Yup, that was a great suggestion, thank you.

Modifying the raw source as follows, I managed to bring it down to 181 characters:

for($r=99;$r;--$r){$s=$r==1?:"s";echo "$r bottle$s of beer on the wall, $r bottle$s of beer.
Take one and pass it around, ",$r-1," bottle$s of beer on the wall.

";}echo "
No more bottles of beer on the wall, no more bottles of beer. 
Go to the store and buy some more, 99 bottles of beer on the wall.";

Basically, this tries to give the dictionary compressor a few more opportunities to match. That's harder to get right (and less intuitive, longer text sometimes compresses to smaller output!) than it sounds. Not only is the compression result a bit hard to predict from the editor, but also better compression may be larger in the end due to escaping.

The level parameter won't change anything, the to-be-compressed string is far too short for that to have any effect (not that I didn't initially try a brute-force packer that chose the smallest of every possible level anyway, but that turned out a complete waste of time ).

As for not escaping all those characters, this was my first thought, too. Actually I thought that escaping \0 as well as single and double quotation marks should be enough. Then it occurred to me that backslash probably needed to be escaped, too. Turned out that still isn't enough, gzinflate will choke on tab and $ as well.

Further, there is indeed at least one character that I'm not escaping although I should to be "formally correct" (I didn't bother to find out which, as the result was longer than the best match anyway -- probably it's the backspace character or some other weird control code?).

I discovered this when trying to rename the loop variable to something that gives more matches than "$n". My first idea was "$e", seeing how e is the most frequent character in English. That was too naive and didn't work out (first, in this particular text, "o" is the most common character, and second it doesn't suffice to be frequent, you also need surrounding characters to match -- which made "r" being the most frequent end-of-word character the best candidate), But what's worst, it broke the decompressor.

Now "r" saved 2 bytes, which is nice, so I stuck with that. And incidentially, it works fine with the small subset of escaped characters, too.

The brackets are need around the ternary operator since it's in a macro.

#include <iostream>
#define a(i) << " bottle" << (i-1?"s":"") << " of beer" << 
#define w " on the wall" <<

int main()
{   
    for(int i=100;i-->1;)
      std::cout << i a(i) w ", " << i a(i) ".\nTake one down and pass it around, " << i-1 a(i-1) w ".\n\n";
    std::cout << "No more" a() w ", no more" a() ".\nGo to the store and buy some more, 99" a() w ".";
}

After some tweaking 273 now

Sadly, it fails for 1 bottle of beer.

C# 230 characters