Determining statistically significant conversion rate for an Ad example I have 3 ads with the resulting data.
Ad 1, 4269 Impressions, 26 link clicks, .61% conversion
Ad 2, 3155 Impressions, 12 link clicks, .38% conversion
Ad 3, 2510 Impressions, 9 link clicks, .36% conversion
Should I be doing a chi-squared test to understand the effect of ad copy on conversion AND then to determine the best ad, do a test between two proportions (Ad 1 and Ad 2) to determine if the best one is statistically significant?
 A: Statsstudent makes a great point when they say "I would use a logistic regression".  The result of a chi-square test tells you that a difference exists, but you have no way of determining which the difference is without resorting to post hoc tests which could possibly deflate power.
However, with a logistic regression, the comparison becomes richer.  Here is the analysis in R.
library(tidyverse)


ad = tribble(
  ~'Ad', ~'Impressions', ~'Clicks',
  '1', 4269, 26,
  '2', 3155, 12,
  '3', 2510, 9
) 

model = glm(cbind(Clicks, Impressions-Clicks) ~ Ad, data = ad, family = binomial())

100*predict(model, type='response')


The benefit of using a logistic regression is that we not only do we get conversion probabilities (0.61%,  0.38%,  0.36% for the ads respectively), we also get uncertainty in that estimate via confidence intervals.  For this experiment, here are 95% confidence intervals for the conversion
     p  low high
1 0.61 0.41 0.90
2 0.38 0.21 0.68
3 0.36 0.18 0.70

So, the first ad's results are consistent with conversions as low as .41% and as high as .9%.
There seems to be some considerable overlap between the confidence intervals.  Let's look at our model coefficients to determine if ads 2 or 3 yielded a statistically significant difference as compared to ad 1.

  term          p.value
  <chr>           <dbl>
1 (Intercept) 6.68e-148
2 Ad2         1.76e-  1
3 Ad3         1.70e-  1

Ad 1 is the reference ad, so the tests for this model compare ad 2 vs ad 1 and ad 3 vs ad 1.  In both cases, we fail to reject the null (p>0.05) and so we can't conclude that ad 2 or 3 results in smaller conversion than ad 1.  In this particular experiment, ad 1 yielded superior conversion, but in a different sample that might not be the case.
I'm being a bit fast and loose with my methods here, but this is the gist.  A little more effort yields a richer set of inferences.
A: Moving beyond a 2-way contingency table introduces a lot more statistical baggage, although is probably the "right" way to handle this.  You'll want to calculate chi-squared tests for each pairwise combination (ad1-ad2, ad1-ad3, ad2-ad3 in this example).  There are tests (like Cochran–Mantel–Haenszel) that endeavor to address this (you can find R packages to handle, or just build a hacky iterative python implementation if you want).
You can always do a chi-squared test with the dataset you've outlined (p1 == p2 ==p3); with that you'll only have a basis for a null hypothesis that all are equal, and it will have a bunch of mathematical shortcomings when you look under the hood.  Given the ratios you've detailed, it doesn't look like you'll have much to hang your hat on trying to assert that any of the versions generate meaningfully different results (not uncommon in ads/CTR's).

For most details: three-way contingency tables
