Two OpenAI embedding models, one plugin, one very different search experience. Here is what the difference actually looks like on a store – and why the paid plans are worth the upgrade if you care about precision or sell in more than one language.
Every search starts with a “meaning fingerprint”
When AI Search indexes your store, it converts each product into an embedding – a numeric fingerprint of what the product actually is. When a user searches, their query gets its own fingerprint, and we match by meaning, not just keywords. That is why “summer blue dress” can find your “light cotton navy sundress”.
The quality of those fingerprints depends on the model that generates them. AI Search uses OpenAI’s third-generation embedding models – and the plan you are on decides which one.
The two models, in plain English
text-embedding-3-small (AI Search Free) produces fingerprints with 1,536 dimensions. Think of dimensions as the number of “traits” the model can use to describe a product. It is fast, efficient, and a big step up from keyword search – the model our Free plan runs on, and it is genuinely good.
text-embedding-3-large (AI Search Standard & Pro) produces fingerprints with 3,072 dimensions – twice the descriptive resolution. More dimensions means finer distinctions: “running shoes” vs “trail running shoes” vs “walking sneakers” occupy clearly separated spots in the model’s map of meaning, instead of blurring together.
What the benchmarks say
On OpenAI’s published benchmarks (see the announcement linked below):
- English retrieval (MTEB): small scores ~62.3%, large scores ~64.6%. A solid, consistent gain – every point here is a query that finds the right product instead of a near-miss.
- Multilingual retrieval (MIRACL): small scores ~44.0%, large scores ~54.9%. This is the headline: an improvement of roughly 10+ points in cross-language search quality.
That multilingual gap matters enormously for stores selling across Europe or in more than one language. AI Search supports 100+ languages on every plan – but the large model is dramatically better at matching a Portuguese query to an English product description, a French search to a German product name, and so on. If you run a multilingual store, this single difference is usually worth the upgrade by itself.
Sources: OpenAI’s “New embedding models and API updates” announcement and embeddings documentation – links at the end of this post.
What this looks like in real searches
Concrete examples of where large pulls ahead:
- Nuance: “something warm but breathable for hiking” – small finds jackets; large separates insulated shells from breathable fleece midlayers.
- Cross-lingual: a user types “vestido de verão às flores” on an English-catalog store – large surfaces the floral summer dresses reliably; small is hit-and-miss.
- Technical proximity: “USB-C hub with HDMI 4K 60Hz” – large distinguishes 30Hz vs 60Hz variants that small treats as near-identical.
You can try each of these on the live demo stores, which run the plugin on real catalogs – clothing, books, and electronics.
Why we do not run large on the Free plan
Honest answer: cost. Large embeddings cost roughly 6-7x more to generate than small ones, and every fingerprint is twice the size to store and compare. The Free plan exists so any store can prove that semantic search works on their own catalog – small does that well. Standard and Pro fund the heavier model and the infrastructure behind it, which is also why every plan has a clear embedding limit rather than a vague “unlimited”.
Upgrading re-indexes for you
Fingerprints from the two models are not compatible – you cannot compare a small query fingerprint against a large product fingerprint. So the first thing to do after upgrading is to regenerate the embeddings for your existing catalog on the large model. Right after you paste the activation key in AI Search → General, the plugin shows a one-click Regenerate embeddings card that runs the whole re-indexing in the background – search stays online on the small-model index until the new one is ready, so nothing on your storefront blinks.
The bottom line
- Free (
small): real semantic search, great for trying it on your own store. Up to 2,000 embeddings. - Standard (
large): the precise model, a 25,000-embedding index, full taxonomy boosting, a year of search insights. €49/year (or €5.90/month). - Pro (
large): the same precision across up to 10 sites with a 100,000-embedding pooled index and priority support. €129/year (or €14.90/month).
Launch offer: 35% off for your first 6 months, applied automatically at checkout.
References
- OpenAI, “New embedding models and API updates” – openai.com/index/new-embedding-models-and-api-updates
- OpenAI embeddings guide – platform.openai.com/docs/guides/embeddings