Detailed comments can be found in the supplement.

L23:  Given the large bias errors in these data sets, comparing the population exposed to a threshold value, like 50 ppb is meaningless.  Do these differences impact "health" as stated or one's analysis of health effects.  This is a bit sloppy writing. 

L25: very good point, but you should compare the gridded data you have with Schnell's gridding of the EU and NAm.  Comparing points to grid-cell averages that you have from the global data sets is a serious science problem -? not the way you treat it here.

How can you get an R2 for surface sites vs grid-cell means?? This is not sensible.

L27:  You just said your data is worst at overestimating O3 at low abundances, but here you say it is worse for >50 ppb??

L29:  "highlighting the importance of continued research on global ozone distributions"

L38:  Oh really.  The number of regions could be much much greater if you picked smaller regions.  The key issue is the area fraction NOT the number of regions.

L42: 30 ppb is basically the minimum background level – this is not a useful statement and it implies that pollution is the cause here.

L44:  The quality of writing (logic, not English) is poor:  You just quoted all these results that rely on estimates of surface ozone and then you say you lack knowledge of surface ozone.

L70:  Great.  This is the most important statement.  Could be up front

L71: "Potential" – there are most assuredly inconsistencies.

L75:  the biases and errors certainly come from the process.  I hope you are not using 'data' to describe the assimilation modeling here.

L85-95:  This exposes the fundamental flaws in the focus of this paper.  The use of OSDMA8 totally washes out the key fundamental information about ozone that can be tested with the real surface direct observations.   The 24-hour diel cycle is a must that needs to be simulated in any modeled ozone product (all of your six sets are modeled products).  Likewise the variability of ozone (including MDA8) is critical in evaluating health/agric. impacts and there needs to be a test of your six 'sets' as to their ability of match extremes.

L195ff:  ibid.  This is a mistake to smooth out the fundamental ozone cycles (diel and synoptic).

L220ff:  "We adopted a point-to-grid evaluation approach, where the data from each TOAR-II observation site was matched with a corresponding grid cell in each dataset. For grid cells with a TOAR-II observation but no valid estimate in a dataset (NA value), we used the nearest valid estimate instead."  This seems to ignore the previous TOAR-related work by Schnell where for the high-density of surface sites in EU and N.Am., a 1ºx1º grid-cell averaged, hourly surface ozone product was created. 

This data set was used to assess extremes and to test the CMIP model's accuracy in seasonal and diel cycle of ozone.  The cell average is the only way to do a fair comparison with the surface sites because of their irregular – sometimes oversampling and sometimes under sampling – many regions.  Comparing surface sites with model cells is dangerous, especially since in this paper their appears to be a lack of understanding of the problems with this approach.  The Schnell data are the obvious choice to validate your six model-data sets, even if it is only for EU and NAm: 

doi:10.5194/acp-14-7721-2014

doi:10.5194/acp-15-10581-2015

doi:10.1002/2016GL068060

doi:10.1002/2017GL073044

doi: 10.1073/pnas.1614453114.

Then you can go after the rest of the world (which is very important).

This paper is based on comparing 6 different modeled surface ozone dataset with one another and with the TOAR set of surface sites (Table 3).  The comparison of individual sites with grid-cell averages fails to recognize the difficulty of the task and ignores the extensive efforts to develop unbiased grid-cell means from high-density observations.  The authors further corrupt the data set by averaging and smoothing to destroy the fundamental information on ozone variability that is critical for testing the modeled ozone products.  The use of these 6 sets, varying in resolution from 0.1 to 2.5 degrees, to map population exposure is premature. 

I can not recommend publication of this work as is.