Why Facebook Advertising is Like Apple Farming

There are two types of front doors to get into the Internet, one is search and one is social.  And each entry point says something about why we came there in the first place – either to search or to discover.  This premise explains why search marketing on Google is so effective – because people come there with intent to buy.  But it also explains the role that social can play in marketing as a platform to help customers discover new products and create more demand.

apple1

Search marketing is like apple picking.  There are a finite number of apples in the orchard, and hiring more pickers doesn’t create any more apples.  If you manage search for Zappos, and your campaign for “riding boots” is doing great, you have a problem.  Nothing you do can get more people to search for riding boots.  Spending more money only makes it cost more to capture the same sales.  Similarly, spending more money to pick apples just costs more money for the same number of apples.  If you want more apples, you need to plant more trees.

In order to “plant more trees” in the marketing context, you must find a channel to create more  demand.  Google is not an ideal environment to create demand because people already know what they want when they go there.  Conversely, on Facebook discovering new things is native to the experience. Marketing on Facebook, when done correctly, should help aide discovery of new products and ideas.

Chart3Our data shows that planting trees on Facebook requires patience.  In some cases, we have found that 82% of purchases came more than 24 hours after the initial click, and 54% came in more than a week after the click.  This means that customers are seeing the product on Facebook, clicking through, and then considering that purchase for a week or more before making the decision to buy.  The return on ad spend (ROAS) for new user acquisition campaigns is often between 100% – 200% on the first day, but can increase to 500% and beyond after thirty days.  As such, aiding discovery should not be measured in isolation.

Recent studies show that by creating demand on Facebook, marketers can drive better results on Google.  Kenshoo, a search marketing company that is now moving into social, recently published a case study demonstrating that using Facebook ads boosted return on ad spend on Google by 30%.  It also increased average order value by 24% and yielded a 7% higher click-through rate.  This data is significant in that it is not just saying that planting more trees creates more apples, it is saying that planting more trees makes your existing apple pickers more efficient.

At SocialWire, our customers still measure success based only on the sales they can directly attribute to our marketing efforts using what is known as “last click attribution.”  While SocialWire consistently meets or exceeds our clients’ goals even on this last-click basis, we believe that marketers could more properly allocate marketing budgets by attributing more value to demand creation and less to demand fulfillment.  In other words, farmers that can attribute more value to planting trees can justify a larger investment in growing their farms.

The marketers that play the long game and invest in demand-creation to grow their business are the ones that are going to thrive on social platforms like Facebook, Twitter and Pinterest.  As a farmer, you cannot look at an empty field full of rich soil and say “that sucks, there’s no apples there.”  Similarly, marketers should not expect social platforms to deliver search-like results immediately.  Marketers should view these social platforms as fertile ground for creating demand that will drive more volume and better results in their fulfillment channels like search marketing.

Winning!

MedalGreat news to share today, the product we just released out of beta – Dynamic Product Ads – was announced today by Facebook as a winner of their PMD Innovation Competition.  We are thrilled about this recognition.  Our true validation has come from seeing some of the world’s largest e-commerce companies signing up to run evergreen campaigns with SocialWire, but the recognition from Facebook tells us that we share a vision for making social advertising more automated and predictable.  Thanks Facebook for the kudos.

Dynamic Product Ads – Now With Less Beta!

Today, we are thrilled to announce the launch of Dynamic Product Ads, a major step forward in making Facebook Ads more relevant for people, and more effective for marketers.

Over the Summer, we worked with a select group of top online retailers to perfect the features of our flagship product, and we are excited to officially take the product out of beta and offer it to a broader group of clients.  Initial beta testers included One Kings Lane, Gilt Groupe, as well as some large traditional retailers who always like to remain nameless.

Our Secret Sauce

blogimageThe secret sauce of Dynamic Product Ads begins with our approach of advertising at the individual SKU level.  This approach requires automation of product selection, targeting, and bidding to help wrangle thousands of SKU’s at a time.  Our software crawls an ecommerce website and extracts all the products and attributes – and then surfaces them in a workflow that empowers the marketer to promote large groups of products at a time.  This capability is particularly relevant for retailers with large catalogs, or flash sales sites that have new products added and removed every day.

