# The Career-Site Source Sweep, Four ATS Hosts to a Deduplicated Daily Intake

*Build saved queries per applicant-system host, run a 30-minute sweep that yields a deduplicated, first-seen-dated intake list, and keep or drop hosts on two-week yield.*

- Canonical URL: https://www.refolk.ai/candidates/guides/career-site-source-sweep
- Pillar: Applying at volume
- Format: Playbook
- Published: 2026-11-08
- Last reviewed: 2026-11-08
- Reading time: 19 min
- Keywords: how to search company career pages directly for jobs, site: search for jobs on applicant tracking systems, find jobs before they hit LinkedIn, search Greenhouse Lever Ashby Workday job boards, how to find newly posted jobs at the source

## Key takeaways

- In Refolk's index, US recruiter profiles listing a Workday skill outnumber those listing Greenhouse 523 to 118, about 4.4x, so a US sweep that skips myworkdayjobs.com skips the host where enterprise hiring clusters.
- On Discord's Greenhouse board, all 44 jobs had an updated_at in the past week but only 1 had a first_published date that recent, so sort Greenhouse rows on first_published and never on updated_at.
- Secondary reports of NBER Working Paper 32320 put 45% of Dice.com applications within 48 hours of posting and 60% within 96 hours, so a sweep cadence longer than two days lands behind most applicants.
- Workday's 'Posted 30+ Days Ago' label is a floor, not a date, so a row carrying it should go to a true-age check rather than be treated as 30 days old.
- Dedupe career-site hits and aggregator alerts on the employer's requisition ID, not title plus location, because Greenhouse merges a second application to the same active job and the resend adds nothing.
- Google documents its after: operator as filtering on last-updated dates, so it reproduces the updated_at trap and cannot stand in for a publish-date filter.

This playbook is for job seekers applying at volume who want to catch new openings for one target role on employers' own career pages before aggregators copy them. It turns the familiar site: search trick into a routine you can run every morning: a saved query set per hosting domain, a named employer watchlist, a timed sweep that ends in a deduplicated intake list with a first-seen date on every row, and a two-week yield number per host that tells you which hosts are worth your minutes.

## Why sweep employer career pages at the source?

Because applications pile into new postings within the first two days, and the employer's own applicant tracking system is where a posting exists first. An applicant tracking system (ATS) is the software an employer uses to publish its job board and collect applications. Greenhouse, Lever, Ashby, and Workday are the four this guide covers.

The best evidence on timing measures where applications go, not who gets hired. Secondary reports of NBER Working Paper 32320, by Steven Davis and Brenda Samaniego de la Parra, describe 125 million linked applications on Dice.com. They found that 45% of applications went to postings made in the previous 48 hours, 60% went to postings made in the previous 96 hours, and the median posting stayed live 7 days. I have not read the primary paper. These figures come from secondary reporting.

**45% - Share of applications sent to postings made in the previous 48 hours**

From secondary reports of NBER Working Paper 32320, covering 125 million Dice.com applications; it measures application flow, not callbacks.

The practical consequence: a sweep cadence longer than two days puts you behind most applicants. That conclusion comes from flow data, so treat it as a floor on urgency, not a promise of better callback odds.

| Study | Sample | Outcome measured | Traceability |
|---|---|---|---|
| NBER WP 32320 | 125M applications | Application flow | High (working paper) |
| TalentWorks 2017 | about 1,600 (one page says 160,000) | Interview rate | Medium; sample conflicts |
| jobloo | 10,000 claimed | Callback | No methodology published |

> **Note:** The callback claims are weaker than they look
>
> TalentWorks reported interview odds up to 8x higher in the first 96 hours, but published accounts disagree on its sample size by a factor of 100. Build your cadence on the flow data, and treat any callback multiplier as unproven.

## Which applicant-system hosts to sweep

Sweep all four hosts, but weight them by who you are targeting: in the US, Workday is where enterprise hiring clusters, and Greenhouse is mostly a coastal-tech sweep. That split comes from recruiter skill listings in Refolk's index.

