Giftin.ai research
We tested what actually changes AI gift ideas.
Which details help? Do past gifts stop repeats? Does saying what not to buy work? Can you trust AI prices and stock?
What we found so far
Answers repeating or nearly repeating something already given, before and after listing three past gifts.
Invented assumptions per answer, bare brief vs full profile.
AI-suggested products that existed and fit the budget when we checked each one by hand.
The experiments
Each study asks one practical question about AI gift ideas. Open one to see the setup, results, examples, limitations and raw data.
Research 001Which details improve AI gift ideas?
The full profile worked best. Interests and everyday habits were the strongest individual details.
4.0→1.7invented assumptions per answer→
Research 002Does telling AI what not to buy help?
Yes. One short avoid list removed every banned category in this test.
33%→0%answers containing a banned category→
Research 003Do past gifts stop AI repeats?
Mostly. Three previous gifts cut exact and near repeats sharply.
67%→19%answers repeating a previous gift→
Research 005Does "partner" vs "coworker" change the ideas?
Less than you might expect. The recipient details mattered much more than the relationship label alone.
2×as many gift cards for a coworker as for a partner→
Research 006Can you trust AI gift prices and stock?
Prices held up better than stock. All 50 products existed and stayed within budget, but availability was less reliable.
50 of 50products were real and within budget→
Research 007Does a full recipient history beat a quick prompt?
Yes. Fuller history produced better-grounded ideas and fewer repeats.
Prose vs structured profile made no reliable difference.
3.3→4.2 / 5how clearly ideas tied to the person→
How we test
Change as little as possible
Most controlled studies start from the same bank of fictional recipients and change one thing at a time.
Decide the rules first
Scoring rules are written before results are reviewed. Studies that need judgement are scored without showing the rater which version produced the answer.
Show the work
Prompts, outputs, scoring and limitations are published with each finished study. Shopping claims are checked by hand.