Big Data Builds a Better "Help Wanted" Ad

Bloomberg View , August 20, 2015

If you want good applicants to respond to your job posting, write it as if you were talking to actual human beings. Banish “proactive” and “value added.” Change “interface with” to “work with.” Scuttle “synergy.” Replace “the [job title] will” and “the successful candidate will” with “you will.” Don’t repeat the same phrases, as though applicants might forget the job title or the company mission. Read what you’ve written out loud to make sure you haven’t pasted in the same paragraph more than once. Better yet, start from scratch instead of remixing last fall’s ad with a few bullet points from three years ago.

No, this isn’t a bunch of out-of-touch advice from a pointy-headed English major who knows nothing about derivatives, Python, HIPAA requirements, or supply-chain optimization. Except for the part about reading aloud, it isn’t coming from me. It’s coming from big data.

Using an enormous and continually updated database of job postings and results, Textio, a Seattle-based startup, applies machine learning and natural language processing techniques to see what works -- and what doesn’t -- in job ads, whether for candidates in general, specific industries, or male or female job seekers. (Coming this fall is a service to help companies write better e-mails to prospective hires.)

“We measure. We look at listings that have been posted in the past and how they’ve actually done,” explains CEO and co-founder Kieran Snyder, who earned her Ph.D. in natural language processing and then spent a decade building product and engineering teams at Microsoft and Amazon. “How long did that role take to fill? Was it closed in a week, meaning you had a good listing and you found a great person? Or did it take eight months and you just couldn't attract the right person?”

The company analyzes which of more than 40,000 distinct phrases attract or repel applicants. It also checks structural characteristics such as sentence length, adjective use, the balance between “we” and “you” phrases, or the mix of bulleted lists and solid paragraphs. The results provide a fascinating peek into contemporary business.

For starters, they reveal that a lot of companies write truly awful ads because hiring managers forego clear, simple writing in favor of a weird mix of stiffness and jargon. “People in general have a really false idea of the level of formality that helps them in a job listing,” says Snyder. “Just write like you really write. It will probably be better.”

Using a free two-week trial, I plucked listings from various job boards and ran them through Textio’s word-processor-style interface to see how they rated on its 0 to 100 scale. Only two out of a couple dozen broke 50, and some scores were so low that Snyder professed shock: A technical editor posting from ICANN, the Internet governance group, rated a measly 2, while a RAND Corp. listing for an information assurance manager scored a 4, as did a Variety Latino ad for an online managing editor. They suffered from corporate clichés (“best practices,” “stakeholders,” “complies with”), bossy language (“must have”), and tons of repetition. (Who in RAND’s department of redundancy department wrote, “U.S. Citizenship is required to obtain a security clearance. U.S. Citizenship is required to obtain a Top Secret/SCI security clearance”?)

The highest score I found was the 76 earned by a Bloomberg LP ad for a senior program manager for global philanthropy and engagement. That doesn’t mean all the company’s ads are especially well-written. One for a software infrastructure developer scored a 1. (Bloomberg Beta, a venture capital fund backed by Bloomberg LP, was an investor in Textio’s $1.5 million seed-funding round.)

Snyder says a typical listing starts with a score between 35 and 65 and improves by 25 to 30 points with about ten minutes of editing. In the company’s A/B tests, she says, revised ads generally see a 15 percent to 40 percent increase in applicants. Much of that editing is just cleaning up the prose.

Textio’s analysis also demonstrates that surfing the fashions in business lingo is as important to hiring as following color trends is to the garment business. Eighteen months ago, the phrase “big data” made tech job postings perform better. Now everybody uses the term, so it’s neutral. Eventually, it could become a negative cliché.

In finance ads, the common phrase “gather insight” has gone from positive to slightly negative. Applicants now respond better to “collect and analyze data” or “practice data science.” Other good finance terms include “mathematical problem solving,” “real-world data,” “empirical,” “statistical analysis,” and -- breaking up the quant fest -- “intuition,” which has shifted from neutral to positive in just the past few months.

When you think about it, that anomaly makes sense. “How do you make sense of all these statistics?” says Snyder. “You use math, but you probably do need to bring some intuition to bear.” The underlying math reports only that a term attracts fewer good applicants. Classifying whether that’s because the phrase is corporate jargon or “too directive” requires editorial judgment.

Textio poses a long-term challenge to organizations. Suppose everyone took its advice. How would employers then stand out?

One answer is to adopt terms that turn off some audiences and excite others, thereby fostering a distinctive corporate culture. Take “dog-friendly workplace.” The term comes out as “neutral” only because it’s so polarizing that the two sides cancel each other out. (Snyder calls herself “neutral on dogs,” but when she arrived at Amazon, where she says there were nearly as many dogs as people, “it was the thing that drove me batty when I started.”) Textio doesn’t yet have enough data to see which industries or demographic groups find dog-filled offices a plus, but companies that promise them clearly want one type of employee and not others.

Similarly, on gender, Textio’s big data approach identifies subtle patterns that traditional checklists, with their emphasis on avoiding macho language (“mission-critical,” “take no prisoners,” “leave it all on the field,” “rock star”), miss. Who knew that women prefer “impact people” -- so violent! -- to “affect people”?

But some of the language Textio flags as gender-biased because it attracts male or female applicants in disproprotionate numbers reveals real differences in how people see themselves and their work. Women prefer to “lead” or “develop” teams, men to “manage” them. These are distinct skills, and throwing them all into the same ad because you want gender balance may not fit the job.

Or take “unrelenting,” a term Snyder herself embraces. It skews male, much to her chagrin. “If you replace it with ‘persistent’ or ‘dedicated,’” she says, “you can remove the bias.” Again, these terms don’t really mean the same thing. Compared with nose-to-the-grindstone “persistent” or slightly passive “dedicated,” “unrelenting” suggests energetically resisting an opposing force. A persistent worker won’t give up on a tough task and a dedicated one will put in long hours -- both desirable qualities -- but an unrelenting person is more likely to found a company. Maybe the real bias isn’t about gender but personality. Some workplaces may in fact be better for those who see themselves as dedicated, while others fit the unrelenting.

Textio reports results. It doesn’t tell employers what to do with them. But it does hope to foster more inclusive workforces. “The tool doesn’t judge,” says Snyder. “When I talk to people in finance or technology, they want to hire more women. When I talk to people hiring nurses or elementary school teachers, they want to hire more men. Both of those are cool.”

In the future, however, textual analysis may become an arms race, with employers trying to out-maneuver one another to target desirable subgroups while potential hires use the same software to suss out which employers offer compatible cultures. (Textio reports that job seekers are already using it to find female-friendly tech companies.) Fashion cycles will likely speed up, as positive terms spread more quickly, neutralizing their effects and rewarding fresher expressions. And, if we’re lucky, we may also see a mathematics-driven appreciation for writing to be read -- or at least an end to proactive interfacing with stakeholders.