In Refolk's index of professional profiles, I counted US recruiter and talent-acquisition profiles that list each ATS as a skill. A self-listed skill is not proof of adoption, because people list tools they used at past jobs and skip tools they use every day. It is still the best public proxy I have for where hiring work happens.

| Market and skill | Profiles | Example employers in sample |
|---|---|---|
| US, Greenhouse | 118 | Meta, Palo Alto Networks, Paraform, Lumos, Zearn |
| US, Workday | 523 | Wells Fargo (2), Boeing, Lifepoint Health, C&S Wholesale Grocers, Cohere |
| UK, Greenhouse | 2 | Beauhurst, Coefficient Giving |

**4.4x - US recruiter profiles listing Workday versus Greenhouse in Refolk's index**

523 against 118, with banks, aerospace, and health systems among the sampled Workday employers.

What this tells you about each host:

- **Workday** is the enterprise host. Large employers standardise on it, which the sampled names reflect. The catch is that Workday shows the least precise date of the four, so the host with the most US recruiter adoption gives you the weakest freshness signal.
- **Greenhouse** skews coastal and tech. In a 25-profile sample of the Greenhouse-skill group, 8 sat in New York, Brooklyn, or the SF Bay Area. In the Workday-skill sample, no region held more than 1 profile. If your target role is a tech role in those metros, Greenhouse earns its place. If not, it may be a low-yield host for you, and the two-week score will show that.
- **Lever and Ashby** are not in my index counts, so I make no adoption claim about them. Keep them on the watchlist if employers you want use them, and let the yield score decide.
- **UK Greenhouse** appears in only 2 profiles. That is too few to conclude anything. If you are searching in the UK, build the host mix from your own watchlist.

## How each host dates a posting

Each host stores dates differently, and the most visible "fresh" date is often wrong. Greenhouse's first_published is the only documented publish timestamp among the four. Workday gives day-level labels with an open-ended 30+ bucket. Lever and Ashby give machine timestamps whose exact field names I could not confirm from primary docs.

| Host | Public date | Precision | Known trap |
|---|---|---|---|
| Greenhouse | first_published, updated_at | Timestamp | updated_at looks fresh |
| Lever | Epoch ms | Millisecond | EU host returns empty on US API |
| Ashby | ISO with offset | Timestamp | No filtering; isListed |
| Workday | Relative label | Day; 30+ bucket | 2,000 cap; floor, not a date |

Reading guide, host by host:

- **Greenhouse.** Use first_published. The updated_at field changes on any edit, so it measures maintenance rather than newness. On Discord's Greenhouse board, all 44 jobs had an updated_at in the past week, but only 1 had a first_published date in that week.
- **Workday.** List views show "Posted Today", "Posted Yesterday", "Posted N Days Ago", and "Posted 30+ Days Ago". The first three are useful at day precision. "30+" is a floor: the posting could be 31 days old or several months old.
- **Lever.** Dates arrive as epoch milliseconds, meaning a count of milliseconds since 1 January 1970, which a spreadsheet must convert before you can read it. Find the field in the response and check one known posting against the date shown on its page before you trust it.
- **Ashby.** Dates arrive as ISO timestamps with a time-zone offset. Postings also carry an isListed true/false value, and unlisted roles should not appear on the public board, so a role a recruiter sent you directly may never show up in your sweep.

> **Watch out:** Do not sort by updated_at
>
> Any edit to a Greenhouse posting bumps updated_at, so a months-old role looks brand new. Sort and filter on first_published, and route anything that looks reposted to the true-age run.

> A fresh updated_at usually means someone edited the posting, not that the job is new.

## Building the saved query set and the employer watchlist

The sweep runs on two artifacts: one saved site: query per host domain, which finds employers you do not know yet, and a watchlist of named employer boards, which you check directly every day. The queries feed the watchlist, and the watchlist is the part that runs daily.

