ARDL is a better method. Here is the explanation: If your variables are all stationary at first difference, then both these methods are equally usable and u should stay indifferent. However, if one or more of the variables is/are stationary at level while others are at first difference, then Johansen Method can not be used but you can still use ARDL to examine the long run relationship (co-integration) among the variables. In short, Johansen is specifically used when all the variables under study are stationary at first difference, while ARDL is a general method and is useable even if your variables are stationary at different levels [I(0) and I(1). 

all depends on the variables characteristics. read both techniques requirements.

I second Denbath, Auto Regressive Distributed Lag (ARDL) Model allows you to analyse the long run relationship amongst variables integrated of different order e.g I(0) and I(1), which you cannot do with the Johansen method.

Use both methods. Stock and Watson ("A Simple Estimator of Cointegrating Vectors in Higher Order Integrated Systems," Econometrica, 1993, p. 811) advise us to use more than one method when testing for cointegration. They advocate (see p. 811 again) the DOLS estimator, which you might also want to try. Hayashi (Econometrics, 2000, pp. 662-665) provides a nice example. Note also that in the Johansen method it is NOT necessary that all of your variables be I(1), as the previous commenters have indicated. According to Hansen and Juselius (CATS in RATS Cointegration Analysis of Time Series, 1995, p. 1), the argument that you can't have an I(0) variable in a cointegration relation is incorrect. Specifically, they write: "not all the individual variables included in z [my note: the p-dimensional vector of the VAR investigated for cointegration] need be I(1), as is often incorrectly assumed. To find cointegration between nonstationary variables, only two of the variables have to be I(1). Often, a stationary variable might a priori play an important role in a hypothetical cointegration relation, for instance, an inflation rate. In particular, variables with a high degree of autocorrelation, also called near-integrated variables [my note: but still I(0)], are often very important in establishing a sensible long-run relation. Note that for every stationary variable included, the cointegration rank will increase accordingly."

@ Dimitriv...thanks for the debate, will review

It also depends on the sample size. ARDL is basically more flexible to small sample sizes than the Johansen method.

Dear my friend!

It is hard to say which model is more convenient.. It entirely depends on your variables and sample size.

ARDL BOUND TEST gives you the desired results for cointegration with different orders I(1) or I (O) .

ARDL is more fitted for time series of different orders unlike Johansen-Juselius approach which requires all the time series in a model to be of first order. However, the two methods can be employed in a study to check for consistence of estimates most especially where the requirements of the two methods are satisfied.

Saka Kamilu: "unlike Johansen-Juselius approach which requires all the time series in a model to be of first order."

This statement is imprecise. According to Hansen and Juselius (CATS in RATS Cointegration Analysis of Time Series, 1995, p. 1), "not all the individual variables included in z [the p-dimensional vector investigated for cointegration] need be I(1), as is often incorrectly assumed. To find cointegration between nonstationary variables, only two of the variables have to be I(1). Often, a stationary variable might a priori play an important role in a hypothetical cointegration relation, for instance, an inflation rate. In particular, variables with a high degree of autocorrelation, also called near-integrated variables [my note: but still I(0)], are often very important in establishing a sensible long-run relation. Note that for every stationary variable included, the cointegration rank will increase accordingly."

Your data will guide you.

Your sample size, and the integration order of your variables will be a good starting point as suggested by other researchers.