We present results from a multisite (`Whole Earth Telescope') photometric campaign on PG 1336-018, the close eclipsing binary system containing a pulsating subdwarf B (sdB) star. The main part of the campaign (1999 April) resulted in ~172 h of observations, representing a coverage of about 47 per cent, and additional data were obtained outside the core campaign. Periodogram analysis shows that the light variations are dominated by three frequencies near 5757, 5585 and 5369 μHz (~174, 179 and 186 s, respectively), although many frequencies are present, particularly in the range 5000-6000 μHz (~200-170 s). We identify, with some confidence, 28 frequencies down to a semi-amplitude of 0.0005 in fractional intensity (equivalent to about 0.5 mmag). It is clear that the pulsation frequencies of PG 1336-018 have changed substantially since the 1996 discovery observations were made, and that amplitude changes occur, at least in the dominant three frequencies, on relatively short time-scales (of the order of a day). On the assumption that the pulsating star is phase-locked in the binary system, we have searched for rotational splitting of frequencies near the orbital and half of the orbital period, but the results are confused by aliasing at those frequencies (due to the data gaps caused by the eclipses). A preliminary model qualitatively matches the distribution of frequencies in PG 1336-018, with some good individual correspondences, but cannot be considered adequate because geometric cancellation should hide some of the modes which are apparently detected. Analysis of the pulsations during eclipse recovers three of the strongest modes, but the limited eclipse data - which can, at best, be only about 9 per cent of the total - do not allow mode identification at this stage. Simulations indicate that an overall coverage of about 80 per cent would be required for this to be viable. An attempt was made to determine phase shifts in the pulsation frequencies as a way of directly measuring the size of the binary orbit, but the uncertainties in the method are comparable to the light travel time across the orbit (probably less than a second).