### The saved query set

Google's site: operator restricts results to one domain. Google's own help page notes that the operator only works with no space after the colon: [site:nytimes.com] works and [site: nytimes.com] does not. Save one query per domain in this set:

**Saved site: query set, one line per host domain**

```
site:boards.greenhouse.io "product manager"
site:job-boards.greenhouse.io "product manager"
site:jobs.lever.co "product manager"
site:jobs.eu.lever.co "product manager"
site:jobs.ashbyhq.com "product manager"
site:myworkdayjobs.com "product manager"
site:myworkdaysite.com "product manager"
```

*Replace "product manager" with your target title, and add a second line per domain for its most common variant title.*

Run this set weekly, not daily. Its job is discovery: each new employer it surfaces with a relevant role becomes a watchlist row. Do not add after: to these queries to get "new" results. Google documents after: as finding documents last updated after a date, which is the updated_at trap again, one layer up.

### The employer watchlist

Each watchlist row needs a host, a board token, and for Lever a region. The board token is the employer's identifier on the host, and it sits in the URL of any job on their careers page. My working estimate is about 2 minutes per employer, done once. No source documents that figure, so time your first ten.

Where the token lives, by host:

- **Greenhouse:** boards.greenhouse.io/{token} or job-boards.greenhouse.io/{token}. Read-only data comes from boards-api.greenhouse.io/v1/boards/{token}/jobs. Greenhouse states no authentication is required for any GET request. Only submitting an application through the API needs credentials.
- **Lever:** jobs.lever.co/{site} for US boards, jobs.eu.lever.co/{site} and api.eu.lever.co/v0/postings/{site} for EU-hosted employers. The postings API pages at 100 jobs per page, up to 100 pages, and asks for a 1-second crawl delay. Lever's postings documentation publishes no rate limit for reading.
- **Ashby:** jobs.ashbyhq.com/{name} for the hosted board, and api.ashbyhq.com/posting-api/job-board/{JOB_BOARD_NAME}?includeCompensation=true for the data. The endpoint has no search or filtering, so you pull the whole board and filter it yourself.
- **Workday:** {tenant}.{pod}.myworkdayjobs.com/{site}, where the pod is a server label such as wd1, wd5, or wd12. Workday's own careers site, workday.wd5.myworkdayjobs.com/en-US/Workday, shows the pattern. Workday has no documented public API, so sweep it through the career site's search box.

If a board comes back empty, the documented causes are a wrong slug, an employer that moved to another ATS, or a board on the EU host. Lever adds one more: many customers disable the public postings endpoint or publish on a vanity URL, meaning the employer's own domain.

## The daily sweep, step by step

The daily sweep is eight steps. Steps 1 and 2 are setup, steps 3 through 7 run every morning, and step 8 runs at the two-week mark. I budget about 30 minutes for the daily part. That is my target, not a measured figure, so log your actual minutes per host from the first sweep, because step 8 depends on them.

#### The career-site source sweep

1. **Identify each employer's host and board token** - Open one job from each employer's careers page and read the URL for host, token, and Lever region. Done means every watchlist row has all three.
2. **Build one saved site: query per host domain** - Save a query for each host domain with no space after the colon. Done means one working query per domain.
3. **Snapshot each watched board** - Capture today's full list of posting IDs from every watched board, because none of these hosts return closed postings. Done means today's ID list sits next to yesterday's.
4. **Diff today's IDs against yesterday's** - Compare the two lists to find rows that are new since the last sweep. Done means a short list of new rows and nothing else.
5. **Read the date field the host's way** - Use first_published on Greenhouse, read Workday labels as day-level with 30+ as a floor, and confirm the publish field on Lever and Ashby. Send anything reposted or stale-looking to the true-age run.
6. **Merge with aggregator alerts and log first-seen** - Match new rows and aggregator alerts on the dedupe key and keep one row per requisition. Stamp the first-seen date and source on every row.
7. **Apply the same day** - Work the intake list top down, checking each requisition ID against your log before submitting. Every qualified row ends the day applied to or skipped with a reason.
8. **Score each host's yield at two weeks** - Count qualified new rows and minutes spent per host across the window. Write keep, trim, slow, or drop next to each host.

