Source of Hire Records the Route, Not Where the Person Heard

Which question is the source field answering?

The source field records the route a submission travelled: which board, aggregator, syndication partner or autofill tool delivered the record into your applicant tracking system. That was a reasonable stand-in for discovery when a person found a job and applied for it in the same place. It is not one now, because those two events happen in different places and often days apart.

The volume underneath the field has also changed character. Application records include duplicates, spam and submissions made by tools on a candidate's behalf, and in one applicant tracking vendor's customer base the average recruiter now processes 291 applications per hire against roughly 100 in early 2021 ³. Route data collected across that population is accurate about routes and quiet about people.

Channel-level volume comparisons inherit a second problem from how the numbers are produced. Apply rate on a paid job ad is completed applications per click, so it measures the friction of the form and the targeting of the ad rather than interest in the role: Appcast put the median on long-apply ATS ads at 6.1% at the end of 2024, against a 19.37% average on one-click easy-apply ads ². Those two rates come from different boards, different ads and different candidate intent, so the gap is a ceiling on what a shorter form buys ². Even read as a ceiling, it is large enough that ranking channels by volume partly ranks their application forms.

So the field is answering a logistics question well and a marketing question badly. Keep it for the logistics question. What to do when a candidate has no memory of applying is the sharpest version of the gap between a route and a person, and it turns up in phone screens well before it turns up in a report.

Why does the standard repair make it worse?

Because it improves the accuracy of the wrong measurement, and accuracy is persuasive. Mandatory source at candidate creation, tags on every posting, capture before the record can be edited: each of those genuinely raises how reliably you know which system delivered a submission. None of them moves you closer to knowing where a person heard about the role, and a clean field reads as a true one.

Watch the failure mode play out. Syndication pushes a posting into feeds that were never a discovery surface, so those feeds show volume, carry perfect tags and convert into nothing. Meanwhile a candidate hears about the role from a former colleague, searches for it, and arrives through whichever board ranked first that morning. The tag on that record is correct and the inference drawn from it is wrong, and no amount of tightening the capture rule changes either fact.

Referrals are where the difference costs the most, because the label travels with the candidate into the review. In Ashby's dataset 52% of referred candidates pass initial screens against 35% overall ¹. Referral status is something the screener can see, so that is not a blind comparison, and part of the gap is the endorsement acting on the reviewer rather than on the candidate ¹. That is also the mechanism by which a referral channel pulls the composition of hires toward the composition of the people already employed. The dataset carries no demographic variable, so it establishes the direction and the reason to measure the effect, never its size ¹.

Conversion and volume are separate questions. Referred candidates convert well at nearly every stage in that data ¹, which leaves open how large a share of your hires that channel actually carries. Count the share in your own numbers before you treat referrals as the main artery. How much of a pipeline referrals can safely carry is the sizing question that follows.

Read source by hires, not by applications

Rank channels by hires made and by conversion into the first human conversation, and never by application volume. Volume is the quantity the mediated route inflates, so a ranking built on it will promote whichever channel is best connected to aggregators. Hires are slow and few, which is the usual objection to using them, and it is also why the ranking they produce holds still.

Three rules make the report readable:

• Collapse every aggregator and syndication partner into one bucket unless you pay them separately. Splitting them measures your plumbing, and the plumbing changes without anybody deciding it should.
• Carry two numbers per channel: hires, and the share of applicants from that channel who replied to a first scheduling message. The second is cheap, arrives fast, and is not inflated by automated submission.
• Never report a channel's cost per application as though it were cost per candidate. Those diverged, and the divergence is exactly what the source field cannot see.

The honest limit is what sits downstream of the hire. In LinkedIn's 2025 survey of recruiting professionals, only 25% reported feeling highly confident in their organization's ability to measure quality of hire, while 89% agreed it would matter more, with the common inputs being job performance ratings at 66%, new hire retention at 60% and hiring manager satisfaction at 44% ⁴. That is practitioners rating their own employers rather than an audit of what anyone measures, and the respondents are a convenience sample of talent professionals across many countries ⁴. Do not wait for a settled quality measure before acting on channel data. Use hires and reply rate now, and add retention when you have enough of it. What tripled the application count on a single opening explains why the volume column stopped deserving the top of the report.

Ask the hire, in the first week

Add one sentence to the offer paperwork or the first-week checklist: how did you first hear about this role? Ask it once, in the person's own words, with no menu attached. A dozen answers a year from people who actually joined tells you more about reach than a mandatory field applied to twelve hundred applications ever did.

Free text matters more than it sounds. A dropdown invites the person to select the last touch they remember, and the last touch is precisely what the tag already recorded, so a menu produces expensive agreement with a number you already had. An open sentence produces named surfaces instead: a former manager, a conference talk, a newsletter, a friend who works two teams over.

The answer usually disagrees with the tag on the same record, and the disagreement is the finding rather than an error to reconcile. The tag is right about the route. The sentence is right about the awareness. When one channel keeps appearing in the sentences and never in the tags, that is a place worth spending on, and no attribution system was going to surface it.

Keep the field for contracts, because at renewal the question genuinely is which contract to renew and the tag can answer it. Spend against reviewed applications and hires rather than raw volume, collapse the aggregator bucket before you compare anything, and write the counting rule at the top of the report so next year's version is comparable. Then start the list of sentences. It grows by roughly one line per hire, and within a year it is the one piece of channel evidence in the building that no aggregator, autofill tool or vendor rep can inflate.

See a sample report >