How we know what we tell you

What our own sample is

30 profiles, audited by us, counted as of 31 July 2026.

30 profiles is enough to say a problem is common among people who ask for help with their profile. It is not enough to say anything about the Upwork marketplace, about a category, or about a country, and the statistics page repeats that limit above the numbers themselves.

How an observation becomes a claim

Five steps, in this order. A claim that cannot survive step four does not get published.

1. 01

   Read the profile against a fixed rubric

   Six sections, the same six every report grades: Title, Overview, Portfolio, Skills, Employment History, and JSS with badges. Each criterion is a yes or a no, recorded before any judgement about what to do next.

2. 02

   Check it against what the platform says

   If Upwork has published something on the mechanic, their wording wins over ours, and the figure is marked as theirs wherever it appears.

3. 03

   Look for an experiment, not an opinion

   A claim about ranking needs someone who changed one thing and measured what happened. A blog post asserting the same thing is not that, however confident it sounds.

4. 04

   Read the source in full before citing it

   If the page cannot be reached, or the number is not actually in it, the claim comes out. That has happened, and the corrections below say when.

5. 05

   Keep field observation separate

   What working specialists report is logged as field observation. It decides what we look at next. It never becomes a number on a page.

Four tiers, and what each one is worth

Every claim carries one of these. They are not interchangeable, and the difference between the strongest and the weakest is the whole point of labelling them.

• platform

  Upwork's own statement

  Something Upwork published: the help centre, product documentation, or an official announcement. We are reporting it, not measuring it.

  Strong on what Upwork says. Silent on whether it holds in practice.

• experiment

  A published experiment

  Someone changed one variable, measured the result, published the method, and we have read the whole thing. One source currently carries our section-weight ordering: a 23-profile test published in October 2024, in which a made-up word was placed in different profile sections and the placements ranked in a fixed order.

  levit.tips

  Strong on relative order under the conditions tested. Weak on anything outside them.

• original

  Our own audit counts

  Counts from the 30 profiles described above. Exact, because they are counts of what we recorded, and narrow, because the sample is small and self-selected.

  Strong on what we saw. Silent on the wider market.

• field

  Field observation

  Patterns from consultations and day to day work with people who run Upwork profiles professionally. The weakest tier, which is exactly why it is named rather than blended into the others.

  Directional only. Never published as a number.

What we cannot claim

This is the section a reader should check us against, so it is written plainly rather than hedged.

That we know how Upwork ranks

We do not have the algorithm. We have public behaviour, published statements, one experiment and patterns from our own audits. Anyone claiming a reconstruction of Upwork's matcher, us included, would be guessing.

That the weights are exact

The experiment we cite established an order, not a magnitude. “Employment History ranked above Portfolio” is what it shows. “Employment History is worth N times Portfolio” is not.

That the ordering still holds today

It was measured in October 2024, with one token. Upwork has changed several things since, including switching Specialized Profiles off in May 2026. An order measured before a change is evidence about before the change.

That our counts describe the marketplace

30 self-selected profiles describe 30 self-selected profiles. Someone who requests an audit has already decided something is wrong, which is not how the average profile arrives.

That the reply-rate bands are causal

The bid-timing figures are observational. Nobody randomised anything. Fast bidders differ from slow bidders in ways other than speed, and those differences sit inside the numbers.

That fixing our findings raises your rank by a set amount

We have not run a controlled before-and-after, so we put no number on it, and neither should anyone quoting us.

What would prove us wrong

A claim nobody can test is not a claim. Each of these is the specific result that would force the corresponding statement off the site.

• we sayEmployment History is the heaviest ranked section

  falsified byA replication of the made-up-token test, on twenty or more current profiles, in which the token ranks higher from Overview or Title than from Employment History. Or Upwork publishing a different order.
• we sayTestimonials and Education do not move matching

  falsified byThe same test showing a placement in either section ranking at all.
• we sayBelow 80% JSS you drop out of organic search

  falsified byA documented profile under 80% appearing in Talent Search results that were not reached through a direct invite.
• we sayBidding within five minutes replies at a higher rate

  falsified byA tracked cohort of two hundred or more proposals in which the sub-five-minute band replies at parity with the sixty-minute-plus band.
• we sayThe problems we count are common

  falsified byAn audit of a comparable number of profiles drawn at random rather than self-selected, finding the same problems at materially lower rates.

If you have run any of these, write to us. A result that contradicts us is worth more to this page than another one that agrees.

What we have already taken back

A correction policy is only worth reading if it has been used. These are dated, and they stay here after the wording on the rest of the site has changed.

• 2026-08-16

  was“Updated weekly”

  nowThe knowledge base's real revision date

  The base is revised when Upwork's behaviour changes, which is not a weekly schedule. A cadence we did not keep had no business being a promise, so every page now shows the actual date instead.

• 2026-08-16

  was“Controlled experiments show…”

  nowA named experiment, with a link and a date

  The claim matched what we had read, but it named no source, which makes it impossible for a reader to check and therefore worth nothing. Every occurrence now cites the specific test.

• 2026-08-16

  wasA third-party figure about how much more likely a complete profile is to be hired

  nowRemoved

  We could not find it in the source it was attributed to. Rather than soften the wording, we took the figure out.

Found something wrong?

Every claim on this site has one person's name on it. If a figure looks wrong, or you have data that contradicts it, write. Corrections get dated and published in the section above, including the ones that do not flatter us.