#### One morning's sweep

1. **Snapshot** - Pull today's posting IDs from every watched board
2. **Diff** - Keep only IDs absent from yesterday's snapshot
3. **Date check** - Read the host's publish field; route stale-looking rows out
4. **Merge** - Fold in aggregator alerts on the requisition ID
5. **Log** - Stamp first-seen date and source on each row
6. **Apply** - Work the list the same day

*The sweep finds new postings by diffing IDs between days, then checks dates, so an edited old posting never enters the intake.*

### Why diff first, then date

Sources disagree on the order. API-oriented guides diff IDs first and filter by date second. Search-oriented pages filter by date first using after:, which reflects update time. I diff first. A new ID is a hard fact: it was not on the board yesterday and it is today. A date is a field the employer controls and edits. The documented method for history is the same reason: none of these APIs return closed postings, so if you want history, you snapshot daily and diff.

For a non-technical sweep, a snapshot can be as simple as pasting the list of job URLs from each board into a sheet tab named with the date. The ID is the number or requisition code at the end of each URL.

### Workday boards over 2,000 postings

Workday search results stop at 2,000 postings per board. On a large employer's board, a full pull can miss new rows. Type your target title into the board's search box and snapshot the filtered results instead of the whole board.

### Applying from the intake list

The intake list is only worth the effort if you apply the same morning, and tailoring each application is where the time goes. [Refolk](/candidates) tailors your resume to each posting on the list and drafts the cover letter from your own history, so the sweep's head start is not spent rewriting bullets.

## The dedupe key: merging career-site hits with aggregator alerts

Key every row on the employer's requisition ID where you can see it, fall back to the host's own posting ID, and never key on title plus location. A requisition ID is the employer's internal code for one approved opening, and it stays the same across boards and across runs.

Why title plus location fails: employers run several Workday sites, and the company name on an aggregator does not map cleanly to a tenant. "Bank of America" is not the same string as bankofamerica.wd1. Two rows with the same title and city can be two openings, and one opening can appear under two titles.

What each host gives you:

- **Workday** URLs embed the requisition, such as JR2028407, and it is stable across runs.
- **Greenhouse** returns a stable numeric id and a requisition_id. Diffing the numeric id between runs detects new and closed roles.
- **Lever and Ashby** give a posting ID in the URL. Use it as the key within that host.

A key that works across all systems is not established publicly. When the same opening shows up on two hosts or on an aggregator without a requisition code, match by hand and note the match in the log.

#### Dedupe key, strongest first

1. **Employer requisition ID** - JR codes in Workday URLs, requisition_id on Greenhouse; stable across boards and runs
2. **Host posting ID** - Greenhouse numeric id, Lever or Ashby posting ID in the URL; stable within one host
3. **Canonical posting URL** - Works only if the employer never reposts under a new URL
4. **Title plus location** - Not a key; use only to flag possible matches for a manual check

*Use the highest layer you can see for each row, and only drop to a weaker key when the stronger one is missing.*

> **Rule:** Log the requisition before you apply
>
> Write the requisition ID into the intake log before you submit, and check it against the log first. Greenhouse merges a second application to the same job while the first is active, so a resend adds nothing.

Greenhouse's documented duplicate behaviour is worth knowing exactly. If you are active on Job A and apply online for Job A again, the two applications are merged. If you were already rejected or hired, every application is preserved. Greenhouse also flags duplicate candidates on matching email, phone, LinkedIn URL, or name, so a new email address does not make you a new applicant.

The intake log below carries the key, the first-seen date, and the field you need to measure aggregator lag yourself.

