I run a visa agency in Bali and build scrapers on Apify on the side. One of them reads Workday career sites — the company.wd5.myworkdayjobs.com pages that NVIDIA, Salesforce, Adobe and a long list of banks and retailers use to publish openings. My own reason for caring is simple: when a company starts hiring a lot in Indonesia, somebody there will soon need a work permit.

Workday is a pleasant target compared to most job boards. There is no login, no browser needed, and every tenant runs the same JSON API. But on 21 November 2026 I noticed that the number the API reports as total is not the number of jobs on the site. This post is about that, plus the other things I had to learn before the rows came out clean.

The endpoint every career site uses

Open any Workday career page with the network tab open and you will see the page call itself:

POST https://nvidia.wd5.myworkdayjobs.com/wday/cxs/nvidia/NVIDIAExternalCareerSite/jobs Content-Type: application/json {"appliedFacets": {}, "limit": 20, "offset": 0, "searchText": "engineer"} 

The path is /wday/cxs/<tenant>/<site>/jobs. The tenant is the first label of the host (nvidia), the site is the first path segment after an optional locale like en-US. Each job in the response has an externalPath, and a GET on /wday/cxs/<tenant>/<site><externalPath> returns the full posting: description HTML, time type, country, additional locations, requisition id, the date it went live.

Two things I learned by getting them wrong:

  • limit is 20, full stop. I tried "limit": 50 on NVIDIA and got HTTP 400. You page in steps of 20.
  • total only comes on the first page. On page two the same field is 0. If your loop re-reads total every page, it will think the search is empty and stop after 20 rows.

The 2,000 that is not 2,000

Here is what NVIDIA's endpoint answered on 21 November:

searchText "" total 2000 searchText "engineer" total 2000 

Same number for "everything" and for one keyword. That alone is suspicious. The first page of the response also carries facets — the filters you see on the left of the career page — with a count for every value. Summing the jobFamilyGroup facet:

jobFamilyGroup 15 values, sum 2645 Engineering 1641 Sales 317 Univ Employment 233 Operations 121 ... workerSubType 6 values, sum 2645 timeType 2 values, sum 2645 (Full time 2644, Part time 1) 

Three independent facets agree on 2,645. For "engineer" the job-family facet summed to 2,516, against a total of 2,000.

So I paged past the ceiling to see what happens:

offset 0 first ids: JR2029456-1, JR2030026, JR2029344 offset 2000 first ids: JR2029456-1, JR2030026, JR2029344 offset 2500 first ids: JR2029456-1, JR2030026, JR2029344 offset 4000 first ids: JR2029456-1, JR2030026, JR2029344 

No error, no empty page — past 2,000 the endpoint hands you page one again. A scraper that trusts total stops at 2,000 and misses 645 jobs. A scraper that ignores total and keeps going until the page is empty never stops, and fills your dataset with duplicates.

The fix is the facets themselves. Filtering by one job family brings the count under the ceiling, and then total is honest:

{"appliedFacets": {"jobFamilyGroup": ["0c40f6bd1d8f10ae43ffaefd46dc7e78"]}, "limit": 20, "offset": 0, "searchText": ""} 

That returned total: 1641 — exactly the Engineering facet count. Split by job family, page each slice, de-duplicate by externalPath, and you get the whole site.

One caution: facet sums are only a good estimate when a job can sit in one value. On Salesforce the first facet is the posting country; it summed to 1,643 while total said 1,614, because a job open in two countries counts twice. Adobe's external site was small enough that both said 253. The ceiling only bites on big tenants.

Honest status of my own tool: today the Workday Jobs Scraper follows total and de-duplicates, so it stops cleanly at 2,000 per search — no duplicates, but also no job 2,001. Splitting by facet when total hits exactly 2,000 is the next change. If you need more than 2,000 rows from one employer before then, run several keywords: each search has its own ceiling.

What one real run looks like

The last run before I wrote this was the default example: NVIDIA, keyword "engineer", 40 jobs, full postings on.

