Tool · · Peter Hanlon
US Vendor Similarity
Not who bought here, also bought there.
US Vendor Similarity is a 3D view of which US brands have audiences that look like each other. Type a vendor, the landscape zooms to it, and the neighbours are brands whose audiences have the same shape, whether or not the same people shopped at both.
Type a vendor into the box. The view zooms to that vendor and shows the vendors whose audiences sit closest to it.
To support the US launch next quarter I've been integrating US audience design into our where+when™ platform. It's been a mixed bag, to be honest.
Deep sample, heavy volumes
On the one hand the sample is deep and the quality is excellent. Really good. On the other, the sample is deep, which makes the volumes quite challenging. Processing, storage, and query costs all need to be managed, even during testing. Challenges are good though, and I've worked hard on this to get what I think is excellent audience definition.
Once the ingestion and audience creation pipeline were hanging together, I did a load of calibration on skew, age, gender, geo, and payment type. I'm very happy with the outcome. After I pushed the final commit and merged, I built about twenty audiences to test all of those features.
Picking the test set
I'm not a US native, so I asked ChatGPT for differentiated retailers acfross a number of dimensions. It picked out half a dozen across different verticals and scale, half a dozen geographically differentiated, half a dozen demographically differentiated, and a few standout brands. I was happy with all of them. Really happy.
Similarity Landscape
I pushed on into the evening, which is where I am now, documenting it after the fact, and thought I'd try to release the Similarity Landscape.
For telco data this is domain similarity. For transaction data it is vendor similarity. The earlier audience similarity piece is the domain version. This is the vendor one.
This is not overlap. It is not who browsed / bought here and then also browsed /bought here. That would be affinity. This is what brand X's audience looks like, and which other brands have an audience that looks like that, regardless of whether the same people show up in both.
The metric is cosine similarity.
We use this on audience builds that need to jump off a seed — a domain on telco data, a vendor on transaction data — and expand the original audience into brands that look like it.
Why it's in 3D
From a coding perspective it's always nice to play around in 3D, and you'll know from the King's Cross simulation and King of the Hill that I'm hugely into 3D, rich, immersive experiences. I thought I'd share this. I was quite pleased with it.
Related in the lab
Questions
What is US vendor similarity?
Cosine similarity between vendor audiences built from US transaction data. It scores how alike two brands' audiences look, without requiring the same people to have shopped at both.
How is vendor similarity different from affinity?
Affinity is overlap: people who bought here also bought there. Vendor similarity ignores that shared membership and compares the shape of each audience.
What is the Similarity Landscape used for?
Audience builds that start from a seed vendor and need to expand into other brands whose audiences look like the seed, whether or not those shoppers overlap.