This blog has been thoughtfully written by a human (Max Sinclair, Founder of Ecomtent) not generically spat out by AI. Max spent +6 years at Amazon, including working closely with the A9 teams whilst owning Browse & Catalog for the country launch of Amazon Singapore. He also studied Deep Learning Essentials at the Mila Institute for Artificial Intelligence.

Spending the weekend at Seller Sessions '24 brought a debate raging throughout the Amazon world into real life and in person – are keywords dead? Almost every speaker, including former Amazon employees and 8/9 figure sellers referenced RUFUS (Amazon’s new AI “shopping assistant” that operates like Chat GPT from within the search bar) or COSMO (Amazon’s “Common Sense” algorithm). On the two latest episodes of the New Frontier Podcasts – Bradley Sutton from Helium 10 and Vanessa Hung of Online Seller Solutions we heard perspectives from both sides of this debate, and in this blog I will offer my take. To be upfront, much like Nicola Sturgeon of the Scottish National Party I am incredibly biased towards ‘Yes’ (I accept non-UK based readers will not appreciate that reference).

‍

Definitions

I want to first off start by defining some terms. By "Keywords", I mean the practice of sellers placing huge emphasis in researching high volume, longtail or some other strategy of keywords for inclusion in their product listing. The goal of this was to optimize for Amazon's traditional A9 Algorithm. Simply put, A9 functions by indexing details such as titles, descriptions, and backend keywords; and matching them with customer search queries. It then ranks results based on relevance by computing various performance metrics like sales history, reviews, CTR, etc. Of course, Keywords will still be relevant to a certain extent. Language is not changing. My focus on describing what is "dying" in this blog is the fanatical obsession, the hours spent in researching Keywords as a priority to launch a new product, or optimize an existing listing.

Now let's define "Context", which I take as matching Amazon COSMO's "common sense reasoning" as to what a customer is searching for based on their intent and their age, gender, geographical location, etc. To quote Amazon Science: "COSMO involves a recursive procedure in which an LLM generates hypotheses about the common-sense of implications of query-purchase and co-purchase data; a combination of human annotation and machine learning models filters out the low-quality hypotheses; human reviewers extract guiding principles from the high-quality hypotheses; and instructions based on those principles are used to prompt the LLM". To put it more simply, COSMO infers customer intent via a training cycle where an AI system makes predictions, gets feedback, and uses the learnings to make better predictions guesses moving forward. Amazon science explains the functionality: "If a customer, for instance, submits a query for “shoes for pregnant women”, the recommendation engine should be able to deduce that pregnant women might want slip-resistant shoes. We’re building a knowledge graph that encodes relationships between products in the Amazon Store and the human contexts in which they play a role — their functions, their audiences, the locations in which they’re used, and the like. For instance, the knowledge graph might use the used_for_audience relationship to link slip-resistant shoes and pregnant women."

‍

The logical case for Yes

The case for yes is a technical yet simple one. If you believe that AI powered search is the future rather than traditional search, either by customers directly using RUFUS, or by Amazon over time putting more emphasis on COSMO logic in the backend rather than A9, than context is inarguably more important than keywords. This is because LLMs do not operate in words. LLMs break down input text into a sequence of tokens, which can be individual characters, words, or subword units (like "un" and "likely"). Each token is mapped to a numeric value for processing. The model then processes this sequence of numeric token values to generate its output, token by token. This token-based approach allows LLMs to handle a vast vocabulary, generate coherent text in multiple languages, and understand rare words or phrases not seen during its training.

What's more, before generation, most LLMs will compress the original prompt to reduce the overall model sizes, computational costs and fit within token windows. Prompt compression techniques remove non-essential tokens from prompts using a smaller language model. This compressed prompt, while difficult for humans to understand, can still be effectively processed by the LLM to generate relevant outputs. So post compression, so-called "longtail keywords" sellers are trying to rank for will likely not even make it past step 1 of the COSMOS workflow diagram below.

