Along with AI-powered price suggestions, Facebook revealed in 2018 how it’s using AI to place listings in the right categories. Marketplace can now “automatically categorize [a] chair as ‘furniture’ based on the photo and description,” Liu wrote. Facebook classifies each listing into one of 29 categories, then gives the seller the option to change the suggested category.
The result is a nearly 100 percent increase in consumer engagement with listings. Automatic categorization has also reduced friction for sellers. “Previously, 7 to 9 percent of sellers abandoned the process without completing the listing,” Facebook added.
What’s coming next for listings and AI at Facebook? Automatic suggestion of listing titles based on comparable products and the ability to search for a product on Marketplace by taking a photo of “something out in the real world, such as headphones you saw in a store or a bag your friend is carrying.”
Seattle-based OfferUp also uses AI to suggest categories. The company’s algorithms look at title, description and historical data. Unlike U.S. rivals Facebook and LetGo, OfferUp doesn’t offer automatic price suggestions, although an OfferUp spokesperson suggested that’s in the works.
Australian automotive site CarSales uses image-recognition software to automatically place vehicles into categories. Its “Cyclops” technology also prompts sellers to upload additional images that CarSales’ software knows buyers will want to see – such as a picture of the back seat or the car’s stereo system. A seller might not think a picture of the back seat is so important, but CarSales machine learning and group AI leader Agustinus Nalwan said that “this is one of the most common features people search for when purchasing an SUV.”
CarSales has three other AI tools that improve the selling process for dealers: Tessa, which automatically verifies vehicle listings (a process that previously was done manually); Hawkins, which extracts the VIN and build date from a build plate; and Mavik, which can identify the make, model, body, fuel transmission and badge of a car from its VIN.
Istanbul-based Sahibinden won first prize at the most recent ICMA (International Classified Marketplace Association) Innovation Awards in Budapest for its AIbased car image-recognition technology. The Turkish company uses AI to categorize a photo of a car by make, model, year and body type, then displays the car listings that match the corresponding search filters. Sahibinden CTO Gökhan Ergül told the AIM Group that the tech “recognizes with over 93 percent accuracy.”
EBay bought Israeli startup Corrigon in 2016 for its technology that helps identify objects within an image to ensure that the image is correctly aligned to its corresponding product.
Classified site KupujemProdajem.com (Serbia’s top horizontal) is using AI for what it calls reclassification. “We had three people dealing with wrongly categorized ads, who spent on average two hours every day on the job,” CEO Bojan Lekovic told the AIM Group. “That adds up to about 1,000 man hours a year. These resources have been freed since 2017, when we launched our AIbased ‘classificator.’”
AI can, similarly, help remove duplicate ads (including those that are designed to fool the system with, for example, a slightly different image or text) as well as those that are illegal, counterfeit or discriminatory.
Styria Digital in Austria shared with the AIM Group data on just how useful image recognition can be to cut down on the number of clicks needed to list an item. Styria’s former head of data science Marko Velic (he left in 2019 to join Facebook) told the AIM Group that instead of 3.1 clicks to list an item, automatic categorization of images can reduce that to just 0.4 clicks while still “retaining 90 percent accuracy.”
Closely related to identifying images of cars and chairs is facial recognition, which some marketplaces are using, as well. Recruitment platform Lagou in China, for example, compares live images of a user’s face with the photo uploaded to the platform as well as the image on China’s state-mandated identity card in order to weed out fake listings. Alibaba does much the same to authenticate uses on its platform.
Matching across categories
“Say you liked your friend’s headphones and wanted your own; you could snap a photo of the headphones and Marketplace’s AI technology could recommend similar listings for sale nearby,” Liu explained how Marketplace is using AI to match listings to buyers. Ditto for outfitting your home: “Upload a photo of your living room and get suggestions on furniture to buy based on your layout and size.”
Facebook combines a ranking system that incorporates Lumos, the company’s image-understanding platform, and DeepText, Facebook’s text-understanding engine, to build an intricate product index that assists with product recommendations. Building in context from photos is key, “since the only text in many of these listings is a price and a description that can be as brief as ‘dresser’ or ‘baby stuff,’” Facebook research scientists Lu Zheng, Rui Li and David Kim wrote.
Austrian marketplace Willhaben has its own spin on product recommendations through AI. Fashion-Cam, launched in 2016, allows shoppers “to take pictures of their favorite dresses or other garments and Willhaben will automatically look for suitable fashion items on the platform that match in terms of garment color, pattern or shape,” the company’s MD Sylvia Dellantonio said.
It’s good for consumers and for a company’s bottom line, too. Davor Anicic, business development and sales manager for Styria, which owns Willhaben jointly with Adevinta, told the AIM Group that Styria’s visual search initiatives have increased ad views per visit by 312 percent, while time spent on the Willhaben app is up 183 percent.
Marketplace app OfferUp uses AI and “machine learning data science models to drive shopping personalization based on a number of inputs such as search behavior, item views, clicks, saves and historical behavioral data,” an OfferUp spokesperson told the AIM Group in an email.