One Kings Lane

Jim Kingsbury, VP of Marketing at One Kings Lane, underscored this point when he said:

“SocialWire’s ability to automate ad creation is a great fit for our highly dynamic inventory model.  Their team also has many interesting ideas as well as an impressive grasp of our business.”

One Kings Lane was one of the first clients to join our beta test and demonstrated our value for flash sales; we have since added more traditional retailers and are now showing the value for less dynamic catalogs as well.

Diamonds in the Rough

SocialWire solves the creative fatigue problem that comes with most traditional marketing campaigns by using the diversity of a catalog to test creatives at a massive scale.  Promoting hundreds of new ads every day is a more scientific way to find the diamonds-in-the-rough products that perform well as ads on Facebook.  Unless retailers can test different products against different audiences, they miss the opportunity to extract the maximum economic value for each product.  Once identified, these high performing ad-to-audience pairs can generate a majority of a campaign’s revenue.  In some cases, we saw ROAS (Return On Ad Spend) in excess of 1000% for these “rock star” ads.

Data and Market Intelligence

Perhaps even more valuable than finding high-performing ads is the market intelligence that can be derived from product-level ad testing.  When our clients told us they were making inventory buying decisions based on our data, we knew we were onto something bigger than just ads.   Traditional campaigns typically use a handful of images and point at a landing page (e.g. “Shop our Summer Sale Now).”  Success or failure is binary and the marketers only learn which audiences have affinity to their brand.  By using individual product ads that point to a product page, marketers can not only drive transaction volume, but also get data about which products resonate with which audiences.

Social Engagement Drives Performance:

Another benefit of product-level ad testing is the ability to find the products that go viral on Facebook.  We ran a test of 450 unpublished photo posts (or “dark posts”), and found that 3% of those ads had high social engagement – defined as 20 likes or more.  These 14 ads generated over 60% of the user acquisition for the whole campaign.  More importantly, they drive the signups at a much lower cost – 71% lower than ads with no likes.  Some clients use the social engagement performance they see on SocialWire ads as an indicator to influence which products to promote on their fan pages.

Making Facebook Marketing More Like SEM

In order for Facebook to become a marketing platform that e-commerce companies can rely on for predictable revenue, the industry needs more ability to create evergreen campaigns that run in mostly automated ways.  SEM is a great example of how marketers generate predictable revenue from a mostly automated and evergreen set of campaigns.  Applying this same principle to Facebook is challenging given the peculiarly high commerce potential of search relative to Facebook, but Dynamic Products Ads are a first step toward the goal of making Facebook a more automated marketing environment like SEM.

 

 

 

 

Our Team Photo

Well, this is embarrassing.  I was about to announce our newest hire and had carefully planned to synchronize the announcement with an updated team panorama photo, and then Valleywag blew the story with their hard hitting coverage of our scandalous photoshopping (actually, we used MS Paint).  I want to take this opportunity to publicly apologize for using photoshop to update our employee photo, and I hereby commit that we will NEVER do that again.  With that, please see our updated team photo with the newest addition to our team, and I assure you, he has some serious Facebook skillz.

sw-panoramic-3

Images and Automatic Color Analysis

 

sushi1

Sushi image source: wikimedia commons

The longstanding intuition of advertisers that the color of an image is a key factor to the success as an advertisement demands an investigation into the relationship between image color and ad performance. Some of our overall best performing colors were DarkOliveGreen, Chocolate, and DarkSeaGreen while Crimson, SteelBlue, Silver, and Tan seemed to repel high performance. We can also use color to think about more specific questions:

  •  Are brightly colored images more effective on the outer edges of a page? (there are indeed colors that perform differently)
  • Are men more partial to red than brown? (not statistically significantly)
  • Does skyblue catch the eye of people who like hiking boots?

The stronger the patterns in the data, the more applicable they are to making optimizations of ads. Should we never show a Crimson ad? Even ignoring the distinction between correlation and causation the answer is no, as there are some crimson ads that do quite well. Instead we will use this information as a factor- at the same time taking into account other details, such as the relationship with product type and audience.

To see a clear picture of the truth behind these intuitions we need to examine a lot of data. For which we need an objective ‘eye’ that can properly judge the color of an image and handle the color labeling of thousands of pictures quickly.

Color Distance

color_distance

Though the two black rectangles are the same size, the colors they connect appear drastically different.

