Two questions:

At what point is the poll statistically meaningful?

When should I publicly identify the sources?

There may be other options, but, off the top of my head, I think you could use either the binomial or the one-dimensional goodness-of-fit chi-square test. The binomial test, an exact test, asks whether or not a single proportion differs from the expectation; the common example is to see if the number of heads you get doing a series of coin flips differs from 50%, but you can also dichotomize other questions. The common example for that is to see if the number of ones you roll on a die differs from 1/6 (so number of ones in a given number of rolls). Given the current poll numbers (5, 18, 7), you would test 18 votes for file 2 out of 30 total votes with an expected frequency of 1/3 (33%). In the R implementation, you get

*P* = 0.003 with an estimated true frequency of 0.60 and a 95% confidence interval between 0.41 and 0.77. [In R: binom.test(18,30,1/3)]

In the chi-square test, you look at whether the observed distribution of proportions differs from the expected distribution of proportions. For the poll, you would expect a proportion of 1/3 for each of the three categories. Testing the proportions of 5/30, 18/30 and 7/30 against 1/3, 1/3, 1/3, you would get

*P* = 0.007. [In R: chisq.test(c(5,18,7)). Note that if you don't specify the expected proportions, R assumes that it is 1/(number of categories) for each category.]

Long story short, both tests show that file 2 was selected significantly more frequently than you would expect by chance. If anyone is really curious (which I kind of doubt!), I can post links or R outputs. And while I am at it

R is a really great, flexible, and free "language and environment for statistical computing and graphics".

Statistically speaking, the more samples the better, wih 30 samples being the minimum.

Definitely better to have larger sample sizes. The number 30, though, is a very general (and somewhat dubious) rule-of-thumb for continuously distributed variables in parametric analyses (unlike this poll). There are lots of caveats. For the binomial test, there are no real sample size requirements (although it's power will be dependent on the number of successes and trials and the expected proportion). For the chi-square test, a minimum of five expected observations per category (15 votes for this poll) is a well-accepted cut-off.