Textio’s Data Sharing Points to Machine Learning Business Model
In the cutthroat game of recruiting, companies are constantly eyeing the competition. Textio, a Seattle startup using machine learning to help improve the performance and reduce the bias of job listings, is taking the lid off a massive data trove it has amassed over the last two years to help its customers see how they stack up.
Textio’s customers can now compare the expected performance of their job listings with that of competitors using a score the company’s “predictive engine” calculates through automated analysis of tens of millions of job listings. Anyone can search through thousands of employers in the Textio Index—the startup says any company with 20 or more public job posts in English is included—to see how a given company’s job listings rate on things like gender bias.
Textio has gathered the data underpinning these predictions to help recruiters write job listings that perform better. Co-founder and CEO Kieran Snyder says Textio created the index to help expose best practices.
“We think transparency around this is beneficial for everybody,” she says. “If you know who is doing well, then you can look at what they’re doing and you can consider how you might modify your approach to perform better.”
Textio went so far as to release a list of the top 50 large companies ranked by the predicted performance of their job listings, whether or not they use the startup’s service. Kay Jewelers came in No. 1. The only tech company in the top 10 was Twilio. Google, at No. 28, was the only tech giant on the list.
The rankings will be interesting to HR professionals and others watching the strategies of various employers to get an edge and make progress on diversifying their workforce. (One of Textio’s top selling points is that it tries to reduce the gender bias in language used in job ads and recruiting e-mails, which can lead to a more diverse applicant pool.)
But take a closer look at the technology Textio employs to create the rankings—and its business model for obtaining the high-quality data it needs to fuel its predictive engine—and you’ll see the direction many machine learning companies are moving, or will be soon.
Textio’s engine crawls the Internet in search of public job posts, ingesting and evaluating the language itself, as well as other attributes about each position, the hiring company, the job’s location, and the like. “If you’ve posted 20 jobs or more in the last year, then Textio has scored them all,” Snyder says.
Earlier this year, the company made a significant investment “to parallelize our architecture,” Snyder says, enabling the software to consider multiple potential improvements to each job listing simultaneously. “Now imagine the engine can not only evaluate millions of permutations at one time, but can do so over millions of documents at the same time,” she says.
(Those permutations power another feature I wrote about earlier this year, in which Textio provides suggested improvements to people’s writing in real-time—taking us a step closer to machines that can do the work of human editors and writers. Snyder will parse what it means to be a human writer in an age of intelligent machines, together with Seattle-based sci-fi author Greg Bear, at our Xconomy Intersect event next week. Details here.)
“The innovation here is primarily a scale innovation,” Snyder says of Textio’s architecture.
As with any machine learning system, it only works if you have enough high-quality training data. So not only does Textio need the job posts, it needs to know which posts were effective. To get that data, the company has implemented a data exchange program with some of its largest customers.
The program consists of more than 50 million job listings and counting that have been analyzed by Textio’s engine and tagged with outcomes data.
Companies share that data with Textio for two primary reasons, Snyder says. First, it helps improve the specific recommendations Textio makes for a given company’s job listings. For example, a recruiter at Cisco wants to know what’s working at Cisco, first and foremost, she says.
Second, Textio gives its data exchange customers a discount on its service.
“If you are contributing to making our engine better with your hiring stats, we think you should pay less for it,” Snyder says.
She calls this one of Textio’s great insights—one that “will become increasingly common for machine-learning oriented companies in the enterprise,” Snyder says. “Those insights from your data partners, and that contribution that they make, really needs to be reflected in your monetization strategy.”