The first issue we have to tackle is how to define the distance between two colors. We can think of colors existing in 3-dimensional space- and think of their distance as a geometric measurement (think back to your geometry class and ((x1-x2)2+(y1-y2)2+(z1-z2)2)). However, there is the complication that the human eye doesn’t see colors in a perfectly even grid structure. Our squishy organics tend to perceive some colors as closer than others when they are “geometrically” the same distance apart.

Charles Poynton has written a good article on a computationally efficient compromise that helps overcome this problem http://www.compuphase.com/cmetric.htm

 

Clustering

KMeans

Image source: wikimedia commons

Once we have established how to measure distance, we can use it to run a clustering machine-learning algorithm (the fancy title makes that sounds harder to understand than it really is). The basic idea is that a bunch of data, eg. different colors of an image, can be divided into groups and the average value of each of those groups can be returned. This means that we can group all the reds (and return the most average red color), group the greens, etc.

The trick to it is that we don’t just want to have the most average red, green, blue… but instead we want to have the most average representative colors of our particular image- in the case of the sushi, we want a red, a brown, and a white.

The ‘learning’ comes in by performing many rounds of dividing into groups and calculating the best average colors. After each round, the image colors are re-divided according to which average color they are closest to. Then the average color is re-calculated from the image colors closest to it, and so on until the average color stops changing between rounds. Andy Shabalin has a great javascript web demonstration of clustering if this wall of text was not enlightening (or even if it was it’s worth checking out for the slightly hypnotic nature of the animation).

 

Binning Data

color_transition

So now we have the color of our images, but have been left with another problem. Our possibility space is HUGE! We have a ton of possible colors (2563 or ~16.7 million) and would need way more images than that to run the statistics that are used in the end of this post. What we need is a way to reduce the space down to a reasonable number of colors. And a human-readable name would be a great added bonus. If we look back to 1996 we see that the W3C released a CSS module that standardized a set of 255 colors. We can group our specific image colors into one of those 255 named X11 colors. Depending on the number of input images you’re working with this may need to be further down-sampled.

χ² Test for Independence

chi_sq_hist

A random sample of the chi-squared distribution.

The Pearson’s Chi-squared Test for Independence is an indicator of how much influence one set of data has on another. In our case, we are trying to see how much influence image color has on whether an ad is effective or not effective.

There are three parts to the Pearson’s Chi-squared Test for Independence:

  1. Calculate the chi-squared statistic, which you can think of as a way to measure how “wobbly” the differences are across the different colors.
  2. Determine the degrees of freedom, the more colors you are trying to compare, the more degrees of freedom there are.
  3. Use the degrees of freedom to determine what chi-squared value means the results are statistically significant (a real result instead of random noise). The more things you are trying to compare the more likely small wobbles are to be just noise.

One of the requirements to perform a valid test is that we need to have enough data in each color bin. A good rule of thumb it that most of the color bins need to have at least as many images as there are total colors (something to keep in mind when you are trying to get your final set of colors). If any of your color bins have a low number of images in them, making a claim whether that particular color really is effective or not is more difficult.

 

Results

It seems there really was something to the old adage that color and clicks are related! A “significant” p-value means that color and performance are not independent. Keep in mind that this test implies a correlation, not necessarily a causal relationship. So maybe “DarkOliveGreen” is a good color. But it could also be that it’s awesome pictures of dinosaurs that happen to all be green that are really what causes the ads to do well.

low high ratio
DarkOliveGreen 17 25 1.47
Chocolate 28 39 1.39
DarkSeaGreen 13 17 1.31
DimGray 81 89 1.10
Gray 90 97 1.08
Sienna 69 74 1.07
SandyBrown 21 21 1.00
RosyBrown 96 93 0.97
SaddleBrown 22 21 0.95
LightSlateGray 26 24 0.92
Peru 76 61 0.80
Brown 26 20 0.77
DarkSlateGray 63 48 0.76
DarkGray 66 50 0.76
BurlyWood 22 16 0.73
MidnightBlue 18 13 0.72
DarkKhaki 38 26 0.68
DarkSlateBlue 22 15 0.68
FireBrick 46 29 0.63
Silver 47 25 0.53
Tan 82 35 0.43
Crimson 57 21 0.37
SteelBlue 28 9 0.32
OliveDrab 6 10 1.67
DarkCyan 9 11 1.22
SkyBlue 14 8 0.57
DarkSalmon 9 5 0.56
MediumVioletRed 15 8 0.53
SlateGray 10 4 0.40
Chi-squared dof P-value
61.6493 28 0.0002497

Note that grayed out values are those that don’t have enough data to be certain about their significance.

