Deliberate Ambiguity

Death of the Search Startup

Death of the Search Startup

View more presentations from philosophygeek.

(note: this is a post based on a talk of the same name at Ignite SMX West)

Search is hard.  Really, hard.  As a veteran of three search engines, some might call me jaded.  However, this post isn’t just an attempt at apologetics.  I’ve learned something from those three experiences that I’d like to share with you.

If you’re looking to start a profitable business, it’s reasonable to wonder: can a startup ever pop up and beat Google?   As entrepreneurs, we always think that there’s a way to innovate our way out of a problem.  I’m going to argue that no amount of innovation can feasibly compete with Google or Bing.  Though marketing concerns like switching costs are certainly a problem for a search startup, I’m going to focus on an even deeper problem: the mighty greenback.  When it comes down to it, building a search engine is an incredibly expensive proposal.

In order to make some calculations, I created an equation that has two major components: hardware and people. I’ll give some explanations for each of the components and, in the process, show you how complicated and expensive a search engine is to build.

 

• Storage/Crawling (SC) – Two years ago, Google estimated the Web to be over 1 trillion documents and it’s growing every day.  A search engine needs to crawl and store the entire Web and keep frequently changing pages up-to-date.  Plus, your index is bigger than just Web pages.  You’ll be storing all sorts of metadata for the page: anchor text, outbound links, and any kind of interesting data you’ve extracted or created.
• Relevance (R) – At bare minimum, your search engine will have to have results as good as Bing or Google’s.  To achieve this, you’ll need servers to run relevance experiments, servers to store vast amounts of click-data, independent judges, and a big team to deal with all of the irrelevant/spammy sites out there.  And don’t think you’re done if you’ve just created 10 blue links.  Users have come to expect lots of other services: image search, stock quotes, weather answers, and news search, just to name a few.  You’re either going to have to license content or build verticals of your own. Both are expensive propositions.
• Runtime (RT) – When you do a search on Bing, you get a list of just 10 relevant results in fractions of a second from a set of possibly billions of Web pages.  How is that possible?  The short answer is: lots, and lots, and lots of computers to calculate your results.  Search engines typically use some kind of divide-and-conquer methodology, which means that whenever you issue a search, hundreds, or possibly thousands, of computers are involved in bringing you back results.
• People (P) – It’s unlikely, if not impossible, that you’re going to build a search engine with your buddy from college in a garage.  Worse, you need really expensive employees: PhDs in computer science, machine learning experts, statisticians, search engine veterans, infrastructure geniuses, etc. I estimate that it takes a bare minimum of 250 people to build a search engine. Yikes!
• Time (t) – Even with the smartest people, it’s going to take time to get everything running properly.  I estimate at least two years to get all of the components working together.
• Johnson Coefficient (Ĵ) – No equation is complete without an eponymous component.  However, I didn’t just stick in Ĵ because I love my last name.  The Johnson coefficient is a real concern for innovative search engines.  It’s all well-and-good if you’ve created a search engine that is equivalent to Bing or Google, but usually you want to make something better.  Unfortunately, better usually translates into more data and/or more processing . . . which translates into more machines.  Thus, the Johnson coefficient is the “tax” that innovative companies will pay for their innovation.

When you add it all up, my guesstimate for the bare minimum cost of a search engine is $100M dollars and could be as much as (or even more than?) $300M dollars.  Numbers like that, for a business that still hasn’t figured out how to switch people away from Google and Bing will terrify investors.

I have two case studies: Powerset and SearchMe.

Powerset can’t be seen as a complete failure, since Powerset was acquired by Microsoft.  At Powerset, all of our linguistic processing caused our index to be at least 10-20x bigger than a standard search engine.  Plus, the linguistic processing and matching we did at the runtime was much more expensive than typical search engine retrieval.  With about $25M injected into the company and 60 people, we were only able to index 2.5M documents of Wikipedia.  To have indexed the whole Web would have been much, much more expensive.

SearchMe was a search engine whose gimmick was to show thumbnails of Web pages in a fancy cover flow interface.  Imagine the cost of processing, storing, and serving all of those images.  It was no surprise to me when SearchMe shut its doors, since it claimed it needed another $50M in capital (!!!) to survive.

