Published: 30 June 2026 by Dan Williams
Series: Notes on eBay & Discovery
Looking beyond Cassini to explore how search, recommendations and buyer discovery appear to fit together.
Whenever eBay visibility is discussed online, one word tends to appear very quickly.
Cassini.
Cassini this. Cassini that. Optimise for Cassini. Feed Cassini. Refresh your listings for Cassini. Whisper softly to Cassini under a full moon.
Some of that advice may contain useful fragments. Much of it probably contains assumptions. A lot of it treats Cassini as though it is a single all-seeing marketplace brain controlling every visibility surface on eBay.
I’m no longer convinced that’s a useful way to think about it.
Not because Cassini is unimportant. Clearly it isn’t.
But because eBay’s own public engineering articles describe something broader and, to me, far more interesting.
Search matters. Best Match matters. Ranking matters. Titles matter. Item specifics matter. Seller reputation probably plays a part. Buyer behaviour feeds into the picture. Recommendations, similarity, images and personalisation all appear to have their own place.
In other words, eBay search and discovery looks less like a single algorithm. eBay’s own engineering articles consistently describe specialised retrieval, filtering, ranking and recommendation systems working together across different buyer experiences.
This is my reading, as a seller, of public eBay material, checked against what I notice from selling specialist GB stamps through Apollo’s Lots. It isn’t inside knowledge, and it certainly isn’t a formula.
First, what actually is Cassini?
The most useful description I’ve found is also the simplest.
In The Architecture of eBay Search, Cassini is described as the eBay search engine.
That wording is useful.
It points towards search infrastructure rather than the entire marketplace.
The paper explains why eBay search is a difficult problem. A normal web search engine deals with a large collection of pages that changes, but not usually in the same way as an online marketplace. eBay inventory changes constantly. Items sell. Auctions end. Prices change. Seller information changes. Buyer context changes. Results may differ depending on buyer location, seller location, postage, item status, price and many other factors.
That alone should make us suspicious of overly simple seller advice.
“Put the right keywords in the title and Cassini will reward you” isn’t completely wrong.
But it’s painfully incomplete.
The search architecture paper also mentions query transformation, active and completed item collections, aggregators, query nodes, ranking, item features, seller features and buyer features. It says ranking uses many hundreds of features derived from the item, seller and buyer, and that multiple ranking functions may exist at the same time to support different search applications, query segments and experiments.
That doesn’t sound like a small black box with one fixed set of seller tips.
It sounds like industrial search.
And industrial search is messy.
Best Match and Cassini aren’t the same thing
Another distinction I’ve gradually come to appreciate is the difference between Cassini and Best Match.
My current understanding looks something like this:
- Cassini = eBay’s core search engine or search infrastructure.
- Best Match = the default search-ranking experience buyers usually see.
In eBay’s article Using Behavioral Data to Improve Search, Best Match is described as the default sort order since 2006. The article explains that it attempts to balance several buyer objectives, including relevance, value, timely auctions and reliable sellers.
That tells us something important.
Best Match was never simply “the listing with the most keywords.”
It is trying to balance different objectives.
The same article discusses behavioural data such as clicks, bids and purchases, together with the problem of position bias. Listings near the top naturally receive more attention simply because they are near the top.
That’s a useful reminder.
If eBay uses behaviour, it can’t simply use raw behaviour.
A listing that gets clicked because it already ranked highly isn’t necessarily the same as a listing that gets clicked because it was genuinely the best result. Behaviour contains useful information, but it also contains noise.
Clicks, bids and purchases may tell the system something.
They almost certainly don’t tell it one simple thing.
Search itself appears to have layers
Even if we look only at search, there seem to be more layers than most seller guides suggest.
In Finding Desirable Items in eBay Search by a Deep Dive into Skipped Items, eBay discusses title intent, query intent and machine learning within core search ranking. One important signal is the intent expressed by the item title itself.
That doesn’t surprise me.
When I sell stamps, the title has to do a lot of work.
A collector searching for a specific SG number, watermark variety, plate number, official overprint or phosphor error isn’t looking for creative copywriting. They need clear identifiers and enough information to recognise that the listing may be relevant.
This is one reason I still think search matters.
I explored that in Search Is Not Dead, But Recommendations Are Taking Over.
For specialist material, search often remains the natural starting point.
But increasingly, it seems to be only the beginning of the buyer journey.
Search can also be personalised
Another eBay article, eBay Makes Search More Efficient Through Personalization, describes Best Match as incorporating quality, demand and market-driven factors, including item popularity, pricing, shipping, regions and seller details.
It also discusses personalising search results based on a buyer’s price preferences.
That’s interesting because it suggests search results are not necessarily one universal list shown to everyone.
