How accurate is ChatGPT on perfume notes? 3 engines tested

For 100 perfumes, ChatGPT returned every note on the official list in 82% of its answers, Perplexity in 78% and Gemini in 69%.¹ Gemini added at least one note not on the official list in 49% of its answers, against 24% for ChatGPT and 25% for Perplexity.¹ Treat answers as a good first look, then check the house page.

• All three engines are usually close on the notes themselves: median recall was 1.00 for each.¹
• The differences show up in extras. In 32% of its answers Gemini added an off-list note that neither other engine named, against 10% for ChatGPT and 11% for Perplexity.¹
• Answers about designer perfumes were 22 to 32 percentage points more likely to carry an off-list note than answers about niche ones, depending on the engine.¹
• For 39 of the 99 perfumes all three engines answered, all three matched the official list exactly.¹
• This measured answers that search the web, not what a model remembers. Claude was not tested.¹

How we tested

We asked ChatGPT, Gemini and Perplexity the same plain question for 100 perfumes from 45 houses: “What are the top, heart and base notes of House Perfume?” Each engine got one call per perfume on 20 October 2026, from a US location, in English. That made 300 calls.¹

• Answer key. Scento’s Perfume Compendium, which held 754 entries on that date.² A perfume was eligible only with a full top, heart and base pyramid whose source is the house’s own page, or two authorised retailers that agree. For 91 of the 100 perfumes the key is a house page; for 9 it is two agreeing retailers.¹
• Sample. 50 designer perfumes (heritage houses included, because too few designer-typed entries qualified) and 50 niche perfumes, with Arabian houses left out of both. Within each group, 25 are well known and 25 lesser known, a judgement written down before any engine was asked. Official lists run from 1 to 17 notes, with a median of 8.5.¹
• Engines. ChatGPT and Gemini as the consumer answers (the API reported the models as gpt-6 and 3.5 Flash-Lite), and Perplexity’s sonar model with web search.¹ Claude was not included.¹
• Grading. A script, with no language model involved, mapped every note an engine named to a glossary of 411 notes and compared it with the official list. Where an answer gave several versions, the best-matching one was graded. A hand check of 20 gradings agreed exactly on 17 and within one note on 19; the reviewer could see the script’s result, so the check was not blind.¹

Four measures matter below. Recall is the share of official notes an answer contains. Precision is the share of the notes it names that are on the official list. A note not on the official list may still be correct, because houses sometimes publish only a few key notes. An exact list has every official note and nothing extra.¹

Results across all 100 perfumes

Gemini answered 99 of the 100 perfumes; one reply was a service error. ChatGPT and Perplexity answered all 100. None refused. Gemini’s answer for Juliette Has a Gun Not a Perfume named the single official note, Cetalox, in prose without a tier list, so it counts for recall only, and Gemini’s shares for off-list notes and exact lists are out of 98 answers.¹

Median recall was 1.00 for all three engines, and median placement accuracy (notes in the right top, heart or base tier) was 100% for all three. Median precision was 1.00 for ChatGPT and Perplexity and 0.97 for Gemini.¹ Because the medians sit at the ceiling, the table uses shares and means to show where the engines differ.

What each engine did

ChatGPT

ChatGPT gave every official note in 82% of answers and an exact list in 70%. It added a note beyond the official list in 24% of answers.¹ It cited the house’s own site in 85 of its 100 answers.¹

Perplexity

Perplexity was close behind: 78% with every official note, 70% exact lists and extras in 25% of answers.¹ It cited the house site in 74 answers.¹

Gemini

Gemini gave every official note in 69% of answers, matched the list exactly in 43% and added an off-list note in 49%. In 32% of its answers it added a note that neither other engine named.¹ It cited the house site in 55 of its 99 answers. When it did, its median precision was 1.00; when it did not, 0.93.¹ For Perplexity the same split was 1.00 against 0.97, and for ChatGPT 1.00 against 1.00.¹

Where the engines agree

Of the 99 perfumes all three answered, all three returned every official note for 59, and all three returned the list exactly for 39. At least one engine returned it exactly for 79, and for 20 none did.¹

Designer versus niche

The clearest pattern is in the extras. On every engine, answers about designer perfumes carried an off-list note more often than answers about niche ones.¹

Full-note recall also slipped for designer perfumes: 74% for ChatGPT against 90% for niche, 64% against 73% for Gemini, and 70% against 86% for Perplexity.¹ Those gaps are less certain; the intervals for the ChatGPT and Perplexity differences touch zero and Gemini’s includes it.¹

A shorter official list could explain part of this, since a house that publishes four notes leaves more room for extras. So we compared designer and niche at equal list length, pooling the three engines. The gap held in all three bands:¹