However, I encourage you to Keep Hope Alive!  Though I don’t think that another general purpose search engine could compete with Google, there’s a lot of opportunity on the periphery for special purpose search engines and vertical search engines.  And heck, if you’re lucky like Powerset, you might get bought by Bing or Google!

The moral of the story is that, next time you hear someone proclaiming how they are going to be the next Bing or Google, smile and nod, and keep your hard earned money far away from an investment in that ill-fated venutre.

Recap of "Don't Call it a Comeback: Semantic Technology and Search" at SES SJ 2009

(c) Ken Yeung

Today I was at SES San Jose and represented Bing on a panel called "Don't Call it a Comeback: Semantic Technology and Search” hosted by Dana Todd.  The format was interesting: the four major search engines acted as the panel and three startups – Twine, WolframAlpha, and HubPages – gave presentations.

SES is very pragmatic conference focused on publisher issues, whereas semantic search discussions typically are forward-looking or, perhaps less generously, speculative.  Luckily for our audience, our conversation was peppered with practical examples.

Nova Spivack of Twine started off by arguing that the Web is now too big for keyword search. He reasoned both that users need more tools to organize the millions of results they get back for a query and that they want to ask more complex questions of that data. He suggested that the next steps in this third decade of the Web (since Tim Berners-Lee wrote his seminal paper in 1989) will be natural language search, semantic search (defined by Nova as “parameterized search”), which will eventually give rise (in Web 4.0) to applications that are able to make inferences and reason.

The most exciting portion of his presentation was his announcement of Twine 2.0: semantic search at Web scale.  He humbly admitted that Twine is a small company and that Twine 2.0 is an incredibly ambitious project (indeed!).  In a thinly veiled plea to the search engines on the panel, he suggested that he'd like to partner with one of us.  I hope to get a demo soon!

Next in line was Lane Soelberg from CoopVentures, who is helping to bring WolframAlpha to market (I wrote a post about WolframAlpha marketing prior to their launch).  Instead of just rehashing what WA does, he gave some compelling examples of how publishers could create mashups using the forthcoming WolframAlpha API and their own data.  One example was a hypothetical insurance company that combined their own internal accident database with WA weather, traffic, and holiday data to create an Accident Predictor.

Finally, Jason Menayan from HubPages gave a very practical example of how semantic technology can be used in SEO.  HubPages is a UGC site that generates over 60MM page views a month.  HubPages tried two experiments.  First, they semantically generated content suggestions for a few pages using Digger.  Though only done with 3 pages that hadn’t shown up in Google results yet, the results were positive.  The content suggestions broadened the topics on the page and caused them to be ranked highly within a few weeks.  HubPages is in the process of experimenting with more pages to get a more statistically significant set.  The second test was to present both semantic tags from Relevad and regularly generated tags as suggestions for the author of the page.  The idea of the semantic keywords is sometimes they surface relationships that are less likely to be found in a typical keyword style. Both sets of keywords seemed to help traffic equally as well up to a point -- the curve started to fall around 35 tags / per page. Not a clear win for semantics, but keep in mind that "folks" were still choosing the tags for the folksonomy.  Maybe automatic (semantic?) tag generation would be more efficient? .

Given the fascinating presentations, there wasn’t a lot of time for the folks from the search engines to comment.  We did get a few chances to respond to questions from Dana and the audience.

Oddly enough, the representatives from Google, Yahoo, and Bing (me!) were all people focused on creating the next generation of rich captions and had great statistics to share.  According to Kevin Haas from Yahoo, “lots and lots” of searches contain SearchMonkey results (try this).  After a bit of prying, he conceded that the number was over 50%, but his demeanor suggested that this number was much higher.  Personally, I’d be more interested in the number of results that have a SearchMonkey result in the top 10 or top 3 results, as a better measure. 

The first question from the audience came from a concerned audience member who worried that search engines could too much control of  content once the content was marked up.  Othar Hansson from Google assured the publisher that, after testing Rich Snippets for months, enhanced captions actually promoted click to the vendor sites.  He also stated that Google focuses on informing the user click and that they don’t believe many user queries can be answered on the SERP itself.  I humbly disagree, but will leave the judgment ultimately up to the users.

