All base wines used in this study were obtained from BevZero (Paarl, South Africa), a commercial de-alcoholisation company. Although each experiment used different base wines available at the time, all wines were de-alcoholised using a Flavourtech Spinning Cone Column (SCC) (Flavourtech Pty Ltd, Griffith, New South Wales, Australia) installed and operated by BevZero. The SCC had a maximum nominal capacity of 10,000 L/h and operated under vacuum at temperatures ranging from 30 to 40 °C (Flavourtech, 2024). The system was fully automated, ensuring reproducibility and control.

For pH and titratable acidity (TA) analysis, a 50 mL wine sample was titrated using a 702 SM Metrohm auto-titrator combined with a 760-sample changer and a 722 stirrer (Metrohm SA (Pty) Ltd, Johannesburg, South Africa). The sample was titrated to pH 7 using 0.33 N sodium hydroxide (NaOH, KIMIX, Cape Town, South Africa) as a titration standard. Subsequently, the volume of NaOH used to titrate the wine was used to calculate pH and TA. All samples were analysed in duplicate (n = 2).

Ethanol, glycerol, sugars, and organic acids were analysed using high-performance liquid chromatography (HPLC, Agilent Technologies, Waldbronn, Germany), following the method described by Eyéghé-Bickong et al. (2012), with modifications to sample preparation and standard handling. A glycerol standard solution was prepared by dissolving 0.5 g of food-grade glycerol (Merck Life Science, Johannesburg, South Africa) in 25 mL of deionised water (dH ₂O). For sample preparation, 400 μL of sample was diluted with 600 μL of deionised water and 1 mL of a single mixed internal standard solution containing ribitol and adipic acid at a combined concentration of 4 g/L (2 g/L of each compound). The final volume of 2 mL resulted in a 5× dilution of the original sample and final concentrations of 1 g/L ribitol and 1 g/L adipic acid. Ribitol was used as the internal standard for ethanol, glycerol, and sugars detected by refractive index detection, whereas adipic acid was used as the internal standard for organic acids detected at 210 nm. Both standards were chromatographically resolved from the target analytes and were used to normalise the measured analyte concentrations (Eyéghé-Bickong et al., 2012). Samples were centrifuged at 8400 g for 5 min to ensure proper separation. Following centrifugation, 1 mL of the supernatant was transferred into HPLC vials and crimp-sealed prior to analysis. Samples were analysed in triplicate or as specified to confirm whether required treatment concentrations were achieved.

The HPLC was performed using an Agilent 1100 Series HPLC system (Agilent Technologies, Waldbronn, Germany) equipped with a diode array detector (DAD) and a refractive index detector (RID). Separation was achieved using a Bio-Rad Aminex HPX-87H ion exchange column (300 mm × 7.8 mm, 3.5 μm particle size), protected by a Bio-Rad Aminex guard cartridge (23 mm × 10 mm, 3.5 μm). The mobile phase consisted of 5 mM sulfuric acid (H ₂SO ₄) in water, run isocratically at a flow rate of 0.5 mL/min for 39 minutes. Column temperature was maintained at 55 °C, and the injection volume was 10 μL (Eyéghé-Bickong et al., 2012). Organic acids were monitored using UV detection at 210 nm, while sugars and alcohols were detected using RID. Peak integration and calibration were performed manually using the prepared standards, with data acquisition and processing carried out using ChemStation Rev. A.10.02 software (Agilent Technologies ^©).

Sensory evaluations were conducted using panels of industry experts, with panel composition and specific protocols detailed under each experiment. Each panellist was seated in an individual sensory booth to minimise bias and ensure consistent evaluation conditions. Wine samples were served at room temperature (20 ± 2 °C) in 25 mL portions using standard ISO tasting glasses (ISO 3591, 1977). To prevent positional bias, the presentation order of the samples was randomised for each panellist by means of a William–Latin square design, and a 10 to 15 min break was provided between flights to reduce sensory fatigue. Each treatment was evaluated in duplicate to enhance the reliability and reproducibility of the sensory data. For sample identification, each wine was assigned a unique three-letter code, ensuring blind tasting conditions. Ethical clearance was obtained from the REC: Social, Behavioural and Education Research Ethics Committee at Stellenbosch University (FESCAGRI-2025-3007).

The sensory data were collected using Compusense ^® cloud software (Compusense, Guelph, Canada) and exported to Microsoft Excel 2023 (Microsoft Corporation, Redmond, Washington, USA). Univariate and multivariate statistical analysis was conducted using TIBCO Statistica ^™ Software 14.0.0.15 (TIBCO Software Inc., Palo Alto, California, USA). Mixed model analysis of variance (ANOVA) type III was performed with “judge” as a random effect and “treatments” and “replicates” as fixed effects. To account for mixed effects, the Kenward–Roger test was used to adjust the degrees of freedom (Luke, 2017). Fisher’s LSD post-hoc test was used to compute pairwise comparisons between the least squares (LS) means of the treatments with p ≤ 0.05.

