The FEC does have an official report of vote totals, state by state.I think Atlas has over-counted the write-in votes by about 433,000 nation-wide. There is no difference between the vote totals that Atlas has compared to the FEC for the states of AL, AK, AR, CO, CT, DE, HI, IA, LA, ME, MD, MA, NE, NV, NH, NM, NC, ND, OK, OR, RI, SC, SD, TN, VT, WI, or WY. That's 27; a little more than half. I think Atlas goes the data at the county-level in order to include, among the votes cast, all of the write-in votes that the counties tallied but which the statewide election official did not tabulate.
I'll explain the difference between tally and tabulate, per what I just said. To tally simply means to count and report how many total write-in votes were cast. There were 100,000 write-in votes in the entire state. To tabulate means to report that, of those 100,000 votes, 22,000 were for Evan McMullin, 20,000 were for Bernie Sanders, 18,000 were for John Kasich, and so on. At the low end of the tabulation, there were some write-in votes for Pee-Wee Herman, Captain America, Darth Vader, and so on.
I'll use Missouri as an example. The largest counties, and many other smaller ones, have a count of all the votes cast, and that can be found online, such as
here. Notice that it says there were 6,458 write-in votes, but it says to "See the official write-in report," which is
here. About 3/4s of the counted write-in votes were invalid - 4828 - which is to say that voters wrote in names of people who were not "declared write-in candidates" -- in Missouri, only McMullin, de la Fuente, Hoefling, Kotlikoff, and Schoenke were "declared" candidates. Writing in names like Bernie Sanders or Jeb Bush gets your ballot treated as "invalid" in Missouri. But I think Atlas counted all of those invalid write-in ballots and included it in the total for Missouri. The Missouri SoS reported, to Washington D.C. (and hence, to the FEC), that the total number of votes cast was 2,808,605. Atlas added about 19,660 write-in votes that Missouri SoS does not include in the total, I assume because whomever did the research over-enthusiastically went to look at every single county's vote totals, disregarding the fact that those 19,660 were treated as invalid write-ins. I think Atlas has done that to a number of states.