Obviously the larger number of observations will provide robust results, however, if you don't have more observation you can carry with the available data. The results may not be robust (stable).
26 observations is a very limited sample, especially for cointegration when you are looking for a long-run relationship between variables. I agree with Fridos that the larger the sample, the more robust your results are. If really you cannot expand your data try to run your estimations and then check the stability and significance of your results.
I think you should try the "Bounds F-test", which proposed by Pesaran et al. (2001), the paper published on the Journal of Applied Econometrics. It's a very robust cointegration test, in particular suitable for small sample. EViews or Stata all can conduct this test.
It is important to remember that cointegration is a test of long-run relationship. In addition to the number of observations, the time span of these observations is also important. While 26 observations are definitely on the low side, the quality of your results will also be impacted by whether these are 26 monthly, quarterly, or annual observations. All else equal, it is better to have 26 annual observations than quarterly observations, which is better than 26 monthly observations, in a cointegration test.