I don't think your question has a closed and easy answer. In a perfect world, you run your chipseq data (as aligned reads, bam or bed) through a peak caller, e.g. macs, and each region identified is a binding site, as mentioned by tangming2005.
However, the situation is typically far from perfect for a number of reasons.
The ChIP enrichment is often quite aspecific and noisy, depending on the quality of the antibody. Consider that it's not unusual to have >90% of the reads in the background, i.e. not in peaks.
Some genomic regions tend to be enriched with whatever antibody you use (an artifact that might be due to the way the reference genome is assembled, especially with respect to repetitive regions).
Different peak callers/algorithms might give different numbers of peaks, this difference can even be orders of magnitude. Same goes for using different parameters within the same peak caller
Typically, the more you sequence the more peaks you identify because small bumps that become significant.
Even if the ChIP works perfectly and the peak callers are ideal, there might be opportunistic sites where the transcription factor binds without having much biological relevance (as an aside, possibly related: some chipseq experiments generate many more peaks than genes in the whole genome).
In practice, you could consider as "true" binding sites the peaks which are identified in different replicates and/or which overlap a known sequence motif recognized by your transcription factor (see also the irreproducible discovery rate).
In my opinion, asking "Where are the binding sites?" is not fruitful for the problems above. Better is to ask which binding sites differ between conditions (might be treatments, stages, tissues whatever). This way the quirks associated to ChIP, peak callers etc are averaged out across replicates and conditions.
A practical way to decide if your peak is a true peak and not an unspecific binding is to check if there is a motif associated to your transcription factor at the peak. This can be done using the meme suite. Of course, this solution assumes that your ChIP is for a protein that directly binds the DNA.