To me the term scientific (bear with me;it's been a few years since Stats 101) with reference to polling means;
1. The sample was large enough and random enough to be a valid sample of the entire population polled, and
2. The questions were free of bias.
For instance, with regards to point 1, if the total population is 100, and comprised of 50 guys and 50 gals, a poll of 10 guys is invalid, as it is unrepresentative of the total population. A poll of 5 gals and 5 guys would be more accurate.
With regards to point 2, if the GOP called you up and asked, "If the election were held today, would you vote for our courageous Commander-in-Chief, or that more-liberal-than-Teddy, French speakin', Hanoi Jane smoochin' Kerry?", the bias would render the poll invalid.
Actually, if they asked, "If the election were held today, would you vote for Kerry or Bush?", that question still has an unacceptable level of bias, as it ignores Nader, the Libertarians (that one's for you, PB!), etc.
You'll see the confidence level in the poll expressed as "the poll has a margin of error of +/- x points." x should be 1 to 3 points, IIRC. As x increases, the value of the poll diminishes greatly.
Harris, AP, etc. work the stat models quite hard to produce accurate, scientific results. If AFA commissioned a reputable outfit to conduct their poll, I'd assume they work just as hard to produce scientific results, or they would not be in business for very long.
Also, in this context, scientific means an independent agency can replicate the results.
Again, the polling outfit would have to play straight pool here, or risk having the bejesus sued out of them.