When do you decide that a kwd is unprofitable?

a statistics question ....

[Algebrator]

I have recently started adding specific kwd phrases to (already profitable) general ones.
Here is a (simplified) description of situation that I find myself in:
There is an ad campaign with 1000 different , very specific kwd phrases. These are low frequency phrases but very targeted; each one costs let's say around $2.00 a month (i.e. avg of 10 clicks).
Within a single month (first month of operation) I get 10 orders, from 10 of the keywords (I know which ones they are), $100 each.
So, I am spending $2000, making $1000 --> not good. However, if you look at just the kwds that produced orders, profit margin is excellent ($100 order per $2 spent).
Common answer (eliminate unprofitable ones) doesn't really apply here because looking at each keyword separately it is too early to tell whether it is going to be profitable or not.
So my question is - at which point should I start eliminating them? How long should I "weather it out" in order to get statistically good enough sample , where I can definitely say this kwd is no good?
I want to make this decision based on profit/cost ratio, not impressions, CPC, CTR. I just don't know where to draw the line.

[loanuniverse]

I don't think is possible since there have been only 10 conversions in the case that you bring up. Now this is probably wrong, but the only thing that occurs to me is that you should be bidding much less in an scenario like the one you mention.

Lets say you go down to the minimum PPC of $0.05 from the $0.20 that you are paying. If you still keep the ratio of conversions you will be ahead of the game.

I am fresh off a graduate level statistic class and can't figure out how the heck to approach this with only 10 conversions.

One thing, I am certain... if one of those keywords converted more than 1 visitor, I suggest you increase the PPC for that one :)

[cline]

Before the internet I was a direct marketing statistician.

You have a common, albeit difficult, direct mail analytical problem.

Segment your key phrases
You've got to aggregate your data to get a higher level view. You say you've got 1000 keyphrases, 10 of them with orders. You need to thematically cluster them to get them to fewer than 10 groups. You'll have to do business judgement to do this, and there may be multiple reasonable groupings (in which case, do them all), but this analysis should help you identify unprofitable themes -- or more conservatively, identify the ones that are most likely profitable.

Your economic structure is wrong
If, as you say, costs are double gross profit, then your initial economic calculations are way off. Potentially you're bidding too much. Or you're targeting keyphrases that are too marginal. Or your ads are too sexy -- you're not delivering what you promised (alternatively, your site is too lame and doesn't fulfill its promise)

BTW, don't get overconfident about those 10 terms that brought in sales. You're dealing with a huge error term that's making those 10 look vastly better than they really are.

[Algebrator]

Thank you for your input.
You are correct in assuming that terms are marginal. I have already exhausted the directly related kwds; i.e. "blue widgets", "red widgets", "any kind of widget" - these have been consistently profitable. I am trying to increase exposure with keywords phrases that are related to "widget" but where "widget" is not an actual part of the keyword phrase.
The suprising fact is that I first tried this on Overture, with exact same kwds and these "marginal" kwds increased my profit by 50%.
I guess Google audience is different.
My overall campaign even in Google is quite profitable, so I am just thinking about taking a hit in the name of research on this particular part of it, and running it for a long period of time to get a better sample. (the data is actually from the first of January, not for the whole month).
I am assuming that in couple of months (within a year?) I should start seeing some coversion clustering around particular keywords. Wouldn't that be the only sure way of telling what is profitable and what is not?
Making kwds aggregate (less specific) would get me quicker results, but it seems, not as accurate.
I should note that in my business, things remain very steady, so information that I gather should be valid for a very long time.
I understand that margin of error is now extermely high. At which point does it become acceptable?

[buckworks]

You said the phrases were specific but you didn't say if you were using exact match. So I have to ask: Are you making use of negative keywords to block impressions (and therefore clicks) that are likely to be unproductive?

If widgets are related to doodads, it makes sense to promote your widgets with doodad phrases, but remember to block all those off-target searches for "free doodads", "doodad clip art", "The Doodads mp3" etc.

If you haven't done so already, it's worthwhile to spend some time messing around with a suggestion tool looking for terms to block as well as to bid on.

[cline]

Algebrator, it's not unusual at all to see big differences in performance between Overture and Adwords.

Yes, you're in a situation where you have to pay to do some unprofitable marketing in order to find some new profitable targets.

BTW, when you do your forecasting of what will actually be profitable and should stay in the campaign, realize that there will be a big regression-to-the-mean effect that will cause your forecasts to be excessively optimistic.

[Algebrator]

Cline, thanks for the comment.
Ignoring my current situtation, at which point do you generally decide that a kwd is not profitable? i.e. let's say my general coversion rate is 2%. When do you say, "I am pretty sure that this kwd is no good" - it is not producing profit after "x" clicks. What is "x"? - 100, 500, 1000...

[FromRocky]

I use X=[Profit/sale]/CPC as a cut-off and it works very well for me.

[cline]

Algebrator, an old direct marketing statistician's rule of thumb is that 7 positive actions -- regardless of the total sample size -- produces forecasts robust enough for segmentation decisions.

So do the math backward to determine what sample size you need to get the 7. If you've accumulated half that, and you haven't got any positive actions, you're not going to hit 7 when you get the other half of the sample.

[danieljean]

cline- are you saying that you need 7 sales to figure out how many impressions you'll need on average, or that you need to be able to segment your sample size into 7 distinct groups?

[cline]

danieljean, neither. I'm saying you need 7 positive actions to get reasonable statistical confidence (it's a rule of thumb, not a statististical "truth"). It doesn't matter whether it's 7 out of 100 or 7 out of 1,000 or 7 out of 10,000 or 7 out of 100,000.