Two buyers searching for the same broad thing may not be identical buyers.
One may favour cheaper material.
One may favour better examples.
One may be more likely to buy auctions.
Another may prefer fixed-price listings.
One may simply be browsing.
Another may be trying to complete a specialised collection.
That changes how I think about stamp listings.
There may not be one perfect listing structure for every buyer. There may simply be clearer listings, better descriptions, stronger images, sensible pricing and inventory that naturally relates together.
Then we leave the search results page
A buyer searches.
They click.
They arrive on an item page.
Now what?
They may see similar items.
They may see promoted items.
They may see complementary recommendations.
They may see the seller’s other items.
They may go back to search.
They may browse the shop.
They may watch the item.
They may add something to their basket.
They may return days or weeks later.
This is no longer just a search-ranking problem.
It becomes a discovery problem.
In Complementary Item Recommendations at eBay Scale, eBay discusses recommendation systems using behavioural signals such as co-purchase, co-view, co-search and popularity alongside content-based information like titles.
That feels very different from the familiar advice to simply “put more keywords in the title for Cassini.”
It suggests eBay is not only trying to understand what an individual listing says.
It is also trying to understand how listings relate to one another.
Viewed together.
Searched together.
Bought together.
Considered together.
That’s exactly the sort of behaviour I keep noticing while selling collectable GB stamps.
Collectors rarely buy one completely isolated item.
They browse related material.
They compare condition.
They compare prices.
They move between reigns, watermarks, phosphors, varieties, overprints, shades and postmarks.
A buyer who starts with one Machin variety may not stop there.
A buyer who lands on an official overprint may become interested in another.
A buyer who first discovers a £1 stamp may later purchase several quite different items.
That was one of the ideas behind Why a £1 Stamp Can Be More Valuable Than a £50 Stamp.
Not because the £1 stamp is literally worth more.
But because it may become the beginning of a journey.
The Similarity Engine is where things become really interesting
The public eBay article that probably changed my thinking more than any other was eBay’s Blazingly Fast Billion-Scale Vector Similarity Engine.
It describes a vector similarity engine capable of item-to-item similarity using text and images, together with personalised user-to-item recommendations based on previous behaviour.
The same article mentions production indexes for item-to-item, query-to-item, query-to-query and user-to-item recommendations across both text and images.
That feels like a significant clue.
It suggests discovery is no longer only about matching search terms with listing titles.
It is also about representing items, queries, images and users in ways that allow meaningful relationships to be identified.
I tried to think through some of those ideas in Vectors, Similarity and Why Two Unrelated Listings End Up Together.
Once you begin thinking in terms of similarity rather than simply keywords, some otherwise puzzling marketplace behaviour becomes easier to imagine.
Two listings may be related because they share words.
Or because they share attributes.
Or because buyers interact with them during the same session.
Or because the images themselves are visually similar.
Or because they sit close together inside a learned representation of marketplace behaviour.
In that world, “similar” doesn’t always mean identical.
And “related” doesn’t always mean obvious to the seller.
Visual discovery blurs the boundary even further
Another useful article is How eBay’s New Search Feature Was Inspired by Window Shopping.
It explicitly mentions Cassini while discussing visual discovery.
The article explains that seller metadata — titles, descriptions, images, attributes and interpreted data — are extracted and indexed for Cassini.
It then moves into visual discovery, embeddings, nearest-neighbour search and helping buyers explore inventory when words alone are no longer enough.
This is where the boundaries become less obvious.
Is visual discovery part of Cassini?
Built alongside it?
Sharing the same infrastructure?
I honestly don’t know.
The public article suggests these systems interact.
It doesn’t suggest they are all simply “Cassini.”
That distinction matters because it changes the way I think about marketplace visibility.
Rather than one algorithm controlling everything, I increasingly picture several specialised systems working together, sometimes sharing information, sometimes serving different purposes.
Recommendations aren’t just decoration
Another reason I think this matters is that recommendation placements no longer feel like decorative extras.
Older e-commerce thinking often looked something like this.
A buyer searches.
A buyer clicks.
A buyer buys.
Recommendations sit somewhere lower down the page.
Maybe someone clicks them.
Maybe they don’t.
The more I observe eBay, the less I think that’s an accurate picture.
Recommendations appear to be part of the discovery system itself.
eBay’s article Evolving Recommendations: A Personalized User-Based Ranking Model describes recommendation rankers balancing user shopping behaviour with wider platform objectives.
Another article, Improving Shopping Recommendations for Customers Through eBay’s Relevance Cascade Model, discusses improving recommendation quality rather than simply favouring popularity.
That fits remarkably well with what I’ve observed while selling stamps.
Related inventory appears to matter.
Thematic density appears to matter.
