Couple reasons. Some folks already mentioned below.

- Manual review - we review all data to filter extreme outliers and spam.

- By-Level info - we collect much more info per data point meaning we can hone in on title / role deeply.

- No Incentive - Not specific to Indeed but there is not incentive to submit your data. Everyone has access to data whether you submit or not. This along with the form being long leads to less spam.

- Staleness - we age out old data aggressively

- Total Comp - we're centered around total compensation. Most sites focus on salary which for tech jobs can sometimes be smallest portion of comp.

- Userbase - we're quite popular with tech demographic and have been able to field much wider range of data from folks at all levels due to it.

...there may be other reasons but these are the largest I can think of off hand

Not sure where indeed is getting their numbers, but I can assure you levels.fyi are well calibrated. Glassdoor also has weirdly low estimated salaries.

Has anyone done the sums of the numbers of people on l1->l9 vs the mean salaries estimated?

What does the estimated salary bill for each of these companies come out as and how does that stack up vs. their public declarations?

I haven't done that specific analysis but stock based compensation is given out in public quarterly financial statements. See here for example that Google spent ~9.35 Billion (!) in stock compensation last year: https://www.macrotrends.net/stocks/charts/GOOG/alphabet/stoc...

Glassdoor tends to underestimate because a lot of the datapoints are from several years ago.

Is it possible Indeed is only counting base salary, and not stock and bonus?

That's what I thought, then I looked at their estimates for Netflix Senior SW Engineer and found them to be similarly way , way too low. At Netflix, bulk of total comp is base salary -- bonuses and stock grants are just not done.

From talking with friends, the levels.fyi estimates are nearly perfect, indeed and glassdoor are way off.