Skip to content
Menu
Home / Article

Hiring signals from job posts: what broke and what was free

Job posts as hiring signals: a swallowed 402 made all but two of our seeds look empty, a cron never ran, and Workday's free JSON covered the employers we wanted.

Our testJul to Nov 2026 · Workday, Greenhouse, Lever and Ashby boards · PredictLeads, TheirStack and SerperAll our tests ↗
The short answer

Job posts work as hiring signals once you can trust the pipeline: most of our problems from July to November 2026 were silent failures, such as a swallowed HTTP 402 that made all but two of about 100 seeds look empty and a weekly cron that never ran. The employers we cared about post on Workday, and its public job JSON is free, so we kept PredictLeads' free tier for tracking accounts we already knew and paid for no job feed.

The question was whether public job posts could tell us which companies were staffing up for the work our product does, soon enough to act on. We spent July to November 2026 on it. The sources mattered less than we expected. What cost us most was data that looked fine and wasn't.

A 402 that looked like "no results"

In September we ran PredictLeads' similar_companies endpoint over about 100 customer seeds. The first results were promising. Then the account ran out of monthly requests and the API started answering HTTP 402, payment required.

Our code caught HTTPError and returned an empty list. So all but two of the seeds printed "returned=0", and for a while we read that as PredictLeads having nothing similar to offer. It was our error handling. A news poll in August had shown a version of the same thing: all 176 calls came back blocked, most likely because the quota was gone, and a blocked call says nothing about whether the company had news.

Keep "no results", "quota hit" and "request failed" as three separate states, and stop a run on the first 402 instead of carrying on.

Closed jobs and frozen archives

PredictLeads returns closed postings by default. Of about 2,100 job rows we pulled in August, 42% carried status="closed". Its own status field didn't catch everything either: one employer's domain returned 100 rows marked open whose median last_seen_at was more than five years old. The company had moved to a new domain, and the old one was serving a frozen archive. The response names the live domain in meta.redirects_to_domain, and re-querying that domain brought the coverage back.

The filter we settled on: drop status = "closed", keep rows seen in the last 21 days, and dedupe on title plus location, because the same posting often appears once from the employer's ATS and once mirrored from LinkedIn.

The cron that never ran

The free Workday poller was the best source we had, and it stopped producing anything without telling us. It ran from cron on a laptop, macOS blocked cron from reading the project folder, and every weekly run in the log, into November, recorded "Operation not permitted". The job was installed, scheduled and doing nothing.

An earlier version had a quieter bug. It mapped Workday tenants to company domains from a hard-coded list of 9 entries, and silently dropped 56 of 65 tenants, large employers among them, until we made the lookup programmatic.

Check that a scheduled job wrote output this week, because a crontab entry only shows that someone scheduled it.

Most keyword matches were noise

On Greenhouse, Lever and Ashby boards in late July, 190 raw matches looked like a healthy weekly flow. Reading them showed otherwise. Generic job-function terms produced 79 of the 190, at close to zero usefulness. One employer's "about us" boilerplate, repeated in every posting, produced another 74. After that, about one useful signal a week remained.

Short product names caused the rest. One that is also an ordinary English word matched 73 of 269 postings, and another matched tens of thousands of Hacker News comments until we added word boundaries to the regex. Collapse matches per company before you count anything.

The employers we care about post on Workday

Our buyers skew toward large employers, and large employers tend to run careers sites on Workday. Each tenant answers a public JSON endpoint with no API key:

POST https://{tenant}.wd{N}.myworkdayjobs.com/wday/cxs/{tenant}/{site}/jobs

In August a handful of Serper queries found 90 tenants, all of which answered. Scanning 17,500 postings turned up 521 matching roles at 65 companies, 135 of them posted in the previous week. Workday also gives real posting dates. Greenhouse's updated_at changes when a company bulk-refreshes its board, so use first_published there. At 12 requests at once, 19 of the 90 tenants timed out, so throttle the poller.