Cars.com revamped its website last year to push buyers towards searching for cars based on lifestyle preferences (are you hauling kids to soccer practice or commuting on the expressway?) rather than by the traditional make and model. AI is used to make the matches, which received a faux-dating ad campaign called “We met on Cars.com,” complete with “likes” and swipes right and left (although we’re not sure it truly deserves the “Tinder for cars” catchphrase that Cars.com has taken to calling it).
The approach seems to be working: Cars.com has seen a 61 percent jump in return visits and a nine-fold increase in profile creation on the site.
Rex Homes is bringing AI to real estate matching. The company’s goal is to (mostly) replace real estate agents with artificial smarts. Software from Rex, which is a licensed broker in several U.S. states, first shoots out an initial batch of online ads to buyers it thinks might be in the market for a new home. As the first clicks come in, Rex’s algorithms figure out the commonalities between interested users. It then takes those criteria and searches out new leads based on the patterns it finds.
Rex also analyzes online behavior to identify potential home buyers and sellers. So, someone making a lot of trips to Home Depot might be fixing up a house to sell. Creepy? For sure. But investors seem to like it: as of 2019, the company has raised over $70 million U.S.
The marketplace vertical that has perhaps most enthusiastically embraced AI-powered matching technology is recruitment. Grupa Pracuj’s “Pracuj Select” sends passive candidate recommendations to recruiters by matching skills with employer requirements. The company, which operates Polish Pracuj.pl website in Poland, also uses AI to help candidates by parsing their resumes and profiles to recommend career steps.
Ringier-owned EJobs.ro – the No. 1 job site in Romania – launched its AI-based matching engine in January after two years of development. The company said the new technology produces matches that are five times more accurate than the legacy keyword-matching algorithm it had used for its first 20 years of business.
EJobs’s title-matching approach, built in-house, is similar to what Google’s Cloud Talent Solution API does, surfacing positions even if the search query is not a direct match or in a different language. On EJobs, a search for “sofer” in Romanian” returns results for “driver” in English and “sofor” in Hungarian. EJobs’ algorithm understands “the meaning of not only words, but of whole sentences or paragraphs,” Bogdan Badea, CEO of EJobs, told the AIM Group.
Ringier and Axel Springer’s Hungarian subsidiary, Profession.hu, has a new AI-powered product called Candidate Recommender that promises to find appropriate candidates for employers “within 10 seconds,” Imre Tüzes, head of business development at Profession.hu, told the AIM Group. The system is fully automated and takes into account which candidates are most active in the labor market to push them to the top of the list. Employers contact candidates through Profession.hu’s ATS. Candidate Recommender is being beta tested now with 100 Hungarian companies.
Zhaopin, the No. 2 job vertical in China, has seen a 30 percent increase in applications per candidate since implementing AI to match candidates with jobs, Jingfeng Li, the company’s CTO, told the AIM Group. The rate of applications marked “suitable” by recruiters is up by 40 percent. Zhaopin’s AI advancements were built in-house from its 11-year-old Zhaopin Lab division. That technology will also be used by Zhaopin’s Australian parent, Seek Ltd., which acquired the Chinese company in 2017.
Finally, Facebook is applying its AI smarts to its jobs initiative. A slide that appeared during a presentation at the 2018 JobG8 Job Board Summit, part of a presentation by RecruitingDaily.com executive editor Matt Charney, hinted that “Facebook’s plan is to build the matching AI + a firehose of jobs globally [with] direct ATS integration [and] one click apply.” An accompanying image showed a Facebook search box for London with a 150-km radius and four matching jobs.
Charney wasn’t able to disclose at the time where he got the screenshot, saying he was under an NDA. And Facebook’s usually chatty communications team clammed up when we asked them for clarification for this article.
AI for jobs can also determine which listings don’t get shown to particular users. This can be helpful – stopping people from seeing jobs they don’t have a chance of getting, which improves the quality of lead flow to recruiters – but also raises discrimination issues.
A study from April 2019 found that Facebook’s algorithms shared job positions along racial and gender lines, even if that’s not what the advertiser specified. The same issues are at the core of the U.S. Department of Housing and Urban Development’s lawsuit against Facebook for violating the Fair Housing Act.
Chatbots powered by AI
Perhaps the most visible (to the average user, at least) form of AI in marketplaces today is the proliferation of chatbots that can, depending on the app, either entirely automate a conversation or collect initial data before handing the conversation off to a real person.
Chaterix, a chatbot developed by Latin American property vertical Properati is an example of the latter: it fields queries that it directs to human real estate agents who close the deal.
Cars.com’s Ana Bot does the same for automotive dealers when the lot is closed. More than 50 percent of car shoppers “are connecting with dealerships during off-hours, between 6 pm at night and 9 am in the morning,” the company found. A live salesperson then follows up in the morning.
Automotive retailer Carvana is adding “a conversational AI platform” (read: a chatbot of its own) into its online customer support channels, the company said in May.
Carvana wants its automated support channels to be “so compelling that the customer may not reach a human, except when that touch really drives the experience forward,” VP of strategy Christina Keiser explained. Carvana bought startup Propel.AI last year for its Mojo chatbot and is in the early stages of testing now.