“Low” vs “High” columns have the number of ads with lower than average click-through-rate (CTR) and higher-than-average CTR, respectively.

Our Latest Funding, and My New Role

Today, we’re excited to announce that we have raised $1 Million in additional funding, and I am taking on a new role as the company’s CEO.  The new round is a Seed Extension, and includes new investor SoftTech VC, as well as participation from First Round Capital, 500 Startups, Accelerator Ventures, and Joi Ito.  We plan to use the funding to develop technology behind our Dynamic Product Ads Platform.

The reason we chose to do a Seed Extension round instead of a larger Series A was because we wanted to prioritize more product development before we scale out our sales and operations.  We believe that technology must be at the core of what we do if we are going to solve important problems – and our seed investors understand that technology is not built overnight.  It takes at least two nights… maybe three.  Regardless, this latest capital infusion will give us the time we need to build a strong technology underpinning to our product that will help us scale more efficiently.

I am thrilled to be entrusted by our Founder Selcuk Atli, and our Board with the role of CEO.  Selcuk did an incredible job getting the company to this stage, having raised two rounds of funding and built an all-star team.  He will keep driving SocialWire’s vision and pushing our product strategy in new directions as Executive Chairman.  We’re all excited for what’s to come.

Introducing Dynamic Product Ads

SocialWire is a recommendation engine for ads.  Our mission is to make ads as personalized and useful as product recommendations that you see on Amazon or other retailers.  One common example that people are familiar with is the ubiquitous “people who bought this, also bought…”  This feature is a great example of using programmatic signals to match the right products to the right audiences.  While retailers have been using these signals to power personalization on their websites for years, the next evolution will be to power personalized product ads based on programmatic signals.

Our Beta Clients

We recently launched DPAs in beta with a few clients, including Living Social, Shoedazzle and a few other major retail brands.  We’re thrilled to have these great brands as clients, and while we’ve been focused on just them during our beta test, we look forward to inviting more clients into our program.  Two consistent themes across these clients that have made them successful are personalization and curation.  We are helping them translate those themes into their marketing campaigns.

Curation and Personalization – Key Success Drivers

We have all seen an ecommerce boom in the last few years for the retailers that have embraced the modern success drivers of online sales – personalization and curation.  But advertising is doing what it always does – lagging behind.  Dynamic Product Ads enable retailers to offer the same level of curation and personalization they do on their websites, in their ads across the web.  As a retailer, your customers spend only a fraction of their time on your site compared to the time they spend on Facebook and other sites.  Rather than wait for them to come to you, why not bring your personalization and curation to them?

Indeed, Daily Deal and Flash Sales sites have seen success bringing their curation to your inbox.  But we’re growing accustomed to skipping over those emails, or letting our smarter email clients filter them out for us.  The technology now exists for us to create a rotation of relevant product ads to show me as I navigate across the web – not just in my inbox.  Even if I’m not a customer yet, you can still curate a rotation of product ads for me based on other signals like my Facebook interests.

Dynamic Ad Creation

Retargeters are familiar with dynamic ad creation – when someone visits a product on your site, and then you dynamically create an ad from that page and retarget that user on other sites. SocialWire is taking that same dynamic ad creation technique and applying it to prospecting as well.  For example, SocialWire can dynamically build ads for every new product on sale, and feature each product in the ad.  Our tool enables a marketer to combine static copy with dynamic parameters like so:

{NAME OF SITE}

Check out this {PRODUCT TITLE} from {BRAND} for {PRICE}.

{IMAGE/LINK}

So the ad itself would then look like this, for thousands (or millions) of products….

DarkPost

Dynamic Keyword and Location Targeting

Most Facebook campaigns use only a handful of ads, so the targeting can be done manually, but when you have thousands of unique ads the targeting must be done in bulk or dynamically.  We extract product information from the page such as brand, and other relevant keywords, that we dynamically convert into Facebook interest targeting.  We have some clients that are doing localized daily deals, so we scrape the location from the page, and dynamically convert that to target the appropriate zip codes on Facebook.  Thus, a daily deal for Golf Lessons in San Francisco would automatically be targeted to Facebook users in SF who are interested in Golf, Tiger Woods, The Masters, and other relevant interests that our algorithm finds.

