Why most gift recommendation engines are bad — and what would have to change for them to be good

Type "gift for sister, $50, birthday" into any of the current gift-recommendation tools and you will get the same six suggestions: a candle, a pair of slippers, a generic skincare set, a wine accessory, a book about cocktails, and a tote bag with a phrase on it. The exact six rotate based on the season and the affiliate margin. The character of the six does not change.

This is not because AI is bad at recommending things. It is because the entire category of gift-recommendation tools is built on a model of gift-giving that does not match how gift-giving actually works. The result is technically functional, structurally hollow, and obviously useless to anyone who has ever bought a gift they were genuinely proud of.

I have been thinking carefully about this problem for the past year — for reasons I'll get to. This piece is the diagnosis. What's wrong with these tools, why they're wrong in the specific ways they are, and what the tool that doesn't have these problems would have to look like.

The first failure: treating all gifts as the same kind of decision

The deepest flaw in current gift recommendation tools is that they treat "what should I get for someone" as a single problem with a single answer. It is not.

There are at least three completely different gift-giving situations, and they call for incompatible reasoning:

• The gift you give your spouse for your tenth anniversary. The constraints are emotional resonance, personal history, and the sense that this person will reread the gift's meaning for years. The wrong gift is one that's too safe — that says "I didn't think very hard about this."

• The gift you give a coworker for the office Secret Santa. The constraints are inverse. The wrong gift is one that's too personal — that says "I read your social media too much." Safety, professionalism, and broad appeal are the actual targets.

• The gift you give your in-laws for a holiday, when you've already given them a meaningful gift last year and you're now in the middle stretch of a long relationship. The constraints are subtler still — neither too bold nor too safe, calibrated to a relationship that runs across decades.

The same recipient — say, your sister — does not get the same kind of gift on her birthday, on the day after her divorce, and on the year she becomes a grandmother. Same person. Different gift-giving situations. Different optimal answers.

Current gift-recommendation tools collapse this entire spectrum into one ranking problem. They are, structurally, an Amazon search with a chat interface bolted on. The output looks personalized; the reasoning is not.

The second failure: no model of the relationship

Most gift tools ask you for the recipient's age, gender, interests, and budget. They do not ask the question that matters most: what is the recipient to you?

A "30-year-old woman who likes cooking, $75 budget" gets the same five suggestions whether she is your wife, your sister-in-law, your coworker, or your dental hygienist. This is absurd. The relationship is the largest single piece of information about what an appropriate gift looks like.

The relationship determines:

• How personal the gift can be without crossing a line
• How much editorial confidence the gift can express (a gift to a spouse can be a bold bet; a gift to an acquaintance should not)
• How much money is appropriate to spend
• Whether the gift can reference shared history
• Whether the gift should be experiential or object

None of this is captured by demographics. All of it is critical. Current tools collect the demographics and skip the part that matters.

The third failure: no model of the giver

Even if you fix the relationship problem, there's a deeper one: gift-recommendation tools have no idea who you are.

Different givers have systematically different styles. Some people are utility-first gifters — they think the right gift is the useful one. Others are sentimentally-first gifters — they think the right gift is the one that signals "I see you." Others lean novelty (gifts that surprise), others indulgence (gifts the recipient wouldn't buy themselves), others identity (gifts that affirm who the recipient is).

There is no universal right answer here. There are styles. A utility-first gifter giving a sentimentally-loaded gift will get it wrong. A sentimentally-first gifter giving a utility gift will be disappointed in themselves.

A good gift-recommendation tool would learn the giver's style over time and produce recommendations that fit that giver's sensibility, not a generic "good gift" template. The current tools do nothing like this. They produce the same output whether you're someone who has handmade every gift you've ever given or someone who has bought everything in their life on Amazon Prime.

The fourth failure: optimizing for click, not satisfaction

Here is the structural reason all of this persists.

Current gift-recommendation tools are funded by affiliate revenue. Affiliate revenue is earned on the click and the purchase. It is not earned on whether the recipient actually loved the gift.

The result is predictable. Tools optimize for "you clicked this and bought it" rather than "this turned out to be the right gift." The metric the tool can measure (click-through, conversion) is uncorrelated with the metric the user actually cares about (satisfaction with the gift). This is a textbook alignment problem — the proxy and the goal are not the same thing.

It explains the candle-slippers-skincare-set output pattern perfectly. Those six suggestions are the local maximum on the "things people will impulse-buy" gradient. They are not the local maximum on the "things people are glad they gave" gradient. The tools know how to find the first. They have no way to measure the second.

What the tool that doesn't have these problems would look like

The good news is that all four failures are addressable. None requires a model breakthrough. They require a different framing of what the tool is trying to do.

It would model the gift situation, not just the recipient. It would ask about the relationship, the occasion, the history, and the social context — and use those to shape what kind of gift is appropriate, not just what objects to suggest.

It would model the giver. Either by asking up front, or — better — by learning from what the giver actually chose in past situations. A giver who consistently picks consumables over objects, sentimental over utility, indulgent over practical, has a style. The tool should reflect it.

It would be calibrated. When the tool is confident, it should say so. When the tool isn't sure — when the recipient profile is sparse, when the relationship is ambiguous, when the occasion is unusual — it should also say so, and either ask for more information or present a wider range of options instead of bluffing.

It would close the loop. It would measure satisfaction, not clicks. It would ask the giver, after the gift was received, whether the recipient actually loved it — and adjust future recommendations based on what worked and what didn't. This is harder to build than click tracking. It is the only metric that matters.

A tool with these properties would still produce some bad recommendations. No tool produces perfect ones. But the failure modes would be qualitatively different — closer to "wrong but defensibly chosen" and farther from "the seventh candle in this category this month."

Why I care about this

I am working on this problem. Quietly, with a small team, with the kind of patient pacing that befits a category that has been done badly by everyone else and that deserves to be done well.

I am not going to say more in this piece, because the work isn't done and because I have a personal commitment to not making promises I cannot yet keep. But Depth Protocol readers who are interested in this category will see more here in the coming months — both from me and from the projects I cover.

In the meantime: the diagnosis above is the diagnosis. If you are evaluating a gift recommendation tool, or a gift-giving service, or a personalized shopping experience of any kind, the questions to ask are: Does it understand the situation, not just the recipient? Does it understand me as the giver, or treat me as anonymous? Does it know what it doesn't know? Does it measure whether the gift actually worked?

If the answer to all four is no, you are looking at an Amazon search with a chat interface bolted on. There are better tools coming.