Background Medications licensed for the treatment of dementia have limited efficacy against cognitive impairment or against the distressed behaviours (behavioural and psychological symptoms, or behaviour that challenges) which are also often the most distressing aspect of the disorder for caregivers. Complementary therapies, including aromatherapy, are attractive to patients, practitioners and families, because they are perceived as being unlikely to cause adverse effects. Therefore there is interest in whether aromatherapy might offer a safe means of alleviating distressed behaviours in dementia. Objectives To assess the efficacy and safety of aromatherapy for people with dementia. Search methods We searched ALOIS, the Cochrane Dementia and Cognitive Improvement Group Specialized Register, on 5 May 2020 using the terms: aromatherapy, lemon, lavender, rose, aroma, alternative therapies, complementary therapies, essential oils. In addition, we searched MEDLINE, Embase, PsycINFO (all via Ovid SP), Web of Science Core Collection (via Thompson Web of Science), LILACS (via BIREME), CENTRAL (via the Cochrane Library), ClinicalTrials.gov and the World Health Organization (WHO) trials portal (ICTRP) on 5 May 2020. Selection criteria We included randomised controlled trials which compared fragrance from plants in an intervention defined as aromatherapy for people with dementia with placebo aromatherapy or with treatment as usual. All doses, frequencies and fragrances of aromatherapy were considered. Participants in the included studies had a diagnosis of dementia of any subtype and severity. Data collection and analysis Two reviewers independently selected studies for inclusion, extracted data and assessed risk of bias in included studies, involving other authors to reach consensus decisions where necessary. We did not perform any meta‐analyses because of heterogeneity between studies, but presented a narrative synthesis of results from the included trials. Because of the heterogeneity of analysis methods and inadequate or absent reporting of data from some trials, we used statistical significance (P ≤ or > 0.5) as a summary metric when synthesising results across studies. As far as possible, we used GRADE methods to assess our confidence in the results of the trials, downgrading for risk of bias and imprecision.