**Intake log, one row per unique opening**

```
dedupe_key | host | employer | board_token | title | location | date_field_used | date_value | first_seen | first_seen_source | aggregator_seen | status
```

*Paste as a spreadsheet header; fill aggregator_seen only when an alert later shows the same requisition.*

The aggregator_seen column matters because the lag between a source-board publish and the same posting reaching an aggregator is not established publicly. After two weeks, the gap between first_seen and aggregator_seen on matched requisitions is your own measured head start, per host. If that gap is zero for a host, the sweep is not buying you time there.

## How the sweep goes wrong

Most sweep failures come from trusting a date or an empty result that means something other than it appears to. Each failure below produces either a false positive, where an old role looks new, or a false negative, where a live employer drops off your list. Each comes with the check that catches it.

1. **Sorting by updated_at.** False positive: an old role edited yesterday looks brand new. On Discord's board that was 44 fresh-looking jobs against 1 genuinely new one. Check: sort on first_published.
2. **Reading Workday "30+" as 30 days.** False positive: a role that is months old looks borderline recent. Check: open the job detail, or send the row to the true-age run. Do not log a date you do not have.
3. **Concluding a Lever board is empty when it is EU-hosted.** False negative: the employer gets dropped from the watchlist. Tools that read only the US API report EU-hosted employers as having no jobs. Check: retry api.eu.lever.co before dropping.
4. **Trusting Google after: as a publish filter.** False positive: Google documents after: as filtering on last-updated dates, so edited postings pass. Check: confirm the date on the host itself.
5. **Pulling a Workday board above 2,000 postings in full.** False negative: new rows past the cap never reach your snapshot, so the diff shows nothing. Check: filter by your target title in the board's search box, then snapshot.
6. **Deduping on title plus location.** Both errors at once: two openings collapse into one, or one opening under two titles becomes two rows. Check: key on the requisition ID.
7. **Re-applying to the same requisition.** Wasted application: Greenhouse merges same-job applications while the first is active. Check: look up the requisition ID in the log before every submission.
8. **Citing the TalentWorks figures as settled.** Bad planning input: reported sample sizes conflict, 1,600 against 160,000. Check: plan cadence on the NBER flow data and treat callback multipliers as unproven.

Two quieter failures deserve a watch. An Ashby role marked unlisted will never appear on the public board, so its absence from your sweep says nothing about whether it exists. And an employer that moves ATS makes its old board go quiet. A board that drops to zero new rows for a week should get a fresh look at its careers page URL before you read the silence as a hiring freeze.

## Scoring each host's two-week yield

Score each host by qualified new rows per sweep minute across two weeks, then keep, trim, slow, or drop it. No public source defines this measure or a minimum sample size, so what follows is my working rule, not an established standard.

Definitions for the score:

- **Qualified new row:** a row that survived the diff and the date check, matches your target role, and passes your own fit bar.
- **Sweep minutes:** the minutes you logged on that host across the window, including snapshot, diff, and date reading.
- **Yield:** qualified new rows divided by sweep minutes, per host.

The fit bar is where people get inconsistent, and an inconsistent bar makes the score meaningless. Refolk scores how well you actually fit each posting, which gives you one consistent bar to count qualified rows against across every host.

#### Keep-or-drop call per host

Horizontal axis runs from Few minutes per sweep to Many minutes per sweep. Vertical axis runs from Few qualified new rows to Many qualified new rows.

| Quadrant | What it means |
| --- | --- |
| Keep | Cheap and productive; sweep daily as is |
| Trim | Productive but slow; cut low-yield employers from this host's watchlist |
| Slow down | Cheap but quiet; move to a weekly check and keep discovery queries running |
| Drop | Costly and quiet; remove from the daily sweep and revisit next quarter |

*A host earns daily sweeps by producing qualified new rows, and its minute cost decides whether you keep it whole or trim it.*

Read the result with its limits in view:

- **Two weeks is a small sample.** A host with one or two qualified rows is not proven low-yield. Before dropping a host whose employers you care about, extend the window.
- **Yield is relative to your role.** The Greenhouse sweep skews toward coastal tech in Refolk's index, so a Greenhouse score that is low for a hospital administrator says nothing about its value for a backend engineer in New York.
- **Pair yield with head start.** A high-yield host whose rows reach aggregators the same day earns less than its count suggests. Use the aggregator_seen column to check.