Another question to the search engines was: what are you doing in semantic search today?  In typical search engine style, we were all a bit cagey with our responses, partially because it’s hard to succinctly describe the technology that goes into any given search.  My response went something like this: given that Bing is a decision engine, it’s critical that we approach semantics from both the query side and the document side.  On the query side, a user should be able to type in a query in a way that’s most natural to her vocabulary, not the specialized vocabulary of a search engine.  Therefore, it’s incumbent upon us to translate the normal language of the user into the language of a search engine (if necessary).  However, the document side is just as important. The more a search engine knows about a document, the better we can rank the document in search results, the better snippet we can generate, and the more information we can present on the SERP.  Bing is certainly working to make strides on both fronts. Usually these advances are opaque to you, the end-user, and so much the better!  We’re all about solving user problems, not about touting our technical prowess.

I’d love to hear comments from readers, both on my impressions of the content of this panel and what you’d like to hear about on panels in the future.  No doubt there will be more semantic  search discussions at conferences in the future: what sorts of issues would you like addressed and what kind of burning questions do you have? Some questions that never got answered (courtesy of Dana Todd):

• Why has semantics struggled so hard for an identity, and what’s necessary to help it get market traction and mass appeal?
• What are the factors that limit/enable success for a semantic project or technology?
• We typically think of semantics in terms of words; WolframAlpha is now attempting to redefine to include numbers and pure data. How does that change things?

Bing: An Insider's Perspective

Today, Steve Ballmer announced at AllThingsD that Live Search will relaunching next week as Bing.  Given the amount of speculation on Twitter and that I'll be out next week on vacation when Bing actually launches, I thought I'd type out a few quick thoughts now.  Keep in mind that this is my own perspective and doesn't represent the view of my employer.  In case you don't know, I came to Microsoft Live Search through the acquisition of Powerset last year and work here as a program manager focused on Powerset integration and development.   

First a few comments on functionality and what I like in Bing.

I use Bing (formerly known internally as "Kumo") every day for all of my searches.  When I first joined Live, I used Live.com because I wanted to gain an intimate knowledge of my new product.  I freely admit that I frequently returned to Yahoo or Google when I couldn't find what I wanted.  Over the past six months, Bing relevance has improved noticably and dramatically. Now, I rarely use any other search engine.  When I have trouble finding something on Bing, my searches on competitors are usually just as sparse.

When I do use a competitor, I tend to miss some of the features on Bing (which, unfortunately, you can only experience through screenshots on other blogs until next week).  We've done a killer job designing the left-hand nav (affectionately known as the "ToC" for Table of Contents).  I love the organization of results (I wish it showed up on more queries!).  I also find the Related Searches to be super-useful and, being a frequent searcher (and re-searcher), Search History is literally my bread crumb trail. Another amazing feature is the Best Match, which tries to determine when I'm entering a navigational query (Walmart, Facebook, etc.) and gives me deep links, a search, sometimes a phone number, etc.  There's lots' more in there and I'm sure you'll get to see next week.  I'm excited to see people's feedback on the features that they find the most interesting.

Also, remember that there's a lot of complicated technology under the hood.  It's not just about the interface, it's also about delivering great results.

Which leads me into a few comments for those writing blog posts about Bing and who want to dive into the technology.

Please, please consider writing two blog posts: your initial impressions and your thoughts after you've used Bing as your primary search engine for a minimum of a week.  As per my admonishment in How *not* to rate a search engine, you can't test out a search engine with just a few queries, you have to use it for some period of time to form an informed opinion.  Give us a chance.

Building a search engine is super hard.  We did it from the ground up at Kosmix and the scientists and engineers here at Bing have been working on it for several years.  Awhile back, a friend told me that Amazon estimated it would cost $50-$70M to build a search engine; I think that number is way too low, both in terms of Capex and human capital.  My point here is that it's hard (impossible?) for a startup to challenge Google (cf. Cuil) and even companies with deep pockets are scared of such a huge investment.  Microsoft is committed to Bing and has an incredible architecture for advancing search.

Which brings me to an important point: give the underdog a chance.  I think folks here at Microsoft know that we're coming late to the game, that we're coming from behind, and that we've got some solid competitoin out there.  But competition breeds innovation and innovation is good for consumers.  As Steve Ballmer said today: search could benefit from a good ol' fashioned feature war!

