Yeah, the GAO study was purely based on the U.S. Extremist Crime Database and that's clearly a limitation, although less of one than reading through a list of notable attacks on a wiki page. The page itself also is generally about terrorism, so it talks about both international and domestic throughout it which can be confusing. That's why I suggested looking at the actual studies linked to on the page, a lot of them do a better job of pointing out their time period and limitations than the brief overview on the page.
Also, the page only has pretty good resources up until like 2020, where it ends with a study from the Center for Strategic and International Studies which reviews data up to May 2020, and DHS which reviews data from 2018-2019. The CSIS one is pretty good because it includes graphs of data over time and really shows the worrying increase across the board but the staggering increase of "right-wing" violence since the mid 2010s.
CSIS has a few more studies more recently it looks like. There's https://www.csis.org/analysis/pushed-extremes-domestic-terro... from 2022, which shows that 49% of events were committed by far-right and 40% were far-left. However, the far-right were more likely to target people with guns and bombs and the far-left were more likely to target property with melee and incendiary weapons, so 28 of the 30 deaths were from the far-right while the far-left accounted for 1.
Then a few more years later, there's https://www.csis.org/analysis/rising-threat-anti-government-.... Which is more about the increase in "partisan political belief" based attacks, and then gives some examples instead of breaking it down further.
However, CSIS likewise uses their own database of attacks, and in between the other studies and the most recent one it appears they changed their methodology of what attacks were included to make it more strictly about an attempt or threat to kill (which would remove a lot of the property based attacks from the previous study), premeditation, and desire to strike fear broadly. I'd be interested in seeing a revisitation of their previous methods with their new datasets, or even to actually be able to see the dataset itself.