Collector behaviour appears to matter.
I explored that idea in Thematic Density.
The basic thought is straightforward.
One isolated listing can certainly sell.
But a collection of related listings gives buyers more opportunities to continue exploring.
A collector may arrive through one listing and gradually discover several others.
That doesn’t require eBay to be “rewarding” the seller.
It may simply mean there is more relevant inventory available for the system to present.
Where auctions may fit
This is also why I’ve started thinking differently about auctions.
In Why eBay Auctions Are About More Than The Final Price, I tried to separate the obvious outcome from the less visible ones.
The obvious outcome is simple.
The auction sold for £x.
Good or bad.
But around that final price may sit views, watchers, bids, comparisons, repeat visits and browsing of other inventory.
Does that directly improve the visibility of other listings?
I can’t say that.
But eBay’s own search engineering articles discuss clicks, bids and purchases as behavioural data used to improve search experiences.
That makes me think of auctions less as pure price-discovery mechanisms and more as behavioural events occurring within a wider marketplace.
A bid isn’t only a bid.
It may also represent interest.
A watcher isn’t only a watcher.
They may simply be part of a longer buyer journey.
A losing bidder isn’t necessarily a lost buyer.
They may simply continue their journey somewhere else.
Again, that isn’t proof.
It’s simply a more useful way for me to think about what I’m observing.
What I think can reasonably be inferred
If I had to reduce everything in this article to a few careful observations, they would be these.
- Cassini appears to be eBay’s search engine or search infrastructure.
- Best Match appears to be the default search-ranking experience.
- Search itself appears to use many signals rather than simply title keywords.
- Search can be personalised, so different buyers may not always see identical results.
- eBay publicly describes recommendation systems, vector similarity, visual discovery and personalised recommendation ranking.
- Those systems appear to interact with search while also serving different purposes.
- Marketplace discovery therefore seems better understood as a system of systems rather than one algorithm.
That’s about as far as I’m prepared to go.
I wouldn’t claim:
- watchers are a direct ranking factor;
- auctions boost an entire shop;
- one sale creates a Green Wave;
- low-value items improve all-store visibility;
- thematic density is a confirmed ranking factor;
- there’s a fixed Cassini formula waiting to be discovered.
Those ideas may contain fragments of truth.
But I can’t demonstrate them from public information alone.
What this changes for me as a seller
The practical conclusion isn’t especially complicated.
I’ve largely stopped thinking about Cassini as though it’s a small god hiding behind the search box.
Instead, I try to think about buyer discovery across multiple systems.
That means concentrating on things I can actually influence:
- clear titles;
- accurate item specifics;
- good photographs;
- honest condition descriptions;
- sensible pricing;
- reasonable postage;
- related inventory;
- buyer confidence;
- stock groupings that make sense to collectors;
- listings that help buyers make decisions rather than forcing them to guess.
That’s essentially where I eventually arrived in Beyond SEO: Visibility, Discovery and Selling on eBay.
The more I read, and the more I observe, the less useful the phrase “eBay SEO” feels.
Not because search no longer matters.
It clearly does.
But because search increasingly appears to be only one route into marketplace visibility.
A listing may be found through search.
It may be compared through recommendations.
It may be reinforced by related inventory.
It may be trusted because the seller describes it clearly.
It may be clicked because the photographs inspire confidence.
It may be purchased because the buyer feels comfortable.
It may even lead to another purchase because the surrounding stock is relevant.
None of those things feels like a single ranking factor.
Together they begin to resemble a buyer journey.
Final thought
I started this series thinking about visibility.
The further I looked, the more I found myself thinking about relationships.
Relationships between search terms and titles.
Relationships between buyers and items.
Relationships between one listing and another.
Relationships between images, attributes, behaviour and recommendations.
Relationships between an initial click and a later purchase.
Perhaps that’s now the more useful way to think about eBay.
Not Cassini as a single answer.
Not recommendations as an afterthought.
Not buyers following one straight line from search to purchase.
But a marketplace made up of interconnected systems, each trying to understand what belongs together, who may find it useful, and which path is most likely to help a buyer reach a successful outcome.
I can’t know exactly how eBay does that.
But after reading the public engineering material, and comparing it with my own observations, I no longer think the old seller folklore is a big enough explanation.
Cassini is part of the picture.
Best Match is part of the picture.
Search is part of the picture.
But the rest of eBay discovery deserves just as much attention.
— Dan Williams
Continue Reading
Previous: Why eBay Auctions Are About More Than The Final Price
Next: Search, Recommendations and the Successful Buyer Session
Read More
This note forms part of Notes on eBay & Discovery: a collection of observations about visibility, recommendations and buyer behaviour on modern marketplaces.