We tested 7 AI gift finders. Most just gave us what Google would.

GiftyWow was the only tool that stayed in the top two across all three scenarios, with a score range of just 1.5 points (72.1 to 73.6%). That kind of consistency didn't appear anywhere else in the test.

The standout quality was discovery depth. Across all three scenarios, between 88% and 94% of GiftyWow's passing recommendations were things you wouldn't find through an obvious Google search. Two gifts in particular showed what photo-based dual-person matching can do that other approaches can't. For Nico, a hickory wood fly fishing box connected his smoking-wood aesthetic with his fishing hobby, a cross-interest match no other tool found. For Hemi, a sunprint cyanotype kit connected the giver's photography interest with the nephew's love of outdoor play, a gift that works because of who's giving it, not just who's receiving it.

The weaknesses were minor. Two recommendations in the Hemi scenario were wrong for the context, and a handful of picks across scenarios were solid but not surprising. The overall sets were rated exceptional in two scenarios and strong in the third.

For GiftyWow, these results reflect only the AI-generated recommendations. The full product experience starts with picks from a pool of over 10,000 gifts that real users have already liked, then layers in the AI-matched ideas. We excluded the pool to test the matching engine alone. To learn more about how GiftyWow works, read about us here.

ChatGPT performed best on the anniversary scenario (71.7% overall score, within 2 points of GiftyWow) and weakest on the hobbyist scenario (52.5–54.7% depending on whether it received a photo). The 19-point swing between its strongest and weakest scenario is worth noting: ChatGPT is excellent when the gift category maps well to its training data, but less reliable for multi-hobby recipients where specificity matters.

We tested ChatGPT with and without a photo for the Nico scenario. The photo helped: pass rate jumped from 50% to 75%. But even with the photo, ChatGPT produced zero surprising finds. Every recommendation was something you'd see on page one of a Google search for "gifts for a guy who fishes and homebrews." The generic rate tells the story: between 16% and 64% of ChatGPT's passing recommendations were things you could have found yourself, compared to GiftyWow's 6–12%.

ChatGPT's strongest quality is thoughtful reasoning. It explains why each gift fits, and many of its recommendations feel considered even when they're not surprising. Its weakest quality is discovery. If you already know roughly what you want and need help refining the idea, ChatGPT is a solid free tool. If you're hoping to be surprised, look elsewhere.

Gemini produced the highest single-scenario overall score in the entire test: 74.6% on the anniversary, edging GiftyWow by 1.6 points. It also achieved zero wrong suggestions in two of three scenarios, which no other tool managed.

The important caveat: several of Gemini's anniversary recommendations couldn't be verified as purchasable products. The overall score doesn't currently penalize hallucinated recommendations (products that sound real but don't exist or can't be found at the stated price). If hallucination verification were part of the scoring, Gemini's anniversary result would likely drop. This is a known limitation of the current evaluation framework, and one we plan to address (see "What's missing" below).

Gemini's performance also varied significantly depending on which model version we used. For the Nico scenario, the Flash model scored 58.8% while the Pro model scored just 32.2%. If you use Gemini for gift ideas, the model matters.

There's also a practical trade-off between Gemini's models: Flash produced stronger recommendations but doesn't include product links, while Pro includes Google Shopping links but scored nearly half as well (32.2% vs 58.8% on the Nico scenario). Using Gemini for gift finding means choosing between better ideas or easier purchasing.

The generic rate ranged from 8% to 22%, sitting between GiftyWow's consistently low 6–12% and ChatGPT's higher 16–64%. Gemini finds more non-obvious gifts than ChatGPT but fewer than GiftyWow, with the added risk that some of those "non-obvious" finds may not be real products.

GiftList Genie tells the most interesting consistency story in the test. For the standard hobbyist scenario (Nico), it performed well: 58.4% overall score, zero wrong suggestions, and practical picks across all four hobbies. But the score dropped 15 points for the Hemi scenario (43.9%), where the unusual relationship dynamic (aunt and 7-year-old nephew at Easter) sat outside the tool's sweet spot.

The weakness is depth. A third of GiftList Genie's passing recommendations for Nico were generic (things like a portable bike floor pump or a digital meat thermometer that a 47-year-old hobbyist almost certainly owns already). The anniversary result was stronger on generic rate (only 8%) but weaker on overall score.

GiftList Genie doubles as a wishlist platform, though GiftyWow also captures wishlists, tracks occasions over time, remembers what's already been recommended so you're not starting fresh each time, and lets you flag gifts as bought to keep a running memory of what you've given. If you're buying for an adult with clearly defined hobbies, GiftList Genie is a reasonable option. For more complex gift-finding situations (kids, unusual occasions, recipients who "already have everything"), the results suggest looking elsewhere.

GiftX's quiz format produced the most dramatic failure in the entire test. For the anniversary scenario ("eclectic taste, into design, already has everything, something he'd never find himself"), GiftX scored 7.6% overall with zero passing recommendations out of 20. Not a single suggestion was worth considering.

The Nico and Hemi scenarios were marginally better but still weak. The Hemi results included a $120 Magna-Tiles set on a $50 budget, the kind of basic constraint violation that erodes trust. The generic rate across passing scenarios sat at 36–38%, meaning more than a third of the adequate recommendations were things you'd find through obvious searches anyway.

The core problem is structural. A quiz captures age, interests, and budget, but it can't capture the subtler signals that determine whether a gift feels "right." When the recipient doesn't fit neatly into standard hobby categories (a design-oriented person with eclectic taste, for example), the quiz has no way to adapt. The 32-point swing between GiftX's strongest and weakest scenarios (39.4% to 7.6%) illustrates this limitation.

Uncommon Goods is a curated retail store, not an AI gift finder in the traditional sense. We included it for two reasons: it appears in many "best AI gift finder" articles that searchers encounter, and it's a genuinely interesting test case. Uncommon Goods sells products specifically designed to be surprising, distinctive, and delightful. If any retail catalog should score well on discovery, it's this one. The results show that even a catalog full of distinctive products can fail when the matching is poor.

The core limitation is inventory: Uncommon Goods can only recommend products it sells, which constrains matching to a relatively narrow range. For the Hemi scenario, this meant 8 out of 20 recommendations were flat-out wrong for a 7-year-old, the harshest single result across any tool in any scenario. The catalog skews toward adults, and the recommendation filter couldn't compensate.

When the catalog happens to match the recipient, the individual products can be interesting (the 0% generic rate for Nico and Hemi reflects genuinely distinctive artisan goods). But low generic rate means nothing when most of the recommendations miss on basic fit.

TheGiftTool was the most consistently mediocre performer. No catastrophic failures like GiftX's anniversary result, but no strong showings either. Overall scores ranged from 35.6% to 43.6% across the three scenarios, never approaching the competitive tier.

The Nico scenario exposed the generic problem sharply: 60% of passing recommendations were things you'd find on a basic Google search. Five suggestions in the Hemi scenario were wrong for a 7-year-old, including items from entirely wrong age categories.

The quiz format is quick (a few questions, instant results), but the speed comes at the cost of depth. Without richer input about the recipient, the tool defaults to category-level matches rather than genuinely personalized recommendations.