Loading Form...
Mar 14, 2008 | 8 minute read
written by Linda Bustos
We just wrapped up our webinar on merchandising (cross-selling and up-selling) with Mike Svatek of Baynote.
This was an incredible session and I'm sure you'll get a lot out of catching the Webinar replay which will be posted within the next few days. The replay will walk you through all of the screenshots used in the presentation - I'll only be using select screenshots for this recap.
Mike chose the king of cross-selling Amazon to illustrate the concepts in the webinar, sharing an impressive statistic:
35% of Amazon Sales come from cross-sells & recommendationsVenturebeat (Dec 06)
How does Amazon do this?
First-Time Visitors - Pre-Intent
If Amazon has no information on you (your first visit, you are not logged in or your cache and cookies are cleared) you'll see default merchandising (pre-intent) within a number of merchandising zones, what Mike refers to as a shotgun approach:
Search Box Query
If you head to a search box and enter a query, you start hinting at what you're looking for, and typical Amazon results appear with an item thumbnail and description. Amazon's algorithm is likely using keyword match and user reviews to deliver results, and perhaps purchase popularity as well.
Your search history doesn't matter at this point. All you care about (and the results depend on) is your current query.
Purchase History Cross-Sells
Once you select an item to view in detail, Amazon pulls cross-sells and displays them on the page like so:
This example was a book about karate for kids, and the cross-sells are other books on childrens' karate based on others' purchase history (transaction data on the back-end). One thing to point out is that people don’t necessarily buy all of them together -- maybe one or two but not the whole set.
There is some affinity between these products, but a lot of comparison data has been lost. This may be fine for books - one may be interested in collecting several books on a similar topic. But for highly considered products - like digital cameras where the customer isn't going to buy 5 or 6 different cameras at once. Just another reason why you shouldn't copy Amazon because it does something - it really has to make sense on a product-by-product basis.
Dynamic Merchandising on Home Page
Once Amazon has gathered a bit of information about your intent, it can show you more relevant products on the home page. When you refresh the page you'll notice the cross-sells move around, indicating that Amazon is likely testing out which merchandizing zones or which offers get the best click through or conversion rates.
Amazon will also show related items from other categories - like karate toys, video games or training videos based on "similar searches." In other merchandising areas Amazon can show other things boys aged 7-12 would like - Spiderman merchandise, for example.
Cross-sells will change on the home page as you click on or search for other items (change your intent). A challenge with collaborative filtering and profiling in general is that people have different purchasing roles. You're an uncle, husband, son, employee, best friend and individual - your search and browse history may be for you or someone else, even in the same visit. So changing cross-sells with each new "intent" at least keeps things current, and remembering past intent is better than ignoring it completely.
When do cross-sells work?
Cross-sells work well for considered purchases (high involvement rather than impulse - typically higher cost) provided they are lower cost accessories related to the product. They also work for smaller purchases with small accessories like Barbie and an outfit. You want to keep the cross-sells at half the price or less. When they are more than 1/2 the price of the item considered, the attach rate is low.
Products with natural bundling are also good, like cameras with lens, cleaner, memory cards and warranty.
When do cross-sells fail?
Don't try to push higher priced items together with lower priced. People who buy a camera may buy a camera lens at the same time, but it's unlikely someone adds a lens to cart and then all of a sudden wants to buy a camera. Same with sports tires - you wouldn't try to upsell a Porsche.
Be careful that you don't just look at correlation in your analytics data - but consider the primary and secondary intent. You may want to manually add constraints to your rules engine so you don't goof your directional selling.
When do up-sells work?
Upselling (suggesting a similar item instead of the item being viewed) must have a small difference in dollar value or a small nominal percentage difference - 10-20% max. You need to show some incremental value for the increase in price.When do up-sells fail?
When important attributes are different (red vs. blue dress) or when you show items that don't have the features the customer is looking for. You can also fail by showing different brands. If a customer owns a Nikon he needs Nikon accessories, not Pentax or Canon.
You also need to consider any contractual agreements you have with suppliers and brands. For instance, you may not be allowed to show certain brands next to each other.
There are 2 sides to this debate:
A. Interruptive merchandizing doesn’t work. People are in context of a shopping experience and are less likely to complete the transaction if distracted.
OR
B. Shopping cart merchandizing is natural, accessories can be added as impulse items, or forgotten items, to increase the average ticket.
Mike says they both work and cited Amazon and Best Buy as examples. Really you must do your own testing. Baynote has A/B testing built into its system. You could also use multi-variate testing and directed revenue tests. Directed revenue tests aim to discover how much revenue you can drive through a recommendation without showing it to different groups or changing the user experience.
It's a major challenge to develop, code and maintain rules for the long tail - not only are there a lot of products but by nature of being long tail, they are less frequently viewed and sold. There is a lot of time lag in working SKU by SKU as you have to take trends in your data, analyze them, come up with a strategy and apply the new rules.
Rather than manually programming your long tail cross-sells, you're better off using dynamic recommendations.
Purchase History - Based on personal history or collective history (collaborative filtering).
Catalog Rules - Using an "If... This..." approach: "If viewing Nikon SLR, then show other SLRs $100 higher and lower." This gives you extreme control which is good, but also carries a lot of responsibility and time lag.
Browsing Behavior - Customers viewed this/that. This can be session-specific or profiled over time. This approach requires less maintenance, you can profile people over time and your cross-sells can change as a person’s behavior changes. The challenge is that your site navigation influences browsing behavior. Clicks don’t necessarily indicate interest, rather they may be the result of your navigation, links and layout.
Keyword Match - Matching search terms with product descriptions to pull recommendations. There is explicit intent but you're just matching lexical terms. People use different terms to describe the same things so you need to integrate thesauri and dictionairies to make associations.
Community/Emergent - Combining behavioral statistics with on page actions like comparative shopping behavior, virtual bookmark, virtual print, scroll rates, browse patterns, referrer context etc. When someone is comparing products, really engaging and reading, this signals more interest. You get a higher fidelity set of recommendations by looking at the way people shop. You can rapidly adapt if this is implemented properly.
When using the "wisdom of crowds," watch out and make sure that you are pushing higher margin items, respecting inventory constraints and any contractual brand associations you may have. Use tools to constrain your set of recommendations.
We believe Mike has coined the term SEOandizing - and we love it!
What is SEOandizing? We can define this term as optimizing your cross-sells based on referring keyword from a search engine.
SEOandizing can help reduce your search engine bounce rate by showing cross-sells for the referring keyword. If a visitor doesn't find what he wants on your site, he's likely going to hit the back button. But if you show keyword-related cross-sells right on your landing page, you give them reason to keep searching on your site. Especially since search engines are going to use their own algorithms to determine which page on your site best matches the keyword typed into its own search box. Whereas your own internal search engine may be programmed to show a best-seller or a higher margin item.
This is a brilliant idea, thank you for sharing this, Mike!
Here's a nice example of what's possible with email cross-sells:
Again, we'll have the replay up for you next week.
You might also find yesterday's Internet Retailer article on merchandising strategy planning interesting, particularly the statistics on how 187 merchants surveyed by the E-Tailing group make merchandising decisions:
Respondents were allowed to cite more than one information source.
If you can't get enough merchandising tips, check out our own posts on cross-selling:
A Good Example of Bad Cross-Selling
Cross-Selling Tips for Online Retailers