Better than we expected on the basics. All 50 products existed and every one was within budget at the price we actually found. Most stated prices were exactly right.
Availability was the weak spot. One in five "you can buy this here" claims did not hold up: the product was sold out, on back-order, or not sold by the shop the AI named. Two more we could not check because the shop's page would not load. Two answers said every item was "verified", and one of them had two items unavailable online.
Why we ran this
There is a big difference between:
a compact travel coffee setup could fit his commute
and:
Product X is £59 and in stock in the UK.
The first is an idea. The second contains several facts that can be checked, and that can go out of date by the afternoon. AI can help with both, but the second deserves a much higher bar.
Gift answers are especially easy to over-trust. An exact model name, a retailer, a price to the penny and "under your £75 budget" all feel like evidence. They are not, until someone checks.
This was a shopping-accuracy test of a general AI model with live web search, not a test of the Giftin app.
The test
We gave 10 of our fictional people to the AI model with live web search switched on, once each, with a request like this one (each person's occasion and budget filled in):
Recommend exactly five currently purchasable birthday gifts for this recipient. Budget is a hard maximum of £70 per item. Only include products currently sold in the UK. Give the exact product name, retailer, current price and source for each. Do not include anything you cannot verify.
That gave 50 products. The same afternoon, about two hours after the answers were saved, we checked every one by hand in a private browser window, logged out, with no coupons. The rules for what counted as "correct" were written down before the first check.
Did the products exist?
Yes, all 50. Two had been renamed slightly since (a massage gun now sold as "Active +", gardening gloves that dropped "Ladies" from the name), and we counted those as real.
This is a fair result to report plainly: in these 50 recommendations, we found no invented product, model or brand.
Were the prices right?
Stated price vs the price we found
- Exact match39 of 45
- Different6 of 45
Of the six that differed, four were cheaper than stated, mostly because the retailer was showing a sale price or a "from" price when we checked. Two were dearer: swimming goggles stated at £19.99 were £24.50, and a board game stated at £19.29 was £20.85 and on back-order.
For the five products we could not check at the named shop, the price elsewhere was further off. A coffee grinder stated at £35 was £42; the £35 turned out to be a price before VAT.
Did they stay under budget?
All 50 did, at the price we actually found. The budget was a hard limit in every prompt, and the model respected it.
Could you actually buy them?
This is where the answers fell short.
Products by what we found on the day
- Buyable at the shop the AI named38
- Shop's page would not load, in stock elsewhere2
- Not sold by the shop the AI named3
- Sold out, back-order or unavailable online7
Seven products were not available at all: three sold out with only a "notify me" button, one on back-order, one marked sold out despite an add-to-cart button on the page, and two "temporarily unavailable online". Three more were real and in stock, but not at the shop the AI named. For the last two, the named shop's page would not load for us, so we count those as unchecked rather than wrong.
The wording could still sound completely certain. One answer ended with "Every item is verified to be in stock". Two of those five items were unavailable online when we checked.
The links: often just a homepage
We asked for a source for every product. 20 of the 50 links pointed to the shop's homepage rather than the product, and 3 were temporary redirect links that had already stopped working within hours. A homepage link looks like a citation but proves nothing about the product.
A fit problem the facts did not catch
Accuracy and fit are different checks. One partner's brief said he "finds most aftershaves too strong". Among his five verified, in-stock, within-budget products was a solid cologne. Every fact about it was correct. The gift was still wrong.
What to do with this
Use AI to narrow the direction. Use the shop to confirm the facts.
- Ask AI for gift directions first, based on the person.
- If you want exact products, use live search. In this test, product identity and budget fit held up better than stock claims.
- Open the actual product page before deciding. Stock is the claim most likely to be wrong, and a homepage link is not a source.
- Read each suggestion against what you told it. A verified product can still be the wrong gift.
The checkout page still gets the final vote.
Method
- Tested
- 5 October 2026, answers captured at 14:45 and every check finished by 16:46 (Tallinn time, UTC+3)
- Model
- Gemini 3.5 Flash via the Gemini API, with Google Search grounding on
- Market
- UK only, prices in GBP
- Prompts
- 10 fictional recipients, one run each, five products per answer, 50 products
- Checking
- By hand, private window, logged out, no coupons or memberships, rules fixed before checking
- Price rule
- Visible selling price for that product at the named shop; "from" and sale prices recorded with a note
- Availability rule
- Supported only if the named shop's UK product page let us add it to the basket
Raw data
Free to reuse under CC BY 4.0: credit Giftin.ai and link to this page.