Without gift history, two out of three answers suggested something the person had already been given, or something very close to it. A third of answers repeated a past gift outright.
Listing three past gifts stopped exact repeats completely and cut near-repeats to one answer in five. The model did not overreact: in no answer did it drop an interest just because it had been gifted before.
What we tested
"Tell the AI what you already gave them" sounds like obvious advice. It solves a real problem: nobody wants a confident suggestion for wireless headphones when they gave the same person headphones last year.
But history raises a subtler question. Does the AI learn not to repeat the gift, or does it learn never to touch that interest again? Those are different outcomes, and the second would make the ideas worse.
Each of 16 fictional people was run three ways:
A. Interests only
She likes films, skincare, hiking and cooking. Suggest four birthday gift directions under £80.
B. Interests plus three past gifts
Previous gifts: cinema membership, skincare set and wireless headphones.
C. Past gifts plus a nuance line
Avoid exact or pointless duplicates, but do not automatically ban an entire category if another gift in it would still genuinely fit.
Every version ran three times in a fresh session, in random order: 144 answers, 576 ideas.
How we counted repeats
Each idea was checked against the person's three previous gifts:
- Exact repeat: the same gift again, like theatre tokens after theatre tokens.
- Near repeat: the same kind of product, a different brand or version. Another pair of earbuds, another cinema membership from a different chain.
- Same area: the same broad interest but a genuinely different gift. A theatre history book after theatre tokens.
Same-area ideas are not failures. Someone who loves cooking can receive more than one kitchen gift in their life.
Did gift history stop the repeats?
Yes, clearly.
Answers with an exact or near repeat of a past gift
- Interests only67%
- Plus past gifts19%
- Plus past gifts and nuance line21%
The repeats without history were not random. They were the obvious gift for the obvious interest, which is exactly why the giver had already bought it. The crossword fan got a crossword book in all three runs. The houseplant collector got a rare houseplant in all three.
- Ergonomic crossword pen set
- Garden kneeling pad
Flagged: National crossword compilation book- Bird feeder or seed
- Garden kneeler and seat
- Refillable mechanical pencil
- Wild bird feeder
- Wooden lap desk
Did the model overcorrect?
This was the risk we cared about most. "Last year: cinema membership" should mean "find a different film idea", not "no more films".
We did not see it happen. In none of the 144 answers did the model abandon an interest that the person clearly still had. Same-area ideas actually went up with history, from 1.5 to 2.1 per answer. The model kept the interest and changed the gift.
The repeats that survived
One answer in five with history still contained a near repeat. Looking at them, most were understandable:
- a craft beer subscription after a craft beer selection;
- a window bird feeder after a garden bird feeder;
- another strategy board game after a two-player board game.
These are "different enough?" judgement calls rather than blunders. Some of them may even be good gifts.
Repeats are not always bad
Our prompts listed past gifts without saying how they went. That hides something important. Two of our people had loved their previous gift:
- the aunt used her theatre tokens for a matinee she loved;
- the sister renewed her cinema membership herself afterwards.
For them, a repeat might be the best idea on the list. With bare history the model treated every past gift as something to avoid. It had no way of knowing which ones were hits.
- Flagged: Theatre gift vouchers
- Insulated travel flask
- Backstage theatre tour
- Flagged: Merino walking socks
- Insulated travel flask
- Theatre opera glasses
- Waterproof sit mat
- Theatre history book or play script
That suggests the useful form of gift history includes the outcome:
2025: theatre tokens. Used them for a matinee she loved.
2024: china ornament. Gave it away when she downsized.
We have not tested outcome notes yet. It is the obvious next experiment.
Did the nuance line help?
Not in the way we expected. Telling the model not to ban whole categories made no difference to repeats (21% against 19%) and there was no overcorrection to fix in the first place.
It did two smaller things. Answers covered slightly more different directions (up 0.27 distinct directions per answer against interests only, 95% interval 0.10 to 0.42). They also contained more invented details used as reasons (up 0.69 per answer, interval 0.21 to 1.21). The pattern was consistent: invited to build on past gifts, the model started reading them as evidence of taste. A vegan cookbook became "she is vegan". A coffee shop gift card became "he loves coffee". Neither was in the brief. In this test, the extra line did not earn its place.
What to tell AI about previous gifts
List the last three or four gifts. You do not need ten years of history to stop the most embarrassing repeats.
Previous gifts: cinema membership (2025), skincare set (2024), wireless headphones (2023).
If you know how a gift went, say so in a few words. "Loved it", "never used it" and "already had one" each point the model somewhere different.
This is the part a saved profile makes easy. Writing down past gifts as you give them, with a word about how they landed, means the next brief starts with evidence instead of memory.
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
- Interests only; plus three past gifts; plus past gifts and a nuance line
- Runs
- 3 per person per condition, random order, 144 answers, 576 ideas
- Scoring
- Each answer scored blind to its condition, by an AI rater (Claude) using a fixed rubric written before any output was read. The rater saw every person's gift history, including for the interests-only answers.
- 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.