Marcia - 10:13 pm on Mar 16, 2008 (gmt 0)

Town is ambiguous, they were lucky to get it at all.

They got beaten out for town & country [google.com] by far when there had to be a choice made. Is that because it appears more together with magazine than car words, numerically in a co-occurrence matrix?

And then, there's the town car [google.com], which is Lincoln (Ford), also co-occurring with cars, vehicles, autos, automobiles, etc. Even the photos for image search are the Lincoln, not Chrysler, though in a search for Lincoln town car there aren't any site links - but then, is that specific enough not to need any assists to further narrow it down for users.

What's interesting is that even though it's Lincoln that's the right choice for relevance for "town car," in those search boxes on top with choices for model, make and location - I see chrysler entered as the default choice even though it isn't Chrysler, it's Lincoln. Is that because I have a Google_tracked history of looking at and for Chrysler vehicles in current/recent searches? Or are they just giving an alternative because the Lincoln search result is right under it?
_________________________________________

Afterthought:

Translated into Google patent-speak, there seems to be a predictive element involved with choice of terms for site links. When the secondary terms (that appear as anchor text in the site links) have co-occured enough times in a co-occurrence matrix together with the search term that triggered the listing with the site links in the search results, it can be statistically deduced that the search term used is predictive of those secondary search terms. In other words, users who did the search could be predicted as being likely to do searches on those secondary terms when refining and narrowing down their searches to be more specific.

For example, when we saw muscle cars as a secondary site link term for Dodge, even though we did a search for Chrysler trucks, couldn't that be because there's so much usage of Dodge and muscle cars together that it's established that they go together? And can't it be predicted that a high percentage of users searching for "dodge" can be predicted as likely to also search for muscle cars?

If they're feasible possibilities, there's no other way to do that statistically without using co-occurrence data because there's no human-like reasoning or logic in a straight mathematical algo.

[edited by: Marcia at 10:49 pm (utc) on Mar. 16, 2008]