Inferring intent on mobile devices

[Google CEO Eric] Schmidt said that while the Google Instant predictive search technology helps shave an average of 2 seconds off users’ queries, the next step is “autonomous search.” This means Google will conduct searches for users without them having to manually conduct searches. As an example, Schmidt said he could be walking down the streets of San Francisco and receive information about the places around him on his mobile phone without having to click any buttons. “Think of it as a serendipity engine,” Schmidt said. “Think of it as a new way of thinking about traditional text search where you don’t even have to type.”  – eWeek

When users type phrases into Google, they are searching, but also expressing intent. To create the “serendipity engine” that Eric Schmidt envisions would require a system that infers users’ intentions.

Here are some of the input signals a mobile device could use to infer intent.

Context

Location: It is helpful to break location down into layers, from the most concrete to the most abstract:

1) lat / long – raw GPS coordinates

2) venue – mapping of lat / long coordinates to a venue.

3) venue relationship to user – is the user at home, at a friend’s house, at work, in her home city etc.

4) user movement – locations the user has visited recently.

5) inferred user activity – if the user is at work during a weekday, she is more likely in the midst of work. If she is walking around a shopping district on a Sunday away from her home city, she is more likely to want to buy something. If she is outside, close to home, and going to multiple locations, she is more likely to be running erands.

Weather: during inclement weather user is less likely to want to move far and more likely to prefer indoor activities.

Time of day & date: around mealtimes the user is more likely to be considering what to eat. On weekends the user is more likely to be doing non-work activities. Outside at night, the user is more likely to be looking for bar/club/movie etc.  Time of days also lets you know what venues are open & closed.

News events near the user: they are at the pro sporting event, an accident happened nearby, etc.

Things around the user: knowing not just venues, but activities (soccer game), inventories (Madden 2011 is in stock at BestBuy across the street), events (concert you might like is nearby), etc.

These are just a few of the contextual signals that could be included as input signals.

Taste

The more you know about users’ tastes, the better you can infer their intent. It is silly to suggest a great Sushi restaurant to someone who dislikes Sushi. At Hunch we model taste with a giant matrix. One axis is every known user (the system is agnostic about which ID system – it could be Facebook, Twitter, a mobile device, etc), the other axis is things, defined very broadly: product, person, place, activity, tag etc.  In the cells of the matrix are either the known or predicted affinity between the person and thing.  (Hunch’s matrix currently has about 500M people, 700M items, and 50B known affinity points).

Past expressed intent

– App actions:  e.g. user just opened Yelp, so is probably looking for a place to go.

– Past search actions: user’s recent (desktop & mobile) web searches could be indications of later intent.

– Past “saved for later” actions:  user explicitly saved something for later e.g. using Foursquare’s “to do” functionality.

Behavior of other people

– Friends:  The fact that a user’s friends are all gathered nearby might make her want to join them.

– Tastemates: That someone with similar tastes just performed some actions suggests the user is more likely to want to perform the same actions.

– Crowds: The user might prefer to go toward or avoid crowds, depending on mood and taste.

How should an algorithm weight all these signals? It is difficult to imagine this being done effectively anyway except empirically through a feedback loop. So the system suggests some intent, the user gives feedback, and then the system learns by adjusting signal weightings and gets smarter.  With a machine learning system like this it is usually impossible to get to 100% accuracy, so the system would need a “fault tolerant” UI.  For example, pushing suggestions through modal dialogs could get very annoying without 100% accuracy, whereas making suggestions when the user opens an application or through subtle push alerts could be non-annoying and useful.

Facebook is about to try to dominate display ads the way Google dominates text ads

It is customary to divide online advertising into two categories: direct response and brand advertising. I prefer instead to divide it according to the mindset of users: whether or not they are actively looking to purchase something (i.e. they have purchasing intent).*

When users are actively looking to purchase something, they typically go to search engines or e-commerce sites. Through advertising or direct sales, these sites harvest intent. Google and Amazon are the biggest financial beneficiaries of intent harvesting.

When the user is not actively looking to buy something, the goal of an online ad is to generate intent. The intent generation market is still fairly fragmented and will grow rapidly over the next few years as brand advertising increasingly moves online. P&G – which alone spends almost $4B/year on brand advertising – needs to convince the next generation of consumers that Crest is better than Colgate. This is why Google paid such a premium for Doubleclick, Yahoo for Right Media, and Microsoft for aQuantive (MS’s biggest acquisition ever).

In 2003, Google introduced AdSense, a program to syndicate their intent harvesting text ads beyond Google’s main property Google.com.  The playbook they followed was: use their popular website to build a critical mass of advertisers; then use that critical mass to run an off-property network that offers the highest payouts to publishers. AdSense became so dominant that competitors like Yahoo quit the syndicated ad business altogether. Today, Google has such a powerful position that they don’t disclose percentage revenue splits to publishers and extract the vast majority of the profits.

It is widely believed that Facebook will soon follow the AdSense playbook by introducing an off-property ad network. They’ll try to use their strong base of advertisers to dominate intent generating ads the way AdSense dominated intent harvesting ads.

But to win the intent generation ad battle, data is as important as a critical mass of advertisers. For intent harvesting, users simply type what they are looking for into a search box. For intent generating ads, you need to use data to make inferences about what might influence the user.

This is what the introduction of the Facebook Like button is all about.  Intent generating ads – which mostly means displays ads – have notoriously low click through rates (well below 1%). Attempts to improve these numbers through demographics have basically failed. Many startups are having success using social data to target ads today. But the holy grail for targeting intent generating ads is taste data – which basically means what the user likes. Knowing, for example, that a user liked Avatar is an incredibly useful datapoint for targeting an Avatar 2 ad.

