When running search query reports for Google Shopping campaigns, we've noticed a large difference between the total number of search impressions vs. the sum of the impressions for the specific search terms listed. In fact, we're only getting data on 15% of the impressions. It looks like Google only reports on queries that result in at least some clicks in the reporting period. 85% of the time Google is showing our products, it's for search queries that will never get clicked on.
From Google's perspective, this makes sense. But for a retailer, it means we are missing out on information that could be used to optimize our shopping campaigns and product titles. For example, knowing which terms result in no clicks could provide us with a list of negative keywords that would help only show products for more relevant searches, thereby improving the click through rates, and reducing our necessary bid amounts.
Since we can't examine the unreported terms directly, we want to look for variations and correlations. If there is no significant variation in the Search Impression Ratio (SIR) of search term impressions vs. total reported impressions, then there is little we can do. However, if we see significant variation, there must be a reason; and if there is a reason, there is something we can do to improve our shopping campaigns and product titles.
If we discover variation, the next step is to look for two kinds of correlation. Does this ratio relate meaningfully to any of our bottom line numbers (total clicks, conversions, conversion rate, cost, CPC)? If it doesn't improve the bottom line, then it can hardly be worth spending much time on.
We can see that there are correlations with some of our key metrics, which means its worth trying to identify what affects it.
So next we check to see if it relates meaningfully to factors we can affect (search terms, campaign structure, etc), or is it the result of something more nebulous.
We do see some correlation with the amount of text in the search, but not enough to declare length of search term to be the primary causal factor. The correlation is logical given the two assumptions that a user who is most likely to click on a product ad is one who knows what they want, and someone who knows what they want is more likely to type in a longer, more specific search string than someone who is just browsing. For example, someone who types in "shoes" is less likely to be ready to click on a product ad than someone who typed in "high top athletic shoes".
Of course, total text doesn't give us much idea on how to improve, so let's split each search string into either product type (what the product is), attribute (features of the product), unidentified (what our system can't automatically assign meaning to).
While having a product type is extremely important for conversions, the number of characters comprising the product type don't seem to relate at all to the SIR. However, we see a definitely correlation with identified attributes, and a weaker correlation with the unidentified text which may be related to specific details about the products that we haven't classified.
In conclusion, search impression ratio is correlated to both CTR and an identifiable pattern in the search terms. We can use to learn that making sure to include the product attributes in the product titles will help improve SEO/SQR clickthrough rates.
Automatically deciding which product attributes are most important for each product type can be done with two methods. The first is analyzing large batches of product titles to see what attributes are standard practice to include in the title (as well as where they normally appear in the title pattern). The second method is examine the SQR report for attributes, correlate each attribute's SQR performance, then look for the most valuable attributes that each product has that are currently not being included in the relevant titles. I've written prototypes for the analyses of both methods. By combining them together, we could automatically enhance product titles for better search performance without requiring a human to edit titles one by one.