Forty one percent of our own database was a single number

Somebody searched Columbia University on this site and found no benefits. Not a thin page. No benefits at all.

We looked. Columbia held exactly one fact: a wage figure. No leave, no health cover, no retirement.

Then we counted how many others were like it.

That is 41% of a benefits database with no benefits in it.

Where it came from

Almost all of them arrived through one route. A collection run read US Department of Labor visa filings, which are public and structured and cover thousands of employers at once, and built a record for every company in them. It was a fast way to get broad. It never went back for the benefits.

The single fact those pages carried was median filed wage, and that number deserves its own warning. It is the median of wages an employer certified for visa-sponsored roles. At a retailer of sixty thousand people it can be the median of forty-one filings, none of them anyone on a shop floor.

We had a page for Nordstrom whose only content was a $165,906 figure describing about forty people.

What we changed

Three things, none of which was "collect more".

The pages now admit it. If we hold one fact or none, the page says so and explains what the one figure actually is, rather than presenting an absence as a slim offer.

A page holding nothing stopped pretending to be a comparison. Putting an uncollected employer next to a documented one and rendering $0 says they give you nothing. Not researched and gives nothing are opposite claims, and we were printing the wrong one.

We found records that existed and were invisible. Nine companies held four or five sourced legal entitlements each and never appeared, because the publishing rule counted the wrong statuses. Those were real facts, withheld by a bug.

The uncomfortable part

The honest reason this happened is that breadth is easy to measure and depth is not. Two thousand companies is a number you can put on a homepage. A median of three facts per company is a number you have to go looking for.

That is better and it is not good. We are publishing it because a benefits site that hides its own coverage gaps is doing the exact thing it was built to stop.

Every page on this site now tells you what we do not know about that employer. That is not a limitation we are apologising for. It is the product.