The wines used in this experiment were sourced from BevZero (Paarl, South Africa), a commercial de-alcoholisation facility. The trial aimed to investigate whether glycerol addition has a perceivable and positive impact on the mouthfeel of de-alcoholised wine. Two fractions of Sauvignon blanc were obtained, including the full-strength (FS) wine prior to de-alcoholisation, as well as the de-alcoholised wine fraction containing essence. Table 1 summarises the wine parameters of the Sauvignon blanc before and after de-alcoholisation.

To assess the impact of glycerol on mouthfeel, the de-alcoholised wine fraction underwent four treatments with varying glycerol concentrations. Due to the typically high acidity and low residual sugar (RS) levels found in de-alcoholised wine, a standard addition of 30 g/L crystal grape must (Naturalia Ingredients, Mazara del Vallo, Italy) was made to each sample prior to glycerol (Glycerin Food Grade, Pound Chemicals, Paarl, South Africa) adjustments. Glycerol additions were prepared at 0, 1, 2, and 4 % v/v, corresponding to target final glycerol concentrations.

The following target glycerol concentrations were applied:

A panel of nine de-alcoholised wine and production specialists (six male and three female), aged 32 – 59 years and with previous experience evaluating de-alcoholised wines, were invited via email to participate in the sensory evaluation.

Sensory evaluations assessed mouthfeel, taste, and overall quality differences between de-alcoholised Sauvignon blanc wines subjected to different glycerol treatments. Each wine sample was evaluated in duplicate for free listing (FL) and quality scoring to ensure repeatable, consistent responses.

The free listing (FL) methodology was used to assess the sensory profile of each glycerol treatment (Vieira et al., 2020). Panellists were instructed to describe the taste and mouthfeel attributes they perceived for each wine sample using a blank sheet of paper. There were no restrictions on the number or type of descriptors to be provided. Following the sensory evaluation, the handwritten responses were manually transcribed and exported to Microsoft ^® Excel 2023 (Microsoft Corporation, Redmond, Washington, USA) for data processing. When no descriptors were recorded for a modality, the term “no comment” was noted to maintain consistency in the dataset. To prepare the data for analysis, the dataset was first screened for Afrikaans descriptors and translated into English equivalents. Spelling and grammatical errors were then corrected to ensure accurate interpretation of responses. Finally, descriptors with overlapping or synonymous meanings were standardised to a single term to reduce redundancy and ensure consistency across panellists and replicates (Campo et al., 2008). Table 2 provides an overview of how we standardised the original mouthfeel descriptors.

Thereafter, the standardised descriptors were grouped into broader sensory categories using a process known as lemmatisation. Lemmatisation reduces words to their base or root form and clusters similar terms (McLaughlin et al., 2022). In the present study, descriptors considered synonymous or representative of the same sensory concept were assigned to a common sensory subcategory. For example, “watery”, “thin”, and “diluted” were grouped under “poor mouthfeel”, as these terms indicated an overall negative perception of insufficient palate weight and mouthfeel intensity. We retained “low body” as a separate subcategory for descriptors that explicitly refer to body or closely related sensations, such as “light palate”. When a response contained more than one distinct sensory concept, we retained each concept separately. Each descriptor was classified as positive, neutral, or negative, based on its evaluative contribution to the sensory profile (Campo et al., 2008; Valentin et al., 2021). The lemmatisation and classification were performed independently by two researchers. Differences were discussed until consensus was reached, and the agreed coding framework was then applied consistently across all panellists, treatments and replicates. This framework enabled a structured interpretation of the data and facilitated the identification of favourable and unfavourable attributes across treatments. Examples of sensory categories used and classifications are outlined in Table 3.

Not all descriptors were included in the lemmatisation and subsequent statistical analysis. Descriptors were excluded on two grounds. First, comparative descriptors that lacked standalone sensory meaning and were dependent on other samples, such as “sweeter than 160” or “more vegetal”, were removed as they could not be interpreted in isolation. Second, entries consisting of illegible handwriting that could not be confidently transcribed were excluded.

Each wine sample was evaluated using the overall 20-point quality scoring system, which is commonly applied in wine sensory analysis (Brand et al., 2018). Panellists were instructed to assign scores for each sensory modality as follows: three points for appearance, seven points for aroma, and ten points for palate. All quality scores were recorded electronically using Compusense ^® Cloud Software (Guelph, Canada) and subsequently exported to Microsoft ^® Excel 2023 (Microsoft Corporation, Redmond, Washington, USA).

