The following analysis path is probaby the quickest to lead you from your MEG data to presentable source estimates, with some compromises with respect to accuracy and statistical analysis. The general idea is this:

1. interpolate your MEG data to a standardised sensory array
2. apply statistics in "signal space" (e.g. using SensorSPM), in order to detect significant contrasts and latency ranges
3. run source analysis using a standardised head model on the grand-averaged MEG data (e.g. using a commercial software package, offering several different approaches to explore together with good visualisation)
4. possibly apply the same source analysis on individual subject data, e.g. for statistical analysis in "source space"

This closely resembles the analysis strategy employed for ERP analysis, where electrodes are usually placed at standardised positions (e.g. the "10/20 system"), such that data for different subjects are very easy to combine (e.g. for grand-averages, statistics on the same eletrodes etc.).

Although using individual structural MRI data sets for source estimation, and applying group statistics in source space on individual subject data, is in principle the most sophisticated approach to EEG/MEG data analysis, there are also some limitations that should be considered before choosing the most efficient analysis path:

1. Some source estimation procedures rely on rather restrictive modeling assumptions. This is the case for dipole models, for example, which make assumptions about the number and approximate locations of sources. These assumptions may not be fullfilled for single-subject data, either because of high noise levels, or because of inter-individual variation of the generator distributions.
2. Some distributed source methods make weaker assumptions about the number active sources, but still implement constraints on focality of sources (e.g. L1-norm, multiple sparse priors). These methods may still require a high SNR in order to produce accurate results, which may not be achieved for single-subject data.
3. Source analysis on the single-subject level can be veeeeery time consuming, in particular if realistic head models and cortex reconstructions are involved. Depending on the research questions, testing the predicitions underlying the experiment may not require this effort.
4. Analysing grand-mean data in signal and source space can provide clues about the optimal analysis strategy on the single-subject level, e.g. with respect to latency ranges, optimal source models, ROIs etc.
5. Many MEG papers do not report any data in signal space (e.g. time course information), which could provide important information about data quality. Furthermore, since different papers often use different source estimation procedures, it is often difficult to judge whether different studies replicate each others' results.

Analysing MEG data in a common sensor frame for all subjects, applying statistics in signal space, and using a standardised head model for source analysis of grand-mean data, can therefore be recommendable for the following reasons:

In MEG recordings, the position and orientation of the sensor array usually differs significantly across subjects (unless special precautions are taken). Until recently, it was not straightforward to interpolate MEG data from different subjects into a common sensor array (e.g. the average across subjects). It has therefore been the standard to analyse MEG data only in source space, i.e. after source estimation procedures have been applied on single-subject data. This is the basis for some of the "fMRI-like" analysis strategies, as e.g. implemented in SPM5.

1. Differences in MEG signal indicate differences in the source distribution. Differences detected only in source estimates, but not in the MEG signal, would be suspicious (although theoretically possible).
2. Anatomical normalisation of source estimates across subjects will lead to a "blurring of results", similar to interpoating MEG data to a common sensor frame.
3. To be continued.