Oh, and for you Powerset fans, there's certainly some Powerset in Bing, but there's a lot more Bing in Bing.  We at Powerset proud to be a part of the Bing team, but most of the credit goes to the incredible team assembled in Redmond.  Watch Powerset's blog next week for more information about some of Powerset's specific contributions.

I'm working for search engine number 4 exactly because I think that search is in its infancy and there's a long way to go.  Try out Bing, send us your feedback and we'll listen.  To me, the launch of Bing as a new Decision Engine today is a huge leap forward, but I'm even more excited about the next 6 months, 12 months, and 2 years.  Hope you hop on board for the journey!

Happy Birthday Powerset.com!

It's hard for me to believe it, but a year ago today, we launched Powerset.com.  Of all the products I've launching in my career, I'm most proud to be a part of the Powerset team. Happy Birthday, Powerset.com!

So, how did it go? 

The launch itself was a resounding success.  Powerlabs and the amazing community (thanks, Powerlabbers!) helped to hone the final product into something highly usable, despite the underlying complex technology. 

Though we mistakenly flipped the switch and went live a few hours earlier than we had planned, the huge engineering effort to ensure launch capacity was worth it: Powerset went down only once for <30 minutes when we got slashdotted.

Press coverage was awesome.  The 30+ press briefings we did prior to the launch helped to calm some of the hype around Powerset ("No, we're not a Google-killer, yes, we have some cool technology, yes, this is a long journey," etc.).  Some of the hype is unavoidable, but I think people realized that we had the seeds of something really cool.

When we started getting real-world queries, we found that users are ready, willing, and able to type in natural language queries.  However, we also discovered that it's really hard to market a Wikipedia search engine to people (lots of people wanted to search for Pizza near their house).  As we predicted, once people had a taste of natural language search, they wanted much, much more.

Then, on July 1, 2008, it was announced that Microsoft had decided to purchase Powerset.

This was an amazing start, but the journey has just begun.  I wrote a post a year ago musing about whether Powerset crossed The Uncanny Valley.  I think it's safe to say that neither Powerset nor anyone after us has been able to create a semantic search experience that meets users' expectations. However, now that we're part of Microsoft, I'm really excited that Powerset will help to transform the Live Search experience.  We've got a lot of things in the pipeline, so keep checking back for more announcements.

Wolfram Alpha: A Marketer's 2 Cents

Wolfram Alpha will be launching on May 18, 2009, almost exactly a year after Powerset launched.  Unlike other big flops launches, I am totally stoked about the Wolfram Alpha launch, especially after reading all of great posts written about pre-demos.  In the spirit of nostalgia for Powerset's launch, I thought I'd toss my hat in a ring (unfortunately never having had a live demo), focusing on Wolfram Alpha's pre-launch from a marketing perspective.

Branding. In launching a new search service, a comparison to Google is inevitable.  We explicitly told reporters that Kosmix and Powerset were not competing with Google, but they cannont resist the siren song of a controversial headline.  Wolfram Alpha is wisely branding themselves as a computation engine, suggesting questions that crunch numbers instead of better search.  Also, despite comparisons to Google, Wolfram Alpha is trying to be humble and remind people that the road to computation is a long journey, not a quick win.  Wolfram's post The Quest for Computable Knowledge: A Longer View sets the right tone in a historical context.

Audience. There have been a lot of complaints that the name "Wolfram Alpha" is too long and complicated.  In demos, they've been showing off chemical formulas, population graphs, and calculus. This will not appeal to my mom.  Instead of going after consumers, they are going after geeks who love data.  Most geeks associate Worlfram with Mathematica (and the ueber-geeks associate him with NKS).  I like the name, for now.

Competitors. Notwithstanding comparisons to Google, I'm surprised that few have compared Wolfram Alpha to True Knowledge, which is currently live and can answer inferential questions like who are queen elizabeth ii's grandchildren and who was the first woman to climb mount everest and what time is it in Boston.  Note that both of these questions require inference, but TK knows about QEII's kids and their kids (and that kid's kids are grandchildren).  Unlike Wolfram Alpha, TK has a system for users to enter and curate information in their database.  If TrueKnowledge can get enough editors, this is potentially a much more flexible system than manually curating government data.  Another obvious competitor is Metaweb's Freebase. It'll be interesting to see who else pops up in this space.