The study cannot say why. The designer and niche perfumes came from different houses and different pages, and that cannot be separated here.¹ Fame made no clear difference: the gap in off-list notes between well-known and lesser-known perfumes was between 4 and 8 points per engine, and every interval included zero.¹

Five examples

Viktor&Rolf Flowerbomb: the older pyramid

The house page lists four notes: peony on top, centifolia rose in the heart, freesia and vanilla in the base. Gemini and Perplexity both returned the same 11-note pyramid, found three of the four official notes (no peony) and gave eight notes that are not on the list. Perplexity’s answer then added that the brand’s current description “highlights peony, Centifolia rose, freesia, and vanilla”, so it held both versions and led with the older one. ChatGPT found all four and added four more: jasmine, orchid, warm amber and patchouli.¹

Prada Infusion de Vétiver: one list tripled

The house page lists four notes: bergamot; ginger and cardamom; vetiver. ChatGPT and Perplexity returned exactly those, in exactly those places. Gemini returned 12 notes, including all four official ones plus eight that neither other engine gave, among them Italian lemon, Italian tangerine, lime, Tunisian neroli, tarragon and Turkish rose. Its precision against the official list was 0.33.¹

Montblanc Explorer: right notes, swapped tiers

The house page puts bergamot, pomarose and sage on top, vetiver, leather and sandalwood in the heart, and patchouli, ambrofix and akigalawood in the base. ChatGPT put patchouli and ambrofix in the heart and vetiver and leather in the base. Of the six notes it got right, two sat in the correct tier, a placement accuracy of 33%. Perplexity reproduced all nine notes in the correct tiers. Gemini found six of nine, wrote Ambroxan where the house lists Ambrofix, and placed every matched note correctly.¹

Jean Paul Gaultier Le Male Elixir: the same swap, three times

Two authorised retailers agree on lavender on top, tonka bean in the heart and benzoin in the base. All three engines returned the same seven notes in the same tiers: lavender and mint on top, vanilla and benzoin in the heart, honey, tonka bean and tobacco in the base. Each found all three official notes, swapped the tiers of tonka bean and benzoin and added four notes not on the list. Each scored 33% placement accuracy and 0.43 precision. The data cannot say why the three answers match.¹

Rochas Eau de Rochas: three outcomes from one question

The house page lists ten notes: bergamot, mandarin, lemon and verbena on top; narcissus, jasmine and lily of the valley in the heart; oakmoss, myrrh and patchouli in the base. Perplexity returned the list exactly. ChatGPT returned nine of ten, without verbena, plus iris. Gemini returned 18 notes, found eight of the ten and gave ten notes that no other engine named, among them lime, basil, coriander, carnation, wild rose, sandalwood and amber.¹

What these results do and do not show

• Retrieval, not memory. All three engines search the web as they answer, and most answers cited the house’s own page. High recall shows these products can find a published list. It does not show that any model knows notes, and it says nothing about models without search.¹
• One moment. One call per perfume per engine on 20 October 2026. We made no repeat runs, so run-to-run variation is unknown.¹
• One published list. A note missing from the official list is not automatically wrong. Precision is a lower bound on correctness.¹
• Versions. The question named no concentration, and the official page covers one version. Best-version grading softens this without removing it.¹
• Sample and grader. The sample is mostly Western designer and niche houses that Scento could verify. The grader was tuned on these same 100 perfumes, and it gives no credit when an answer and the list name a note at different levels of detail, such as “cinnamon” and “spices”.¹
• Not tested. Claude, engines without web search, other languages or countries, and whether the official lists themselves are right.¹

What it means for shoppers

For most perfumes, an answer engine returns every official note. In this sample, all three engines returned every official note for 59 of 99 perfumes. But a quarter to a half of answers carried at least one note that is not on the official list, and the extras were most common for designer perfumes.¹ Three habits help.

• Check the house’s own page. It is where 91 of the 100 official lists in this study came from, and the first place to look for a reformulation or a new pyramid.
• Read a long list with care. When an answer gives many more notes than the house page, it may be describing an older or different version of the perfume.
• Smell before you commit. Notes on paper never tell you how a perfume sits on your skin. Perfume samples let you try one before you buy a full bottle.

The official lists in this study come from Scento’s Perfume Compendium, which stores a saved source page for every entry. Developers can read it through the Perfume Compendium and API page.²

Sources

1. 1Scento analysis of ChatGPT, Gemini and Perplexity answers for 100 perfumes, graded against the Perfume Compendium, Scento, 20 October 2026.
2. 2Scento Perfume Compendium, Scento. scento.com/developers, read 21 October 2026.