When the AI had the whole history, its ideas tied more clearly to the person and repeated past gifts less often than with the quick brief. Both held at every one of the five occasions.
Whether that history was pasted as prose or formatted as a profile made no reliable difference. The benefit came from the information being there, not from the format. The catch is effort: by the fifth occasion, matching the full context meant supplying about 210 words of background every time.
What we tested
The hardest part of asking AI for a gift idea is often not the AI. It is rebuilding the person from memory every time.
My sister is 31. She likes films and skincare. She recently moved. She hates clutter. Last year I gave her...
Then Christmas comes and you do it again. A saved profile changes the question to "what is different this time?" We wanted to know whether that changes the ideas themselves, not just the typing.
To isolate the effect of context, we used one model throughout and changed only how much recipient history it got, and how that history was formatted.
Five occasions per person
Each of 6 fictional people moved through five occasions: birthday, Christmas, birthday, Christmas, birthday. Before each one there was something new to know, and after each one a gift was given and we noted how it went. For one sister:
- Birthday. She mentioned she would love a proper camera one day. Gift: a pottery workshop, a big success.
- Christmas. She bought herself a good camera. Gift: a camera strap, which she uses.
- Birthday. She now avoids fragranced skincare after a reaction.
- Christmas. She started working from a small home office two days a week.
- Birthday. She mentioned cold hands cycling last month.
Every occasion was run three ways, three times each, in fresh sessions:
- Quick brief: the basics, one interest line and the newest detail.
- Full history, prose: everything we knew, written out as a paragraph.
- Full history, structured profile: exactly the same facts, laid out as a recipient profile.
That is 270 answers, scored without the rater knowing which version produced them.
Did the full history help?
Yes, clearly, and at every occasion.
How clearly the ideas tied to the person (1 to 5)
- Quick brief3.28
- Full history, prose4.17
- Full history, structured profile4.18
Answers that repeated or nearly repeated a past gift
- Quick brief64%
- Full history, prose38%
- Full history, structured profile28%
Full history also produced more varied answers (about 3.9 different directions out of 4, against 3.5). The prose version used fewer invented facts than the quick brief (down 0.67 per answer, interval 0.26 to 1.07); the structured-profile version sat in between. On average, more than one idea in each full-history answer was built directly on a past gift or how it went.
Not every remaining repeat was a mistake. One partner's baking course had been "the best thing he did all year", and the model, seeing that, suggested an advanced masterclass. That is history used well.
Did the profile format matter?
Not in any way we would rely on. Both versions contained the same facts, and they scored almost identically on how well the ideas fitted (difference 0.01, interval -0.08 to 0.12). The structured profile had slightly fewer repeats and slightly more guesses. Small, and pulling in opposite directions.
That is the useful result: the format was not the advantage. The information was. Pasting the same current history as prose worked just as well in this test. A saved profile is useful because it can keep that history available without rebuilding it from scratch every time.
The practical cost of rebuilding the context
Words of background needed per prompt, by occasion
- Occasion 1118
- Occasion 2145
- Occasion 3168
- Occasion 4189
- Occasion 5212
The history grows with every occasion. By the fifth, matching the full-history condition meant supplying about 210 words and remembering every gift and how each one went. Our test shows the gap when that context is omitted; it does not tell us how often people would actually type the full history by hand.
Out-of-date details: the risk we tested for
A saved profile can be wrong for longer than a fresh prompt. "Wants a camera" is useful until she buys one.
We built that trap in. The sister's profile kept both notes: she wanted a camera, then she bought one. In every full-history answer after the purchase, the model treated the camera as owned. It suggested a wrap or a bag for it, never a camera. We found no ideas that relied on an outdated detail and none that contradicted a known dislike.
There was a milder pattern. Once the camera was in the profile, a camera accessory turned up in most answers for the next three occasions. Profiles do not only stop repeats, they can keep pulling the model back to the same clue. Removing a note once it has done its job is part of keeping a profile useful.
How this maps to Giftin
Giftin: Birthday & Gift Ideas keeps the longer timeline for each person: birthdays and other dates, reminders, interests and notes, things to avoid, saved ideas and past gifts. Gift ideas can start from that saved context instead of a blank box.
This study isolates the part of that workflow we can test cleanly: having the full, current recipient history available. It does not compare Giftin head-to-head with another gift tool. That would need a separate test.
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 product memory; context supplied explicitly in the prompt
- People
- 6 fictional recipients, 5 occasions each, with updates and gift outcomes written before any run
- Conditions
- Quick brief; full history as prose; the same full history as a structured profile
- Runs
- 3 per person, occasion and version, random order, 270 answers, 1,080 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 and occasion (average of the three runs), with 95% bootstrap intervals
Raw data
Free to reuse under CC BY 4.0: credit Giftin.ai and link to this page.