{ "careerSiteUrls": ["https://nvidia.wd5.myworkdayjobs.com/NVIDIAExternalCareerSite"], "keywords": ["engineer"], "maxJobsPerSite": 40, "includeDescription": true } 

It finished in about 25 seconds, made 43 HTTP requests (two search pages, 40 postings, one for the company name) and charged $0.054. One row, trimmed:

{ "type": "job", "title": "Senior System Engineer, Solution Engineering", "company": "NVIDIA", "location": "3 Locations", "locationFull": "Poland, Remote", "additionalLocations": ["Germany, Remote", "Netherlands, Remote"], "country": "Poland", "postedOn": "Posted Yesterday", "postedAt": "2026-11-19", "requisitionId": "JR2026094", "timeType": "Full time", "hiringOrganization": "PL01 Nvidia Poland sp. z o.o.", "applyUrl": "https://nvidia.wd5.myworkdayjobs.com/NVIDIAExternalCareerSite/job/Poland-Remote/Senior-System-Engineer--Solution-Engineering_JR2026094" } 

The median description was about 3,600 characters of plain text, with the HTML kept next to it.

Three fields that look simple and are not

postedOn is a label, not a date. The search returns "Posted Yesterday", "Posted 4 Days Ago", "Posted 30+ Days Ago". In those 40 rows, 17 said "30+ Days Ago". The posting endpoint has the real startDate, and for those 17 it ranged from 20 October back to 2 March — anywhere from one month to almost nine months. If you sort by freshness using the label, a March posting and an October one are tied. Take the date from the posting call, which means paying for the second request per job.

There is a sharp edge here in my own Actor that I found while writing this: the "posted within N days" filter reads the label, before the posting is fetched, and treats "30+" as 30. With a window of 30 days or less that is fine — "30+" posts drop out. With a 60-day window, a job from March gets in. Until that is fixed, use 30 days or less, or filter on postedAt yourself.

hiringOrganization is the legal entity, not the brand. 40 NVIDIA rows named twelve different employers: "2100 NVIDIA USA" (18), "IL00 Mellanox Technologies, Ltd." (9), "IN01 NVIDIA Graphics Bengaluru" (3), and nine more, each with an internal code in front. Nine of the 40 engineering jobs are formally Mellanox, the Israeli company NVIDIA bought years ago. For a recruiter that is noise. For someone working out which local entity will sponsor a visa, it is the most useful field in the row. The Actor keeps it as is and puts the brand from the career page's og:title in company.

remoteType was empty on all 40. Twelve of those rows had "Remote" in the location itself ("US, CA, Remote", "Poland, Remote"). NVIDIA simply does not fill the remote field; it encodes remote work as a location. If you filter remote jobs by remoteType, you get zero on this tenant. Check the location strings too.

Blocking

Workday sits behind Cloudflare, and its 429 is often not a rate limit but a bot challenge. On 16 November that challenge was triggered by the TLS handshake of a newer Node.js image on every route, residential proxies included; going back to Node 20 passed four requests out of four. If you write your own client and get 429 from the first request, try a different runtime before buying proxies. Workday also puts tenants into scheduled maintenance, mostly at weekends; the API then redirects to a maintenance page instead of returning JSON, and the only fix is to wait.

Calling it

curl -X POST "https://api.apify.com/v2/acts/lergassy~workday-jobs-scraper/run-sync-get-dataset-items?token=YOUR_TOKEN" \ -H "Content-Type: application/json" \ -d '{"careerSiteUrls":["https://salesforce.wd12.myworkdayjobs.com/External_Career_Site"],"keywords":["account executive"],"maxJobsPerSite":100}' 

The dataset ends with a free type: "notice" row summarizing the run; filter on type == "job".

When not to use this

  • The employer is not on Workday. Greenhouse, Lever, iCIMS and SuccessFactors have their own APIs; a URL that is not *.myworkdayjobs.com or *.myworkdaysite.com is rejected.
  • You need more than 2,000 jobs from one search right now. See above — split by keyword until the facet split ships.
  • You want salary. NVIDIA puts pay ranges inside the description text, if at all; there is no structured salary field to rely on.
  • Exact posting age from the listing alone. The listing only knows "30+ days". The date costs one request per job.

The first thing I would check on any Workday tenant now is whether total is exactly 2,000. If it is, the real number is in the facets.