Benefits.  This is where Wolfram Alpha is going to have some problems.  It's just not very often that I have questions about chemical formulas and when I type in "Seattle, WA" the graph of the temperature might be interesting, but probably not what I'm looking for.  What is it that I should use a "computation engine" for?  At least in the short term, they're going to rely on geeks getting lost in all of the cool things that Wolfram Alpha makes possible. 

Scope.  As a corrolary to the above, Wolfram Alpha better do a phenomenal job on day one of explaining what data is availabe and what kind of computations can be done on that data.  If they're smart, they'll have tons of example queries that show off the power of the engine, lots of screencast and tutorials, and maybe even a contest to get users to generate cool queries.  Based on what I've seen, here are a few demos that I'm gonna try (of course I'm going to take my own advice and not rate Wolfram Alpha based on the results):

• Chemical formula of water
• Number of moles in a gallon of water
• What happens when you mix bleach and ammonia?
• What states are in the Pacific Daylight time zone?
• What's the time difference between Tokyo and New York City?
• What animals have a cloven hoof?
• Is it colder in San Francisco or Seattle in May?
• When did the wooly mammoth live?
• What were the top 10 largest cities in the US in 1990?

*waiting with bated breath*

How *not* to rate a search engine

While demoing Live Search at the Web 2.0 Expo, people continually asked the same questions: “What makes Live different?” or “Show me some features that will make me want to switch from my search engine” or the extremely confrontational “Why do you think you’re better than Google?”

My first instinct was to dive in and show people the coolest features in Live Search (e.g., demoing  Virtual Earth with an Xbox controller) or to let them play around with their own queries.

However, given my experience working for several startup search engines, I’ve come to realize that it's extremely difficult to convince someone that you’re better than another engine with words, features, or few carefully chosen queries.

So, after awhile, I started my demos with a caveat about the nature of a search engine: I implored my audience to try out Live Search for a week so that, in the words of the immortal Lavar Burton of Reading Rainbow, “But, you don’t have to take my word for it.”

Is this a cop-out?  Why is demoing search so hard?

Search “Features”

When showing off a new version of Microsoft Word or Typepad or Yahoo Messenger, a good product marketing person will not just demonstrate features, but analyze their audience and demonstrate benefits that help users accomplish specific tasks. (This is just product marketing basics.)

A search engine, by contrast, has an extremely simple interface: you type in some words and hope that the engine will cough up pointers to helpful Web sites or give you a direct answer.  The inner workings of a search engine, i.e. how those results were produced, are completely opaque to the user.  Hundreds of features are used to rank results so that the right Web sites and answers show up on a page when you type in some string of words.  Those features don't surfaces as demonstrable chunks that can be easily summarized or understood.

Common mistakes when evaluating a search solution

 Which brings me to the biggest mistake people make: judging a search engine by typing in a few queries and analyzing the results. There are many interrelated reasons that this methodology fails:

• A few good/bad results don’t mean that all results will be good/bad – even if you try out five searches and all are good, how do you know if your sixth is also going to be good?  That is, since you don’t know what is going on under the hood, you can’t make any predictions about the quality of future results.
• It’s hard to select a representative cross section of queries – people usually try out a few navigational queries, a vanity query, and a few queries that are either damn-near impossible or extremely obscure.  None of these sets represents an accurate cross-section of your monthly query log.
• What you think is “good” may not be good for the majority of users – for navigational queries (e.g. “CNN.com”) the top result is clear.  For more complicated queries, the top results are rarely obvious.
• Queries are out of context – we had this problem at SideStep all of the time.  During usability studies, users who were simply evaluating the look-and-feel of the product and scanning for cheap flights without any end-goal were never as good as users who were actually trying to buy a flight for a real trip.  A search engine should help you complete tasks, not just give you a pretty page or have links that look useful.
• People tend to focus on the first result – some queries just require one result.  But many queries should be judged by the diversity of interesting results.

There are probably countless other mistakes that are made during solo evaluations of search.   Therefore, search engines big and small realize that problems of ranking and relevance – the core of any search project – are solved only by lots and lots and lots of data from lots and lots and lots of people.  To solve this data problem, we need to collect data from real users.  For example, we run many thousands of queries past human judges and look at mountains of click data from the production site.  After applying apply advanced statistical techniques to this data, we get the information we need to create algorithms that turn your few (mispelled) words and turn them into a useful page of results.

