These figures represent the root mean square background measurement noise per unit bandwidth, in other words if you had an ideal bandpass filter placed after the instrument whose bandwidth is B [Hz] around some frequency at which those specified figures are given, say, $\sigma [T/\sqrt {Hz}]$ then the output will have some additive fluctuation whose standard deviation is $\sigma \sqrt{B}$
If your $\sigma$ is frequency varying, which is a common occurrence for no other reasons than that all instruments have $1/f$ noise, and you characterize the measuring instrument before it is sampled by a linear transfer function $H(f)$ normalized so that its peak is $max|H|=1$ then the total mean square fluctuation is
$$\sigma_M^2 = \int_0^\infty \sigma^2 (f) |H(f)|df$$
This $\sigma_M$ is then the rms fluctuation created by the instrument itself and is added to the actual measured sample. (Here we, of course, assume tacitly that the instrument noise is additive and Gaussian, as is indeed case usually.)
The equation for $\sigma_M$ is dimensionally consistent because if $\sigma (f)$ is measured in $[T/\sqrt{Hz}]$ then $\sigma^2(f)$ is measured in $[T^2/Hz]$, $H$ is dimensionless for it is normalized to $1$, so the dimension of $\sigma_M$ is then $[T]$ as it should be.