The Greenhouse and Lever boards we scanned matched far less often, but that tells you who posts there more than it tells you about the ATS. We found the Workday tenants with our own target keywords, and the Greenhouse and Lever boards we had skew toward startups that rarely hire for these roles.

What PredictLeads covered

PredictLeads has no discovery endpoint: the global job search path returned 404, so you need the company domains before you start. On a set of large Workday employers, it returned roughly a third of the matching roles the poller saw. Most of that gap came from the page size, since one call returns 100 rows in no useful order, and a large employer needs five or six calls to reach parity. Our first pass also used limit=50, which made it look thinner still until we fixed it.

On 12 employers we picked because they don't use Workday, it found 14 matching roles in total, and four of the 12 had any. Part of that is the population: several of those companies mostly hire drivers and salespeople.

Two things were worth having on the free tier. Every response names the company's parent and subsidiaries, which gave us 71 subsidiary domains from 14 parents at one call each. And tracking worked: a re-test in mid-August caught a funding round, a product launch and a promotion on the day they happened, along with one invented CTO hire attached to the wrong company.

Prices we saw in July 2026: a free tier of 100 calls a month, then a $40 minimum plus $0.04 per credit. We use the free tier to track known accounts and didn't pay for discovery.

Search engines index titles

Searching job titles through Serper hit 1 time in 33, because Google indexes a posting's title and very little of its body. Full descriptions from an Apify Indeed scrape fixed the data and not the outcome: the companies that used our keywords most in job descriptions were the ones building similar products.

TheirStack: checked, not bought

We tested TheirStack with curl in July 2026. It returns the company domain in the same response as the job, job_description_pattern_and accepts regex with (?i), and blur_company_data: true returns a preview that costs zero credits, so you can tune a query before paying. We set a rule not to buy below 20% precision and never ran the paid test, so we have no precision figure for it.

Job changes are a different signal

A champion leaving a customer is a signal too, and it needs a different source. Ocean.io's monitor caught moves among our past champions (Ocean.io review), and re-verifying old addresses gives a rough decay rate (ZeroBounce review). People Data Labs would have cost about $0.28 a record for this in July 2026, and Crustdata quoted about $200 a month; we didn't trial either.

Who it's for

Approach Use it if Skip it if
Free Workday poller Your buyers are large employers Your market is startups
PredictLeads free tier You already have the account list You need it to find new companies
TheirStack You need many boards and will measure precision You can't write a regex

Test it yourself in 30 days

  1. Find Workday tenants for your target accounts with a few Serper queries.
  2. Poll the JSON endpoint weekly, throttled, and store postedOn dates.
  3. Match on a word-boundary regex and read the first 50 matches by hand.
  4. Drop generic terms and collapse each company's boilerplate before counting.
  5. Make every API wrapper report quota errors separately from empty results.
  6. Check each week that the scheduled job wrote output.

The TAM list we built with Serper is where the account list came from.

Questions buyers ask

What are hiring signals?

Job posts that show a company is staffing up for the work your product does. We used them to decide which accounts to contact first, after filtering out closed roles, boilerplate and generic keywords.

Is there a free source of job posting data?

Workday's public job JSON needs no key, and all 90 tenants we tried answered. Greenhouse, Lever and Ashby boards are also free, though for our market they held few of the employers we wanted.

Is PredictLeads worth it for job data?

For tracking known accounts its free tier was useful, and for discovery it wasn't, because it has no discovery endpoint. Filter out closed postings, which were 42% of the rows we pulled.

Is TheirStack worth it?

We curl-tested it in July 2026 and didn't buy. It returns the company domain with each job and has a zero-credit preview for tuning queries, so measure precision on the preview before paying.

Why does my job signal pipeline return no results?

Check for swallowed errors first. Our wrapper turned an HTTP 402 quota error into empty lists, so all but two of our seeds looked like they had no results.

Sources

Sources and evidence behind this page