As one of my colleagues at Powerset always likes to remind me: this is rocket science.

Overview of Semantic Search Engine Panel

As previously reported, I was lucky enough to host a panel at Alternative Search Engine Day last Monday.  I thought I’d type up some of my notes from the panel and talk about what I thought was most interesting.

For the demos, we got a sneak peek into TrueKnowledge, which is still behind a closed beta.  TK currently houses over 106M facts and is constantly growing.  TK does inference, which means that it can derive facts from knowledge it has cataloged and knowledge it has about the world.  For example, TK might not have an explicit fact for Queen Elizabeth II’s grand-daughters, but it knows her children and their children and can derive the fact.  Users can query the system in natural language and, when a fact isn’t returned, can add that knowledge to TK.

Truvert is a vertical search engine for the green market (a popular topic at Alt Search Engine Day) which his built on technology from OrcaTec.  When a user types in an ambiguous term like “palm” into Truevert, the results are centered around “palm oil,” not the PDA.  OrcaTec claims that it can build a vertical with a small training set of a few thousand documents in just a few hours on a single machine.  Their search is based on Yahoo BOSS.

I was especially excited to see TextDigger, which I’d never seen demoed live before.   TD is currently in private beta with about 1000 users and a long waiting list.  Launch is anticipated in Q3 of this year.  TextDigger is meant to be a high-end research tool where users have lot of power over the semantics of their search.  TextDigger shows all of the possible senses of words in a given query and the user can select which sense they really meant, which will filter the search results accordingly.  I found this especially interesting because it’s a feature we always considered at Powerset, but ended up leaving out of the final product because of concerns about confusion of users (and that we thought that the heavy lifting should be done by the engine, not the end user).  I’m really excited to get an invite to the beta and I’m sure I’ll have a more detailed analysis once I get in and play around.

During the Q&A session, a number of interesting topics came up.  Both TextDigger and Truevert both rely on results from a bigger search engine and rerank them based on semantic algorithms.  TrueKnowledge has built their own index of facts, which is focused on structured information as opposed to free text.

When talking about what “semantic search” is defined as, Dr. Musgrove of TextDigger delineated a difference between “soft” and “hard” semantic search (reminiscent of strong vs. weak AI and the hard vs. easy problem of consciousness).  Better query suggestions, similar to what Google recently announced as soft, whereas actually trying to understand queries and read text is hard.  The distinction still isn’t completely clear in my mind, though my instinct suggests that it would be useful to discover.  I’m going to noodle that and see if I can come up with something.

Overall, great panel and I look forward to the upcoming Semantic Demo Session on 4/14 at Microsoft in Mountain View, Web 3.0, and the Semantic Technology Conference.

Olympicifying Search Engine Homepages

Type in "medal count" into Google, Yahoo, and Live, and you'll get helpful instant answer that gives you a medal count.  But, of course, every search engine has decided to theme their logo, Beijing Olympics Style.

Google's is a traditional play on the Google logo, typical of what you'd see for St. Patrick's Day or Easter.  Artsy, but boring.

Yahoo's is traditional Yahoo, with an animated gif of little cartoon swimmer dudes.  Kinda goofy.

Microsoft's Live Search is my favorite (not just because I work here now), showing off gorgeous Olympic photos, which change on a daily basis.

>

See if you can find the "hot spots" on this.

Powerset crosses the Uncanny Valley

In case you haven't heard the news, Powerset launched today.  Since I can't sleep due to launch adrenaline, I suppose it's time to weigh in on Powerset unofficially.  It's fun to write a press release, but it's more fun to philosophize on my own time!

Last year, I came across the concept of the Uncanny Valley though an employee of PARC.  The basic concept runs something like this: as you make a robot more like a human being, people tend to respond with more empathy.  However, there's a dip in the empathy curve when the robot looks enough like a human, but has behavior incongruent with what we'd expect from a real human being.  We tend to forgive weird mannerisms (jerkiness, oddities in speech, etc.) when a robot looks like a robot.  When a robot looks like a human, we expect it to act like a human.

As an example, cartoon characters are usually more believable than computer generated characters, unless the CG is so perfect as to be indistinguishable from reality.

