It does sound like stats one way or another, but as your non-ANSI joins are not simple equality checks, you could easily end up with a more stable plan by converting these to ANSI joins. I've given a quick example below - please check the results before using:
SELECT pixel ,
SUM(imps) impressions ,
FROM #temp4 t
INNER JOIN VINDICOSUITE_NETACUITY.dbo.GeoDatabase geo WITH ( NOLOCK) ON t.ipNumber BETWEEN geo.ipFrom AND geo.ipTo
INNER JOIN VINDICOSUITE_NETACUITY.dbo.Metro dma ON geo.[metro-code] = dma.metroKey
WHERE geo.[metro-code] NOT LIKE 'metro-code0'
GROUP BY syndicationOutletId ,
I'd also suggest not using the NOLOCK clause if you can help it. It's pretty dangerous unless you don't mind the results being inaccurate.