Sensory data were collected using Compusense® cloud software (Compusense, Guelph, Canada), exported to Microsoft Excel 2023, and analysed using TIBCO Statistica™ Software (version 14.0.0.15). Numerical quality comparison scores were analysed using mixed-model ANOVA type III. Glycerol treatment and replicate were specified as fixed effects, while panellist was included as a random effect to account for repeated evaluations. Denominator degrees of freedom were estimated using the Kenward–Roger approximation (Luke, 2017). When the ANOVA indicated a significant treatment effect, Fisher’s least significant difference test was used for pairwise comparisons between treatment least squares means at p ≤ 0.05. Results were presented as least squares means with 95 % confidence intervals.

The original free listing responses were summarised as citation frequencies before correspondence analysis. Correspondence analysis (CA) was performed using a 4 × 13 frequency table, in which the four rows represented the glycerol treatments and the 13 columns represented the sensory descriptor categories obtained after standardisation and lemmatisation. Each cell contained the total number of citations of a descriptor category for a particular treatment across panellists and replicates. The first two dimensions accounted for 78.92 % and 13.41 % of the total inertia, respectively, representing 92.33 % of the overall variation in the descriptor frequency data. CA was used as an exploratory technique to visualise associations between glycerol treatments and sensory descriptors.

To investigate the effects of glycerol (Merck Life Science, Johannesburg, South Africa), mannoproteins (MANNOFEEL ^®, Laffort, Paarl, South Africa), and gum Arabic (STAVIN ^® SP, Laffort, South Africa) on the mouthfeel of de-alcoholised Chenin blanc and Shiraz wines, a reduced treatment combination design was used. Selected combinations of the three additives were evaluated at industry-relevant concentrations. The number of combinations was limited to reduce sensory fatigue and support reliable sensory evaluation. Each formulation was treated as a separate treatment, and the design was not used to estimate the independent main effects or statistical interactions of the three additives.

Glycerol was evaluated at two levels, namely no addition and a final addition of approximately 35 g/L. Mannoproteins (0, 50, and 150 mL/hL) and gum Arabic (0, 200, and 350 mL/hL) were evaluated at three levels. The levels for mannoproteins and gum Arabic were selected to represent no addition, an industry-recommended dosage, and a maximum dosage based on supplier guidelines. Glycerol dosages used for this experiment were determined based on findings from Experiment 1. A complete factorial design would have included 18 additive combinations. However, only 14 combinations were selected to limit the number of samples presented to panellists and reduce sensory fatigue. Each selected treatment was applied to both de-alcoholised Chenin blanc and Shiraz wines, with a standard addition of 30 g/L crystal grape must.

Fifteen de-alcoholised wine and production specialists (10 male and 5 female), aged 27 to 50, participated in the sensory evaluation.

A modified Rate-All-That-Apply (RATA) methodology was used to assess seven predefined taste and mouthfeel attributes: sweetness, sourness, bitterness, astringency, length, body, and balance (Vidal et al., 2018). All attributes were presented for each sample, and panellists were required to provide a response for every attribute. A score of 0 indicated that the attribute was not perceived, while perceived attributes were rated on a five-point intensity scale (1 = low, 3 = medium, and 5 = extremely intense), with scores of 2 and 4 representing intermediate intensity levels. The 0 response option distinguished attributes that were not perceived from unanswered items and ensured complete data capture. Each de-alcoholised wine sample was evaluated in duplicate.

A linear mixed model analysis of variance (ANOVA) with Type III sums of squares was performed for each sensory response. Treatment and replicate were included as fixed effects, while panellist was included as a random effect. Each panellist evaluated every treatment on two occasions, and both scores were retained as individual observations rather than averaged. Evaluation occasion was included to account for possible systematic differences between the first and second assessments, while the random panellist effect accounted for repeated observations from the same individual. The Kenward–Roger approximation was applied when testing the fixed effects in the mixed model (Luke, 2017). Fisher’s least significant difference test was used for pairwise comparisons between treatment least squares means at p ≤ 0.05.

To complement the mixed model ANOVA, principal component analyses (PCA) were performed for the de-alcoholised Chenin blanc and Shiraz wines using the correlation matrix. For each wine, the input matrix consisted of 14 rows representing the additive combinations and 7 columns representing the mean intensity scores for sweetness, sourness, bitterness, astringency, length, body, and balance. PCA biplots were used to explore relationships among the sensory attributes and to visualise similarities and differences in the sensory profiles of the additive combinations. For Chenin blanc, the first two principal components explained 74 % and 10 % of the total variance, respectively, representing 84 % cumulatively. For Shiraz, the first two principal components explained 76 % and 9 % of the total variance, respectively, representing 85 % cumulatively.