Custom Audiences

In addition to using data from a retailers product pages, we can also use a retailer’s customer data to dynamically target product ads.  We use Facebook’s Custom Audiences platform to match all the data that a retailer has on their users to match those specific products to the right customers.  This data could include products purchased, products abandoned in the cart, etc.  For sites that do highly personalized experiences like Shoedazzle, they can now target every one of their customers with a unique set of product ads tailored specifically for that user.  It is like a recommendation engine for ads.

Dynamic Conversion Optimization

Not every product or deal is going to work as an ad – so it is important to be able to optimize for the ones that are getting traction.  We do this through dynamic conversion optimization.  By enabling marketers to promote every product, we need to enable them to also tell us how much they’re willing to pay to advertise that product before someone buys it.  For example, for a $100 product, a retailer might be willing to spend $30 to promote it; while for a $50 product, the retailer would only want to spend $15.  We let the marketer simply input their margin (e.g. 30%), and then we dynamically apply that margin to the price that we pull for every product we promote.  For the products that are not selling profitably, we pause the associated ads, and reallocate that budget toward the profitable ads.

Requirements:

Getting started with Dynamic Products Ads is very simple and requires no integration.  Our standard product crawler will pick up any Facebook open graph tags by default, and for larger clients we will customize our crawler to pick up objects and attributes in whatever format they are currently tagged.  We have a waiting list at the moment, and are taking on new clients as soon as we can.  You can join the waiting list here.

We’re a PMD!

We’re thrilled to announce that last week, Facebook accepted SocialWire into the Preferred Marketing Developer (PMD) program with the Apps badge.

We at SocialWire remain a one-two punch for Facebook Ads, but increasingly find ourselves focused on the second punch – the advertising. We’re excited by the challenge of helping Facebook Ads become a more relevant and native experience.  Facebook is enabling some truly groundbreaking ways of reaching new and existing audiences, and we’re eager to bring our dynamic display platform to the table as an option for marketers.

SocialWire’s philosophy is to enable more automated and scalable ways for marketers to promote their products.  For example, (more…)

A Case for SKU-level Ads

Sperry
Imagine Facebook as a shopping mall.  You hang out with your friends there all day, and retailers are trying to get you to come to their store and shop.  In the shopping mall, retailers use display windows.  On Facebook, they use display ads.  The strategy is the same – pick the most provocative products, display them, and try to get the customer to walk in the door or click through to the site and look at all the other products.  But there’s one key difference.

In a shopping mall, retailers can display three to five products at a time, but on Facebook retailers can display every product in their catalog.  But retailers are still using the same approach online as they are in shopping malls – they promote three to five products.  On Facebook, it’s now possible to programmatically create thousands of ads – one for each product – and target each product to exactly the right customer.  This is SKU-level advertising.

Google Product Listing Ads:

Google recently launched a SKU-level ad product called Product Listing Ads (PLA’s). (more…)

Making the World a Better Place, One Facebook Ad at a Time

Our Mission: Make Advertising a Better User Experience

Our mission in launching SocialWire is to make advertising on the web and mobile a better user experience. When I left Digg a couple years ago, one of the biggest regrets was leaving behind DiggAds – the highly successful native ad product that I’d been responsible for creating and bringing to market. DiggAds went on to become the legacy for other social native ad products like Promoted Tweets and Facebook Sponsored Stories. Seeing the success of DiggAds made me a believer that advertising is not a zero sum game – if users like the ad experience, marketers will do better.

The Problem: Marketers Can’t Promote Individual Products

The problem we are setting out to solve is that display advertising is still a relic of magazine ads – the process of placing ads is labor intensive and still a lot of guesswork. What this means for marketers is that they have to make compromises, they can’t advertise their entire product catalog, they need to artificially choose a handful of products to market, and they need to guess at what their target audience is going to be. For example, eBay gets over a million new products added to its catalog every day, it would take an army of people to create ads for each product and figure out whom to target. But what if marketers could advertise every single product to exactly the right people – market the long tail of their products to the long tail of their customers?

(more…)