‍

The evidence

The evidence? Amazon have explicitly stated they are rolling out this new system in their Original COSMO paper published earlier this year. To quote "The integration of COSMO into our online search navigation system has led to significant business improvements, underscoring the power and potential of COSMO-LM based applications. This conclusion is drawn from meticulously conducted Amazon online A/B tests carried out over several months in total, targeting approximately 10% of Amazon’s U.S. traffic. These well-structured tests revealed a notable 0.7% relative increase in product sales within this segment, translating to hundreds of million dollars in annual revenue surge. Additionally, an 8% increase in navigation engagement rate was observed within the same traffic segment, highlighting improved customer interaction and satisfaction. The success of this initial implementation indicates a tremendous opportunity: by extending the adaptation of COSMO-LM to encompass all traffic for navigation, we anticipate the potential to generate a revenue increase in the billions". I highly doubt Andy Jassy is going to waste any time in implementing something that Amazon has publicly said is worth billions in incremental revenue.

Some of the 8/9 figure sellers and Agencies I spoke to at Seller Sessions had already seen Amazon experiment with COSMO and context matching in the Advertising world. For example Oana Padurariu, Head of Amazon at Trivium Group, told me that last year across multiple different categories and seller accounts in the US market, exact match Sponsored Products were no longer matching keywords word-for-word. In escalations, rather than trying to troubleshoot, Amazon was offering reimbursements - an early indication something new was being tested. In the new semantics matching, Sellers can no longer provide negative keywords to a product targeting campaign, only negative ASINs. This again suggests a shift in the underlying logic is already upon us.

‍

How can Sellers optimize for context?

To quote Jon Derkits at the start of his "9 Hacks Presentation" at Seller Sessions, the best way is to build great products and focus on the customer. And this is exactly what I would recommend sellers to do, focus on understanding their audience, their ideal customer profile, and build products and product listings that speak directly to them. A good framework to help you do this is Bob Moesta's 'Jobs to be done'.

I had a zoom call with Bob, as he is close friends with one of my investors. On the call he defined Jobs To Be Done (and I quote directly from the meeting's transcript): “Jobs To Be Done is about understanding the progress a customer is trying to make. You need to understand where people want to go, and what causes them to get there. Value is defined by the context a customer is in and the outcome they seek. It starts with understanding why people say ‘today is the day I need a new piece of software, a new product, etc’. It’s not random, you need to understand why today is the day they decided it was time for something new. Industries are structured by categories, customers' mind aren’t. Most people don’t know what they want. You are trying to understand what the customer wants, without talking about the technology or solution at all. If you understand that, you will build better products that fit into their lives.”

Bob’s framework starts from the decision point, and works backwards to try and understand the underlying motivations. The classic example of this framework inaction was Bob’s work was Snickers. In the 2000s Bob had been drafted in by the Mars to help with their failing product line, Snickers. During his consulting period, he observed a man in the airport purchase a Snickers. Bob followed this man to his gate, where he sat down and scoffed it down whilst replying to emails. When Bob inquired why he choose to buy a snickers bar, the man responded that he needed something to eat quickly, which will enable him to power through his work. In this regard, Snickers wasn’t competing with other chocolate bars, but rather a can of red bull, a coffee, or an energy bar. Indeed, following gaining this valuable insight, he dug deeper and found the average time people spent consuming a Snickers was 60-90 seconds, compared to other similar products such as Milkyways, which was 10-15 minutes. The problem Snickers customer faced in this instant was “when I’m on the go and hungry, I need something to give me energy quickly so I can concentrate on the task in hand”. Zeroing in on why customers ‘hire’ a snickers bar led to a rebrand of Snickers, a campaign you probably recognise - “Snickers, because you’re not yourself when you’re hungry”. The impact of understanding and nailing its job was profound – Snickers went from a failing product line to claim the #1 candy bar spot in the world. Understanding the context of why customers' enjoy Snickers, and writing the listing accordingly, would be more important for discoverability on Amazon in 2024 than trying to hit keywords such as "chocolate bar" or "peanut bar".

Beyond this fundamental piece, I can share what we are helping sellers to focus on at Ecomtent, from our own testing and experimentation.

1. Enhanced Visual Content: COSMO is multimodal, thus can understand images not just text. Sellers should use the maximum allowed number of images per listing and A+ Content, including multiple demonstrative lifestyle images matching intent of customer queries. We are helping sellers to do this with our AI Generated Product Images.
2. Monitoring & Optimizing: We are using our data across millions of product listings to monitor conversion and continuously optimize as we see these AI models drift (which they do).
3. Backend Attributes: This was my bread and butter whilst working at Amazon. We helping to optimize backend attributes for discoverability, more details of which I will share in a subsequent blog.

You can test how optimized your current product listings are for COSMO's new algorithm for free at www.amazoncosmoaudit.site