Publishers who adopt Facebook’s Like feature may get more traffic and perhaps a better user experience as a result.  But they should hope the intent generation ad market doesn’t end up like the intent harvesting ad market – with one dominant player commanding the lion’s share of the profits.

* Most text ads are about intent harvesting and most display ads are about intent generation, but they are not coreferential distinctions. For example, with techniques like “search retargeting” (you do a Google search for washing machines and the later on another site see a display ad for washing machines), sometimes intent harvesting is delivered through display ads.

Stickiness is bad for business

It is common to hear entrepreneurs and investors talk about the high level of engagement (what we used to call “stickiness”) of their website.  They quite rightly believe that it’s better to have a more engaging user experience, as that generally means happy users. Unfortunately, the dominant advertising model on the web – Cost per Click (CPC) – rewards un-sticky websites.  As Randall Lucas said in response to one of my earlier posts:

The paradox, it seems is this: in a pay-per-click driven world, site visitors who want to stay on your site — due to it having the once-much-lauded quality of “stickiness” — are worth much less than those who want to flee your site because it’s clearly not valuable, and hence will click through to somewhere else.

Facebook recently became the most visited site on the web. Yet their revenues are rumored to around $1B – about 1/30 of what Google’s revenues will be this year. Google has the perfect revenue-generating combination:  people come to the site often, leave quickly, and often have purchasing intent. Facebook has tons of visitors but they generally come to socialize, not to buy things, and they rarely click on ads that take them to other sites. Facebook is like a Starbucks where everyone hangs out for hours but almost never buys anything.

The revenue gap between sites like Facebook and Google should narrow over time.  Cost-per-click search ads are extremely good at harvesting intent, but bad at generating intent.  The vast majority of money spent on intent-generating advertising — brand advertising — still happens offline. Eventually this money will have to go where people spend time, which is increasingly online, at sites like Facebook. Somehow Coke, Tide, Nike, Budweiser etc. will have to convince the next generation to buy their mostly commodity products. Expect the online Starbucks of the future to have a lot more – and more effective – ads.

A massive misallocation of online advertising dollars

In an earlier blog post, I talked about how sites that generate purchasing intent (mainly “content” sites) are being under-allocated advertising dollars versus sites that harvest purchasing intent (search engines, coupon sites, comparison shopping sites, etc).  As a result, most content sites are left haggling over CPM-based brand advertising instead of sponsored links for the bulk of their revenue.

But there is an additional problem:  even among sites that monetize via sponsored links there is a large overallocation of advertising spending on links that are near the “end of the purchasing process” (or “end of the funnel”). For example, an average camera buyer takes 30 days and clicks on approximately 3 sponsored links from the beginning of researching cameras to actually purchasing one.   Yet in most cases only the last click gets credit, by which I mean:  1) if it’s an affiliate (CPA) deal, it is literally usually the case that only the last affiliate (the site that drops the last cookie) gets paid, 2) if it’s a CPC or CPM deal, most advertisers don’t properly track the users across multiple site visits so simply attribute conversion to the most recent click, causing them to over-allocate to end-of-funnel links 3) if it’s a non-sponsored link (like Google natural search links) the advertiser might over-credit SEO when in fact the natural search click was just the final navigational step in a long process that involved sponsored links along the way.

What this means is there are two huge misallocations of advertising dollars online: the first from intent generators to intent harvesters; the second from intent harvesters that are at the beginning or middle of the purchasing process to those at the end of the purchasing process.  This is not just a problem for internet advertisers and businesses – it affects all internet users.  Where advertising dollars flow, money gets invested. It is well known that content sites are suffering, many are even on their way to dying. Additionally, product/service sites that started off focusing on research are forced to move more and more toward end-of-funnel activities.  Take a look at how sites like TripAdvisor and CNET have devoted increasing real estate to the final purchasing click instead of research.  For the most part, you don’t get paid for the actual research since it’s too high in the funnel.

As with all large problems, this misallocation of advertising dollars also presents a number of opportunities.  One opportunity is for advertisers to correctly attribute their spending by tracking users through the entire purchasing process (in the case of cameras, the full 30 days and multiple sponsored clicks).  Very likely, these sites are currently overpaying end-of-funnel sites (e.g. coupon sites) and underpaying top-of-funnel sites (e.g. research sites). There is also an opportunity for companies that provide technology to help track this better. Finally, if over time advertising dollars do indeed shift to being correctly allocated, this will allow research sites to be pure research sites, content sites to be pure content sites, etc instead of everyone trying to clutter their sites with repetitive, “last click” functionality.


Why content sites are getting ripped off

A commenter on my blog the other day (Tim Ogilvie) mentioned a distinction that I found really interesting between intent generation and intent harvesting.  This distinction is critical for understanding how internet advertising works and why it is broken.  It also helps explain why sites like the newspapers, blogs, and social networks are getting unfairly low advertising revenues.

Today’s link economy is built around purchasing intent harvesting.  (Worse still, it’s all based on last click intent harvesting- but that is for another blog post).  Most of this happens on search engines or through affiliate programs.  Almost no one decides which products to buy based on Google searches or affiliate referrers.  They decide based on content sites – Gizmodo, New York Times, Twitter, etc.  Those sites generate intent, which is the most important part of creating purchasing intent, which is directly correlated to high advertising revenues.

But content sites have no way to track their role in generating purchasing intent.  Often intent generation doesn’t involve a single trackable click.  Even if there were some direct way to measure intent generation, doing so would be seen by many today as a blurring of the the advertising/editorial line.  So content sites are left only with impression-based display ads, haggling over CPMs without a meaningful measurement of their impact on generating purchasing intent.

All of this has caused a massive shift in revenues from the top to the bottom of the purchasing funnel – from intent generators to intent harvesters.  Somehow this needs to get fixed.