Keyword search engines sit somewhere on the left of the curve.  We're willing to accept the choppy "keywordese" language of a search engine because we appreciate the results that come back and we can make excuses when it goes awry: "It's just a computer.  It doesn't really understand me."  As with most technology products, the utility curve and the expectation curve are aligned exactly because we treat keyword search engines as an unfeeling technology.

Oddly enough, when a system is endowed with an understanding of meaning (call it "semantics"), we suddenly demand that the system act, think, infer, and comprehend just like another human.  Instead of just being a cold of technology, we begin to treat the system with agency and want it to act like other agents (i.e., humans). 

The question shouldn't (and can't) be about releasing a system that has a perfect understanding of language, because that perfect understanding is many years down the road.  How is it that companies like Radar, TrueKnowledge, and Powerset believe that we can create products based on imperfect technology?

There are two ways to stay out of the Uncanny Valley.  First, create unique features that are so compelling that users are willing to deal with imperfection.  Keyword search hasn't survived because we always get the right answer - plenty of searches yield nothing.  But, though we may get wrong answers, we often enough get answers that would never have been possible before keyword search.  Second, access to semantic features shouldn't require a significant change in user behavior.  No matter how good (or "natural") a feature is, if a user has trouble finding it, it won't get used.

People often think that NL search is doomed to failure both because computers don't understand enough and, even if they did, users won't start entering natural language queries.

I think Powerset has scored on both counts.  As an example, for topical queries like moses or machiavelli, Powerset assembles a summary of Factz based on our deep linguistic analysis of each sentence.  This kind of automatic aggregation of content just wasn't possible before Powerset.  But, users don't have to do anything special.  They just get Factz "for free" by using their keyword usage paradigm.

I'm really proud of Powerset's first product.  The debate on whether or not we've crossed the valley will continue, but one thing is certain: a better, more natural way of interacting with technology is coming. (Of course, whether we sit to the left or the right of the Uncanny Valley is the subject for another post.)

Bed time, finally!

The Ask Lesson: Don't Go Head-to-Head with Google

Ask.com Seeks Makeover as Woman's Site is the real title of an article and the real strategy of a public company.  Wow!  Talk about a shift in focus!

To me, this move indicates that Google, Yahoo, and Microsoft (GYM) have complete market domination.  The capex of building a search engine is too high, not to mention the marketing problem of getting users to change their internet front door.  If IAC-backed Ask couldn't compete with innovative features, comparable result quality, and a big marketing campaign, then cash-strapped startups aren't going to fare much better.

I'm certainly not suggesting that Google will reign forever, considering I'm now at my second search startup (Kosmix, then Powerset).  Rather, I now realize that going head-to-head with Google in search is as ill-fated going head-to-head with Microsoft in operating systems.

What are then the conditions for success?  I obviously don't have the answer, but here are some lessons I've gathered from my experience.

• All of the core components to build a search engine are readily available.  Only a radical shift in technology built on top of this that will drive a significantly different user experience.  Just creating a better SERP based on the same technologies won't be noticeable to user and is easy for GYM to copy.  But, who knows what will herald this Kuhnian shift?

• Users won't remember your vertical focus.  Consumers have trouble remembering Kayak for travel, Trulia for Real Estate, Indeed for Jobs, and dozens of other engines.  The default is to type a problem into GYM. Therefore, use search to drive traffic.  GYM owns front door traffic to the Web.  Don't fight, play along in the ecosystem.  If you can get some repeat users along the way, then good for you.

• Plan a series of interim products.  A Google index is probably hundreds, if not thousands of times as large as the biggest startup index.  Oh, and it's lightning fast.   And it never stops working.  Only after years will all of your components to congeal into a final product.  Better be prepared with a long runway and some good interim steps

 

I suspect most search startups will find their niche in something related to search, much like software companies began playing in the Microsoft ecosystem instead of building competing operating systems.  For example, Kosmix  is using its technology to build a homepage for every page on the Web.  There are definitely still holdouts like Hakia and Mahalo.  Hakia is headed for a battle against the big boys: their experience still isn't even equivalent to Google.  Maholo might be able to market its user-generated pages in Google, but then it's sounding like a Squidoo. 

Ask?  Sounds like they're quitting the search engine business and turn into a social network/content play.  Oh well.  Glad that things didn't work out when I interviewed there last year.