AI Gift Finder and Product Recommendation API for Ecommerce
A gift buyer can describe a person but not a product. TasteRay ranks your catalog against that description. A generated sentence on your page is your claim, so this page also shows how to review them.
TasteRay is a product recommendation API for ecommerce, suited to gift finders and shoppers who can describe a need. You send what they said and up to ten products, and one request returns them ranked with a written reason. Show price, stock and specifications from your own data, and review generated sentences before they go live.
"He has just retired, cooks every weekend and owns too many gadgets" is a clear brief for a shop assistant and useless to a search box. A filter cannot read it, and past orders belong to the buyer, not the recipient. That is where a language model helps a store. It also brings a new risk, which we cover below with evidence.
A birthday, a retired cook, a budget of about $80
We wrote ten generic kitchen and electronics products with store prices, then sent the buyer's description. The smart thermometer, milk frother and fitness tracker did not come back after the buyer said "not another gadget". The $89 air fryer and the $120 course voucher, both above budget, did not come back either.
This is a separate capture from the one on our stores page, made on the same day to show the gift case.
{
"vertical": "gifts",
"context": {
"profile": "Buying a birthday present for my dad. He has just retired, cooks every weekend, and already owns too many gadgets. Budget about 80 USD.",
"preferences": [
"useful",
"will last",
"not another gadget"
]
},
"items": [
{"id": "k5", "name": "Chef's knife, 20 cm", "description": "Forged stainless steel knife with a riveted handle. Store price 75 USD.", "metadata": {"category": "kitchen", "price_usd": 75}},
{"id": "k3", "name": "Smart meat thermometer", "description": "Wireless probe with a phone app. Store price 79 USD.", "metadata": {"category": "gadget", "price_usd": 79}},
// ...and 8 more: an air fryer, a fitness tracker, a cooking course voucher and others
],
"options": {"count": 3}
}meta.processing_time_ms 1,714. The model's sentences are verbatim.Chef's knife, 20 cm
98 out of 100, taste_match_score
You are buying him the one tool that defines a cook's confidence. It is not a gadget, it is a craft.
- He cooks every weekend. A forged steel knife is the fundamental instrument of that ritual. It connects him to the food, not to an app.
- At $75, it fits your budget perfectly and signals respect. It says: 'I see what you do, and I value the quality of your craft.'
Cast-iron skillet, 26 cm
92 out of 100, taste_match_score
Cast iron is the antidote to disposable culture. It is a heirloom in the making, perfect for a new retirement chapter.
- At $45, it leaves budget for a nice bottle of wine or a dinner out to celebrate the birthday. It is practical, not just decorative.
Beech chopping board, large
85 out of 100, taste_match_score
It turns his kitchen workspace into a stage. A large, solid wood board is a sign of a serious cook who prepares with care.
- It is strictly 'not a gadget'. It is 100% wood and function. No buttons, no screens.
Fields shown: item.name, metadata.taste_match_score, explanation.why_match, explanation.key_factors.
What we rejected in a branded category
We also ran a headphones request with real brands for a frequent flyer, and read all twelve key factors and three reasons as a shop's legal reviewer would. This is the capture excerpted on the stores page. Most of it we would not publish on a product page, and the problems fell into three groups.
- Traceable to our list: fine
- Prices and the sum left from the budget were right, because we sent them. Anything arithmetic on your own fields can be checked automatically.
- Opinion: your call
- Statements that a pair is an industry benchmark, or feels more comfortable than a rival, are judgement. They may be fair, and your brand would be saying them, so decide with the brands you stock in mind.
- Invented or unsupported: remove
- The budget pair was credited with a percentage of the silence, a figure nobody measured. One pair was said to be engineered specifically for long-haul travel, another promised for hours of pain-free wear, and an over-ear pair described as fitting in a pocket. We publish none of these.
What we would do before launch
Keep the facts in your own fields
Show price, stock, size and specifications from your product database, and never parse them out of generated text.
Keep product names plain where you can
The generic names in our gift test gave the model little to embellish. With famous brands it reached for what it remembered. One comparison is thin evidence, so try this on your own catalog.
Read a sample per category
Have someone read about fifty responses per category and note each sentence that makes a checkable claim. Categories with regulated or safety claims deserve the closest look, or no generated text at all.
Start with short text
Set
options.explanation_depthtobrief, or showwhy_matchonly, then widen it once the sample reads clean.Use the caveats
potential_concernscarries the response's own caveats. A shopper trusts a reason more when the downside sits next to it.
What changes in a store integration
- Preselect by hard facts
- Filter to in stock, shippable and in budget, then send ten. The API only returns items from the list you send, so it will not suggest what you cannot sell.
- Whose taste?
- A buyer's orders say little about the recipient. Leave account history out of a gift request and describe the recipient in
context.profile. - Margin and promotion
- The API ranks on fit and knows nothing about margin. If you must weigh commercial factors, do it when you choose the candidates, and keep the reason honest about why an item appears.
- Timing
- Plan for a few seconds, with outliers, returned in
meta.processing_time_ms. A gift finder can show a short wait after a question. Category tiles on every page view should stay on your existing recommender.
ForMe, or other vendors
For a shopper already on one product, ForMe needs one script tag and no server call. For "frequently bought together" on every page, compare Algolia Recommend or Recombee. Larger retailers can start a pilot on one category through Enterprise.
Sources and further reading
FAQs about AI Gift Finder and Product Recommendation API for Ecommerce
How does an AI gift finder work with TasteRay?
The buyer describes the recipient in their own words, and your server turns that into a request with up to ten in-stock products in budget. TasteRay returns them ranked, each with a written reason the buyer can repeat. Leave the buyer's own purchase history out, since it says little about the recipient.
Can the model make false claims about my products?
Yes, and you should plan for it. In a headphones capture on 5 November 2026 the model invented a percentage and made claims about comfort and design that we would not publish. Show specifications from your own data, review a sample per category, and consider shorter text or no text for regulated goods.
Do generated reasons need review if I send product data myself?
Yes, though good data helps. Prices we sent came back correct, and the sums based on them were right. The model can still add what it remembers about a known brand, so a sample read by a person is worth doing per category before launch, and again whenever your catalog changes materially.