We asked forty recruiters to review the same twenty profiles for the same fictional role, then watched where their eyes went and in what order. Half were agency, half in-house. Experience ranged from eighteen months to nineteen years.
The point was not to catch anyone out. It was to find out what a resume review actually is, because every product decision we make about screening rests on an assumption about that, and we would rather the assumption were checked.
Nobody reads top to bottom
Not one participant read a profile in document order. The dominant pattern, in thirty-three of forty sessions, was:
- Current title and current company
- The list of company names, read as a column, ignoring everything between them
- Total years, computed by subtracting the earliest date from today
- Then, only if the first three passed, back to the top to read anything in prose
The median time spent before the first accept or reject lean was 6.4 seconds. The median time spent on a profile that got rejected was 11 seconds in total. Profiles that got advanced received a median of 74 seconds, which is where nearly all the reading happened.
This has an obvious implication that we had somehow never stated plainly: for the large majority of candidates, the review is not a review. It is a company-name lookup with a duration check attached.
Company names are doing most of the work
When we asked participants afterwards what they had based the decision on, the answers were about skills, relevance, and trajectory. The recordings showed something narrower. Recognisable company names produced a measurable dwell increase on the rest of the profile; unrecognisable ones produced a scroll.
The gap between the two columns below is the finding. Nobody was lying to us; people are simply poor witnesses to their own attention.
| What they said they weighed | Said it | Gaze pattern supported it |
|---|---|---|
| Skills and tools listed | 68% | 22% |
| Relevance of recent work | 55% | 40% |
| Career trajectory | 43% | 31% |
| Company names | 12% | 79% |
The effect was strongest in the least experienced group and did not disappear in the most experienced group. It changed shape. Senior recruiters were more likely to recognise a small company as a strong signal, which is a better heuristic, but it is still a heuristic about the employer rather than about the person.
The attention curve is steeper than anyone admits
We ordered the twenty profiles randomly per participant, which let us measure attention against position rather than against the candidate.
Attention fell by roughly three quarters across twenty profiles. Twenty. The reqs these people work on have between two hundred and a thousand.
More uncomfortable: when we re-showed three profiles late in the session that had appeared early, eleven participants gave a different verdict the second time, and nine of those eleven were more negative. Nobody noticed they had already seen the profile.
What we changed because of it
Three things, all of them narrower than the study might suggest.
- Rank before anyone reads. If the first five profiles get four times the attention of the last five, the only responsible thing to do is make sure the first five are the ones most likely to deserve it. This is the entire argument for scoring a stage before review, and it is a stronger argument than the time saving that usually gets quoted.
- Lead with evidence, not with the company. Our review panel now opens on the criteria and the passages that satisfied them. The employment history is one scroll away rather than the first thing in the eye's path. Small change, and it moved reviewer agreement on the same profile from 61 percent to 78 percent in a follow-up round.
- Show the reviewer their own drift. Recording the verdicts is not only for the audit trail. If your accept rate over the last thirty profiles has fallen off a cliff relative to the thirty before, that is worth surfacing, because the alternative is finding out at offer stage that the back half of the pile never got a fair look.
The thing we did not change
We did not build anything that hides the resume. Several people, on hearing about the study, suggested the obvious follow-up: strip the company names, strip the dates, show only evidence against criteria.
We tried it. Reviewer confidence collapsed and review time went up, because people started hunting for the context they had been denied. Trajectory is real information. The shape of a career tells you things no individual line does, and taking it away does not remove bias so much as relocate it somewhere you can no longer see.
The better answer is not less context. It is making sure the context arrives after the evidence rather than instead of it, and that the person on profile three hundred is looking at a list that was already sorted by something other than the order the applications came in.
