Published: 05 June 2026 by Dan Williams
Series: Notes on eBay & Discovery
Observations on recommendation systems, similarity and why marketplaces increasingly connect items in unexpected ways.
Strange Things Start Appearing Together
One of the interesting things I’ve noticed while selling on eBay is that listings sometimes appear to become connected in ways that are not immediately obvious.
A buyer could view one item.
Then a seemingly unrelated item appears in a recommendation module.
A sale occurs.
Then another item from a similar area sells shortly afterwards – something I explored previously in The Green Wave.
At first glance, these events can appear random.
However, the more I read about recommendation systems and the more I think about buyer behaviour, the less random they seem.
Similar Does Not Always Mean Identical
Traditionally, search engines worked largely by matching words.
If a buyer searched for a particular phrase, the platform looked for listings containing those words.
That model still exists.
However, modern recommendation systems are appearing to work differently.
Rather than asking:
Do these listings contain the same words?
they often ask:
How similar are these listings?
Which sounds simple in theory.
But similarity can mean many things.
Two listings might be considered similar because:
- buyers frequently view them together
- buyers purchase them together
- they share attributes
- they belong to related categories
- they contain similar images
- they attract similar types of buyers
Similarity is often broader than exact matching.
The Problem Of Scale
This becomes difficult when a marketplace contains billions of listings.
No human could manually connect every related item.
This is one reason large marketplaces are now relying on machine learning and recommendation systems.
Rather than creating relationships manually, they attempt to learn relationships from behaviour.
If enough buyers repeatedly interact with certain groups of items, patterns begin to emerge.
Those patterns can then be used to influence future recommendations.
Enter Vectors
One of the publicly available eBay engineering articles that helped me think about this topic was:
eBay’s Blazingly Fast Billion-Scale Vector Similarity Engine
The article discusses how listings can be represented in mathematical forms known as embeddings, allowing the platform to search for similar items extremely quickly across billions of listings.
The details are highly technical but the important idea is not.
Instead of storing a listing simply as text, a recommendation system can represent it as a collection of characteristics and relationships.
Listings that share similar characteristics end up closer together.
Listings that differ significantly end up further apart.
The system then becomes capable of finding similar items even when the wording is different.
Read more at eBay:
eBay’s Blazingly Fast Billion-Scale Vector Similarity Engine
Why This Matters To Collectors
This becomes more interesting in specialist areas.
A collector looking at a watermark variety may not necessarily want another example of the exact same stamp.
They may be interested in:
- a related variety
- another issue from the same reign
- a booklet pane
- a specialist Machin
- a similar collecting area
To a collector, those relationships often feel natural.
To a recommendation system, they become patterns that can potentially be learned.
The recommendation system does not need to understand philately.
It simply needs to recognise that similar buyers repeatedly interact with similar groups of listings.
Similarity Is Not Always Obvious
One reason recommendation systems sometimes appear mysterious is that the relationships they identify are not always visible.
Two listings may seem unrelated to a seller.
Yet buyers repeatedly view them together.
Or purchase them within the same session.
Or belong to a broader collecting theme.
The relationship may exist in buyer behaviour rather than in the listings themselves.
This can create surprising recommendations.
It can also create unexpected pathways through inventory.
The Human Side
Despite all the discussion of vectors and machine learning, I think it is important not to lose sight of the human element.
The system is not inventing relationships from nowhere.
It is usually observing relationships created by buyers.
Collectors.
Browsers.
Researchers.
People comparing items.
People building collections.
The technology may be sophisticated.
The underlying behaviour is often very human.
Why I Find This Interesting
The more I observe recommendations, the more I find myself thinking less about individual listings and more about relationships between listings.
Some relationships are obvious and some are not, and some may only exist because buyers repeatedly create them through their actions.
This does not mean I understand exactly how eBay works – far from it.
But it does help explain why certain items seem to become connected over time.
And why recommendation systems sometimes appear to know that two listings belong together before a seller realises it.
Closing Thoughts
Modern marketplaces appear to work in terms of similarity rather than exact matching.
That similarity may come from:
- titles
- attributes
- images
- buyer behaviour
- recommendation history
or a combination of all of them.
The technology behind these systems is complex.
But the underlying idea is surprisingly simple.
Items do not exist in isolation.
They exist within networks of relationships.
And the better a marketplace becomes at understanding those relationships, the more interesting discovery will likely become.
— Dan Williams
Continue Reading
Previous: Thematic Density
Next: Why eBay Auctions Are About More Than The Final Price
Read More
This note forms part of Notes on eBay & Discovery: a collection of observations about visibility, recommendations and buyer behaviour on modern marketplaces.