Which details actually improve AI gift ideas?

What we tested

"Tell the AI more about them" is easy advice. The useful follow-up is harder: which details actually change the recommendations?

An age might change the wording without changing the idea. An interest might push every suggestion into one hobby. A dislike might clear away the clichés without adding anything. So instead of assuming more is better, we tested one kind of detail at a time.

Every one of our 16 fictional people had a full profile. Each run, the model saw only one part of it:

1. Baseline: age, gender, occasion and budget.
2. Baseline plus interests.
3. Baseline plus habits and current life: what they actually do, what has changed lately.
4. Baseline plus dislikes.
5. Baseline plus past gifts and what they already own.
6. Baseline plus the relationship, in a sentence.
7. Everything.

Each version ran three times in a fresh session, in random order: 336 answers and 1,344 gift ideas, each scored without the rater knowing which version produced it.

What each detail did on its own

The main measure was recipient grounding: can you trace the idea back to something we told the model about this person? 1 means no, 5 means it uses several real facts without inventing new ones.

Genericness told the same story from the other side. With demographics alone, almost every idea could have gone to anyone of that age (4.9 out of 5 on our "could fit anyone" scale). The full profile brought that down to 2.3.

Demographics alone: the clichés we expected

Without real information the model did not just pick safe gifts. It filled the gap with guesses, averaging four invented facts per answer. For this mum the model wrote them down itself: "assuming she enjoys self-care bath rituals", "assuming she has children, enjoys sentimental keepsakes". Every added type of real detail lowered that number in this test. The full profile cut it to 1.7.

Interests vs habits

Interests are the obvious input. "Likes coffee" gives the model more to work with than "woman, late 50s".

Habits work differently. Compare:

likes coffee

with:

makes coffee before the rest of the house wakes up and complains that the grinder is loud

The second describes a problem to solve. It led straight to a quiet manual grinder, which "likes coffee" never produced. On average the two kinds of detail scored the same, but they found different gifts: interests found things for the hobby, habits found things for the day.

Interests also had a small habit of taking over. In 6 of 48 interest-only answers, three or more of the four ideas came from the same category. With habits it was 3 of 48, with the full profile 1. A single strong clue can turn a whole person into "the cycling guy".

Why dislikes and the relationship scored low on their own

Dislikes and relationship context made the smallest difference by themselves. That does not make them useless. "Hates clutter" tells the model what not to buy, but on its own gives it nothing to build from, so it falls back on defaults that avoid clutter.

In the full profile they seem to do their most useful job: filtering the ideas the other details produce. Our separate tests point the same way. One "avoid" sentence removed every unwanted category (Research 002), and listing past gifts cut repeats from 67% of answers to 19% (Research 003).

Did the full profile help, or just add noise?

It helped. A common worry is that more context gives the model more ways to get distracted. Within the size of our profiles, about 100 words, we did not see it.

The full profile scored higher on grounding than the best single detail (up 0.7, 95% interval 0.5 to 0.9), was the least generic, and invented the fewest facts. The ideas did not get less varied either: between 3.6 and 3.9 different directions out of 4 in every version.

We did not test very long profiles. A page of notes might behave differently.

What to tell AI when you want better gift ideas

In order of how much each did on its own:

1. What they actually do, and anything that has changed lately. A habit or a small complaint often beats a hobby label.
2. One or two interests that still matter now.
3. Recent gifts and things they already own, so the obvious repeats disappear.
4. A few things to avoid, if you know them.
5. The relationship in a sentence, when it changes what would be appropriate.

Then ask the model to show its working:

For each idea, tell me which detail from my description led to it, and separate what I told you from anything you are assuming.

That does not make the model know the person. It makes the gap between your evidence and its guesses easy to see.

Method

Tested

5 October 2026

Model

Gemini 3.5 Flash via the Gemini API

Settings

Fresh call per run, no system instruction, no search, no memory, default temperature

People

16 fictional recipients, UK, budgets £25 to £100

Conditions

7, from age, gender, occasion and budget alone up to the full profile

Runs

3 per person per version, random order, 336 answers, 1,344 ideas

Scoring

Each answer scored blind to its version, by an AI rater (Claude) using a fixed rubric written before any output was read

Analysis

Differences compared within the same person (average of the three runs), with 95% bootstrap intervals over people

Raw data

Free to reuse under CC BY 4.0: credit Giftin.ai and link to this page.

Keep reading