Credit: Google News
Etsy plans to improve search by leveraging Machine Learning technology and by personalizing search results, according to its Chief Technology Officer Mike Fisher. But that could make sellers nervous based on their experiences with similar attempts made by other marketplaces. (Mention the word “Cassini” to an eBay seller and watch their reaction.)
During Etsy’s recent Investor Day presentation, the executive talked about the current state of search and initiatives that are under way.
One of the most complex challenges for Etsy engineering is creating best-in-class search, Fisher said. Search is more difficult for Etsy than all the other etailers because many of the things that make Etsy so special makes it extremely complex: there are no SKUs, there’s no catalog, items are one of a kind, and they’re often customized. “That means when they’re gone, they’re really gone,” he said. “We don’t get the chance to sell the same shirt 500 or 5,000 times.”
“Sellers label and tag items how they want to, how they want to describe the items. Combined with the fact there are 60 million items in inventory makes it a more interesting tech challenge.”
He also said sellers try to game the system – for example, he said, by putting the name of a holiday in listing titles. He used the example of a seller who added the term “Valentine’s Day” to the title of a listing for a pillow, and in an accompanying slide, he showed a listing for a pillow with the title, “Colorful Earth Tone Pillow, 24×24, Valentines Day Gift, Rustic Pillow, Desert Trend Pillow” to drive home his point.
Etsy builds its search indexes using only the title and tags. Descriptions currently are too unstructured – “we allow sellers to put whatever they want in their descriptions, they can tell their story, how they built it, where they sourced it, anything they want.”
And image quality is currently too varied – “until recently, image recognition was extremely difficult,” Fisher explained.
After retrieving all the items based on the title and the tags, Etsy then use predictive conversion rates.
Etsy takes search terms from the previous several weeks and analyzes them to see how shoppers interacted with the results (did the shopper click on them, add to cart, favorite them, etc.), then it builds Machine Learning models that can rearrange the top search results, in real time, so that the first page shows the most relevant item.
“The first page of search results generates 83% of the search purchases. It’s critically important to get the first few listings right. So moving a relevant listing from page 5 onto page 1 is important for our business,” he said.
Fisher believes there are many other factors Etsy can use in addition to titles and tags. For example, he believes Etsy will be able to leverage Machine Learning to help it understand the information in images and to pull useful information from the unstructured data in descriptions.
He said there are dozens of other factors Etsy could be using, including seller rating, shipping times, and price. That sounds a lot like eBay’s Best Match default search order, which has been panned by many sellers.
He then discussed plans to personalize search results. “Relevance is all relative. The results should look different depending on who’s searching.”
If any of Fisher’s goals sound familiar, it’s because eBay faced similar challenges with unstructured data. When Devin Wenig replaced John Donahoe as CEO of eBay in 2015, he launched a Structured Data Initiative and promised it would turn around eBay’s performance. Like his predecessor’s Cassini project, it hasn’t been a magic elixir.
Credit: Google News
