The only innovation recently seems to be application-specific. "Better interface for search results for x" where x is flights, consumer products, images, news stories, etc. These ought to be taken to another level of abstraction somehow (like a unified theory of physics).

Non-textual search: images, videos, sounds, and not just metadata/tags but actual data in them.  Youtube will let you click on the face of a person in a video and find other videos with that person in them.  Or other recordings of that particular Steinway or singer. And hopefully much cooler stuff I haven't thought of.

Accuracy: when I get a fact I should be able to quickly get a better citation than "Wikipedia." Wikipedia is getting worse and worse in terms of accuracy, relevance, organization, even grammar. Google is starting to do this with obvious things like U.S. Census Data.  Wolfram Alpha is designed almost exclusively to do this really well, but only for small facts (AFAICT).

Timeseries: Locational data and search is getting more widespread, but temporal data is lagging far behind. Archive.org and the federal government have a lot of the data that ought to be indexed better.

Mobile: I'm sure we'll see some cool products in mobile search in the next few years, though beyond a better/faster interface for iPhone Yelp, I can't think of what it'll be. But if nothing cool happens, some VC investors are going to be really disappointed, because they've funded a lot.

I think "natural language search" is becoming like artificial intelligence -- over-promised until people realized that it either wasn't gonna happen or wouldn't actually be useful, or both. But I'd still like to see it.

Tom, great ideas.  I hit a weird mental block when I posted this, but your comment really got my brain working how it should.

Of the things you mentioned, I think the one that is most needed (and most possible right now) is a really good image search.  Even if it only searched for text that appears in images, it would be great.  I know that there are already basic versions of that, but it can get way better.

Related to your idea about abstracting this another level, it would be really useful if a tool could aggregate the results from a number of other search engines.  For example, if I search for "Rams Football" I would want to see the rams game I'm going to next fall in google calendar, the chats I have with my brother about the rams, posts from my favorite rams, blog, the voicemail a rams sales rep left on my phone, etc.  This would truely be a universal search bar.

I could also see more localized search.  Maybe you're at the grocery store and you use your phone to find where the Totinos pizzas are.

In the computer vision community there is a lot of work being done on image search.  The type I think will be the most practical is where search is started off with a text search, but then the search is refined using image analysis.

For instance you search "mountain" and it goes and finds images from pages that have the word mountain near the image.  Then it displays a sampling of images like this, as google image search would.  Then you get the option to refine the search to get more images similar to a certain image in that set.  This is done through analysis of properties of the images themselves, such as maybe its a close up of a mountain or a far shot.  Snowy or in summer, etc.

I don't think pure image based image search could ever be practical because if you already had an image to start it out with, it wouldn't make much sense to do the search to begin with.

Frodo, when i'm mentioned image search, I didn't really mean you'd start with an image.  I meant that you'd type "mountain" and it would show you results that look like a mountain rather than images that have title attributes with the word mountain in them.

The idea of refining searches makes a lot of sense.  Google is already doing that with colors.  I hadn't thought of doing the same thing with distances or other parameters like that

I talked to a friend at Google about this post and she made me realize how bad (and difficult) local search is (Spotlight on Mac and whatever Windows calls their search). One obstacle is the small amount of data and links (algorithms like PageRank aren't possible, and very little is statistically significant).

If you accept as inevitable that eventually all our data will be in the cloud, this could get a lot better.  (Similar to your idea of aggregating personal search results.)

Tom, I want to make sure I understand what you're saying.  Basically, local search would be improved by using data from all over the cloud, right?

That makes a lot of sense, although I'm still not sure PageRank would work.  Other than websites, not many things link to one another so it would be tough to build an index based on something like that.

I'm reading the book "The Search" which is about Google.  One of the things that it says all the search engines before Google did wrong is they tried to determine the usefulness of a site based on content alone (rather than links to that site).  Maybe the answer to local search is to go back to what Alta Vista was trying, but do a better job of it.