## Keeping the sweep current

The sweep decays in three predictable ways: employers move ATS, boards move region, and your target role drifts. Re-check each on a schedule rather than waiting for a quiet week to tell you something broke.

- **Weekly:** run the saved query set to find new employers, and add any with a relevant role to the watchlist.
- **Every two weeks:** score yield per host and act on the matrix.
- **When a board goes empty:** check the careers page URL for a new host or a vanity URL, and retry the Lever EU host.
- **When a host's date field looks different:** check one known posting's page against the field before trusting the field again. Field names and formats are set by the host and can change.

#### Before you call a sweep done

- [ ] Every watchlist row has a host, a board token, and for Lever a region.
- [ ] Today's ID snapshot is saved beside yesterday's for every watched board.
- [ ] No row in the intake list was dated from updated_at or from Google after:.
- [ ] Every Workday row labelled 30+ Days Ago went to the true-age run, not the intake.
- [ ] Every intake row has a dedupe key, a first-seen date, and a first-seen source.
- [ ] No requisition ID appears twice in the intake log.
- [ ] Every qualified row is marked applied or skipped with a reason.
- [ ] Minutes spent per host are logged for the two-week yield score.

When a hit looks reposted or stale, do not re-derive its age inside this routine. Hand it to the library's true-age run, keep the sweep moving, and let the intake list stay short enough to apply to the same day.

## Frequently asked questions

### Can I search Greenhouse, Lever, Ashby, and Workday job boards without an account?

Yes for reading postings on three of them, with caveats on the fourth. Greenhouse states its job board data is public and needs no authentication for GET requests. Lever has a keyless postings API, though many customers disable it or use a vanity URL. Ashby documents a public posting API with no search or filtering. Workday has no documented public API, so sweep it through the hosted career site's own search box in a browser.

### Does Google's after: operator find newly posted jobs?

No, not reliably. Google's help page describes after: as finding documents last updated after a given date, which means it filters on update time, not publish time. An old posting that an employer edited yesterday passes the filter. Use site: queries to find boards, then confirm freshness on the host itself, using first_published on Greenhouse or the day label on Workday.

### How often should I sweep company career sites for new jobs?

Daily, and never less often than every two days. Secondary reports of NBER Working Paper 32320 found 45% of applications on Dice.com went to postings made in the previous 48 hours, and the median posting stayed live 7 days. That measures where applications go rather than who gets called back, but a slower cadence means arriving after most of the pile.

### What happens if I apply twice to the same job on Greenhouse?

If your first application to that job is still active, Greenhouse merges the second one into it, so the resend adds nothing. If you were already rejected or hired, every application is preserved on your record. Greenhouse also flags duplicate candidates on matching email, phone, LinkedIn URL, or name, so a different email address does not make you a new applicant.

### How many days after a career-site post does a job show up on LinkedIn or other aggregators?

That lag is not established publicly, and it likely varies by employer and aggregator. Measure it yourself: log the first-seen date and source for every requisition, then record the date the same requisition arrives in an aggregator alert. After two weeks you have a local lag figure per host that tells you how much head start the sweep actually buys.

### Why does a Lever company's board come back empty?

Usually one of four things: the board slug is wrong, the company moved to a different applicant system, the board sits on Lever's EU host, or the employer disabled the public postings endpoint or publishes on a vanity URL. Retry the EU host at api.eu.lever.co and check the careers page URL before dropping the employer from your watchlist.

---

*From the Refolk guide library. I revise these guides rather than replacing them, so the current version is always at https://www.refolk.ai/candidates/guides/career-site-source-sweep*