EBay’s ShopBot is one of the more veteran bots, acting as a virtual personal shopping assistant. ShopBot launched in 2016, initially as a way for Facebook Messenger users to more easily find items to buy on EBay. It was expanded in 2017 to Google Home. You can type, speak or take a picture to jumpstart a query.
Australian real estate portal Domain uses a Facebook Messenger bot to give home-seekers access to location-based listing information and property price data before they talk to an agent.
REA Group, which operates the Australian property site RealEstate.com.au is working on chatbots that can pass the “Property Turing Test.” “If you can go 10 minutes in a conversation with someone over a phoneline, and you can’t tell you’re talking to a computer, that’s the line in the sand where a computer is good enough to be classified as AI,” the company’s chief inventor Nigel Dalton told the AIM Group. REA hopes to release its tech by 2020.
Tokyo-based Mercari launched its Karakuri chatbot (Karakuri is Japanese for “mechanized robot”) at the end of 2018 to answer questions relating to packaging and delivery. The company says as the bot learns from customer feedback and logs, it will be able to handle more complicated queries. Mercari is ramping up its AI activities; the company has plans to recruit 30 members to its AI team in 2019. The new team has plenty to do: Mercari has amassed some 1.1 billion listings (text and images).
OfferUp mixes bots into live chat by using AI to automatically present buyers and sellers with safety tips and meeting suggestions. For example, “the word ‘Moneygram’ will prompt an automatic in-app message from OfferUp that says: ‘Sending money through wire services can be risky and isn’t recommended,” an OfferUp spokesperson told the AIM Group. When OfferUp detects that users are chatting about where to get together to make an exchange, the OfferUp chatbot “suggest[s] a nearby Community MeetUp Spot.”
Chatbots are particularly popular for recruitment classifieds.
TalkPush and My Ally’s Alex are third-party chatbots that job recruiters can use. TalkPush will ask for pictures, videos and documents from potential candidates. It collects audio responses and asks open-ended “tell me about yourself” questions. If a candidate passes the pre-screening, TalkPush will send an invitation to a live interview with a map to the company location.
My Ally’s bot works similarly, responding to candidate questions and scheduling interviews; it can even book conference rooms. Other recruitment chatbot vendors include Grayscale, AllyO, Mya, GoHire and JobPal.
SonicJobs, based in London, has a chatbot named Julie that screens potential employees specifically for hospitality industry clients. Julie asks questions about the basics (experience and location) but gleans its most relevant data from how jobseekers behave during their chat “interview” rather than strictly from what the job-seeker tells Julie. One leading behavioral indicator: how long it takes a job-seeker to respond to an ad.
There are also chatbots that allow users to buy and sell. In Turkey, for example, the Paym.es bot works with marketplaces such as LetGo and Sahibinden.com. Paym takes a 5 percent cut on sales after the first $100.
The proliferation of chatbots is part of a larger trend that TechCrunch has dubbed “app fatigue.” With so many apps to download, many users prefer staying in just a select few. chatbots bring additional functionality without requiring a separate download. They target consumers in the place they already spend most of their time.
That was what then-EBay vice president of engineering Japjit Tulsi told the AIM Group in 2017. “People are spending more time in fewer applications,” Tulsi said. “As these applications start to become platforms, our thesis was, can we be present where users are aggregating, instead of trying to bring them to our own website or app?”
Tulsi added that AI is key for EBay – and not just with chatbots. “When we thought what will be the key technologies in the next two to five years … No. 1 on our list was AI.”
AI can be used for much more than predicting prices, categorizing items and streamlining conversation, of course. Ultimately, it’s about predicting future behavior by analyzing data from the past. AI holds the promise of helping organizations make sense of the huge databases of unstructured user behavior they’ve recorded.
To wit: EBay’s 2016 acquisition of Israel-based SalesPredict with its “deep expertise in predictive analytics and machine learning” that will help EBay predict which buyers are likely to end up making a purchase, EBay’s VP and GM of structured data Amit Menipaz said at the time.
AI can also help classified companies flag items that are likely to sell fast – Cars.com’s “hot deals” functionality is an example here.
Even more enticing in the recruitment classified advertising space: Predicting when an employee is most likely to leave a job or when one from a rival firm would be most responsive to a job offer.
LinkedIn has long used predictive analytics to make job and connection recommendations. Facebook is essentially doing the same each time it recommends a new friend. India-based hiring platform Belong says it uses predictive analytics to canvas social networks, candidate profiles and other public internet sources to make candidate recommendations to recruiters. Amazon, PayPal, Cisco and Adobe are among Belong’s high-profile clients.
Predictive analytics is also finding its way into programmatic buying and selling of job ads. While programmatic advertising and real-time bidding have in the past been more about automation than AI, that’s changing, as can be seen at companies like Appcast, OnRecruit and Recruitics.
Clearly, there is plenty of AI activity across marketplaces, both verticals and horizontals.
“The jury is still out on whether current marketplaces will be smart enough to adapt to [AI], or if there will be newer, bolder entrants that will leverage AI to outcompete the current incumbents,” OLX’s Grinda pointed out. Either way, though, “the future belongs to AI-powered marketplaces.”