Grokking Findory

I have been experimenting for the past couple of weeks with Findory, a web search company that provides implicit personalization of search results. Findory's algorithm orders search results based on three criteria beyond your explicit search terms: your previous search queries; the articles that you read in response to those previous queries; and the articles read by other people with similar search queries.

Unlike most search engines, Findory will present you with search results even if you don't enter a search term. In this sense, using Findory has more in common with using a newspaper than using Google. However, unlike a newspaper, Findory will gradually learn which articles you want on the front page.

Findory also has a feature that I really like: you can subscribe to a daily RSS feeds for keywords that interest you. However (unlike Technorati and PubSub which offer a similar feature) the RSS feed is generated based on Findory's model of your interests -- and reading articles from the RSS feed provides feedback to the personalization engine. That's a big win if you want a generalized tool that can find interesting articles for search terms that interest you.

But enough of an overview -- let's get to the meat.

The Good:

1. Findory turns up articles that I find interesting. Many of those articles (unsurprisingly) appear related to my previous search queries and the articles that I read. But not all. One article (that was completely unrelated to my search history) was entitled "Virginia to Shoot Gulls Nesting on Highway". Sort of a mashup of Google does BoingBoing.
2. You get lots of control over your search history -- if you perform a search and you don't want Findory to remember it, you can simply delete that search.
3. Findory provides substantially fewer search results for an average search term than, for example, Google or Yahoo. I'm going to assume that this is because the underlying algorithm allows them to throw away a lot of keyword results that aren't personal enough.

A few complaints:

Perhaps my most fundamental criticism of Findory revolves around its core strength: Findory is focused on implicit personalization. Yet I kept wanting my homepage to reflect my interests faster. The way to teach Findory what you like is to run searches. But I kept wanting to tell it more explicitly what I wanted.

Let me digress. Those of us who blog are creating ever-expanding streams of meta-data about ourselves, including the keywords and concepts that preoccupy us and the sites and articles that capture our attention. Why should a personalization engine wait around for the fairly paltry information that it can derive from 5 or 10 search queries over a two week period? Why not allow me to explicitly point a search engine to my meta data? (Especially when this meta data is already formatted into a well-understood xml format?) Better yet, why not say "I like gadgets. Personalize my searches based on Gizmodo." Couldn't a personalized search engine make some interesting use of my blogroll or my del.icio.us bookmarks? Why not allow me to "prime the pump" so to speak when I first arrive at Findory?

I also had a minor implementation complaint with Findory's RSS feeds. Suppose I see a summary of an interesting article in my personalized feed, and I click on the title. That click brings me directly to Findory where I see the same article summary that I just saw in my feed aggregator. Why does it bring me to Findory? I can't think of a reason (besides inflating Findory's pageview count). Furthermore, Findory does not consider this behavior to qualify as "reading" the article (and therefore presumably does not customize my future search results based on this click). Why not have the link point to Findory (so that Findory can record that I read the article), but then immediately redirect my request to the full blog article?

Finally, let me wrap up with a bit of speculation on Findory's future.

Let me argue, briefly, that personalization such as Findory's will not find a killer app in serving personalized news or search results. The killer app for this sort of personalization is in serving search engine advertisements. Remember, the current model of an AdWords-type advertising engine is that the advertisers that pay the most get a boost in their ad's ranking, and the ads that have the highest click-thru-rate (a form of collaborative filtering) get a boost in their ranking. But the underlying relevance model for which ads get displayed is based on keyword (and synonym) matching.

Instead, what if the search engine created a mental model of the concept you were searching for, based on your interests as exposed by your search history? Could the search engine serve more accurate ads?

Here's a brief experiment to try at home: Run a search on google for "tiger". When I did this the first ad was for worldwildlife.org: "Learn about big cats". The second ad was for tigerdirect.com: "Low Prices on PCs, LCDs and More!". Now run a second search for "Mac OS", and then a third search for "tiger". Notice that the ads displayed on this third search are still split between felines and computers. Google has missed an opportunity to learn something from my intervening search for "Mac OS", and therefore lost an opportunity to show me more relevant ads.

In fact, the ads that you click on could themselves provide additional feedback back to the personalization engine -- feedback that is at least as relevant as any article that you read.

Here are some practical reasons why one of the top four search engines would be more likely to buy Findory for its effect on advertising than its effect on search results:

1. The return on investment is straightforward to calculate: Success in advertising could be defined as a higher click-thru-rate. To estimate the ROI you need an estimate of the percentage of searches that employ commercially viable but ambiguous search terms; an estimate of the percentage of such queries that can be clarified by the personalization engine; and an estimate of the increase in click-thru-rates that might result. Compare that to estimating the ROI for an search results: Success in this case is defined by a user being incrementally more likely to return to your search engine and use it again in the future. (Not to mention that more accurate search results may actually decrease the number of ads that users click on, since the organic search results may address their needs without them turning to the paid advertisements.)
2. Integrating a personalization engine into an advertising engine would be trivial compared to integrating one into a general search engine. Instead of hundreds of weighting factors affecting relevancy you deal with a handful.
3. Measuring the quality of a particular set of search results is a difficult task -- one for which it is hard to agree upon common metrics. End users wouldn't necessarily notice the results of personalization in an average web search. Search engine marketers, on the other hand, live and die on incremental changes to click-thru-rates.

Overall, I will probably continue to use Findory's RSS service -- I just wonder if its focus on personalized search results will result in a viable business.