Amazon's A9 Algorithm is set for a major shakeup. Amazon is upgrading its search with AI. Forget keyword stuffing; the new system understands context and matches queries to products in a smarter way. Existing keyword tools may become outdated. Thanks, Kevin.
Kevin King
➡️ Post | Profile
COMING TO THE AMAZON A9 ALGORITHM
Amazon is working on both physical GPU-based technology and its own unique e-commerce LLM. It can process massive AI-related search analyses to serve up results in the blink of an eye.
This could take away the need to be indexed for everything, eliminate keyword stuffing, and change how we try to optimize listings to rank today.
Let me introduce you to Amazon’s new LLM:
Formerly introduced by Amazon this past May in Osaka Japan at the 27th Pacific-Asia Conference on Knowledge Discovery and Data Mining.
Traditional methods of matching products with search queries just look for exact words or phrases in the product data & descriptions. This is fast and simple but has some problems. For example, a small spelling mistake or a synonym, might not get the results someone wants.
To fix this, newer methods use something called "semantic matching" which tries to understand the meaning behind words rather than just the words themselves. Like a search for "sneakers," would also show "running shoes."
Now Amazon will soon be launching a new method to improve search results using a large language model called BERT.
There is a four-step method to train an e-commerce LLM model that can efficiently match queries to products to buy:
1. Domain-Specific Pretraining
Train it on e-commerce data. E-commerce language is different.
Amazon created a special vocabulary from around 1 billion product titles and descriptions from 14 languages, fine-tuning with 330 million query-product pairs using Amazon Web Services (AWS).
Tools like Deepspeed and PyTorch were used, and training was done on AWS P3DN instances.
2. Query-Product Interaction Pre-finetuning
They improve the model by making it understand how search queries relate to products. They use a dataset where the queries and products have a strong relation, and they mask parts to train the model to predict the missing parts.
3. Finetuning for Matching
This stage trains the model to match queries to products. They employ a bi-encoder system which is efficient for large-scale data.
4. Knowledge Distillation to a Smaller Model
The final step is to transfer the knowledge from the large, well-trained model to a smaller one that can work quickly in real-time for product search.
Bottomline: A single sentence in a product listing could soon match hundreds of search phrases without keywords needed explicitly mentioning them all.
Tools like Helium 10, Jungle Scout, and Data Dive will need to pivot with their keyword tools, or they could become irrelevant.
The rest of this story and many more tips for Amazon sellers are published 2x per week in my FREE Billion Dollar Sellers dot com newsletter (link in comments). Many of the industry elite and loads of big sellers read it religiously.
TAKE NOTE: There is no such thing as an A10 algorithm - that’s marketing click bait. It’s the A9. It evolves over time like you did during puberty. But its name doesn’t change.