Introduction

The Bordeaux region is one of the largest producers of protected designation of origin (PDO) wines in the world, with an average annual production over the last 10 years of around 5 million hectolitres. The red Bordeaux region comprises 31 PDOs covering 110,800 hectares and includes over 5,500 vineyards that produce more than 10,000 references. These include the regional appellations (Bordeaux, Bordeaux supérieur), whose wines may be produced in almost all the wine-growing communes, and the other appellations, which are generally classified according to the geographical and climatic organisation of the terroir (Figure S1). Bordeaux PDO is the prominent wine type and represents about 40 % of the total production of Bordeaux wines. Due to their large surface area, the regional appellations express a wide variety of terroirs and offer a wide range of wine styles. Red Bordeaux wines are renowned for their symbolic “red Bordeaux region wine style” characterised by “an intense colour, a fresh and complex aroma that is both fruity and floral, partially transforming during aging into a mineral, empyreumatic and spicy bouquet, a smooth and fresh taste, velvety on the attack, dense and silky on the finish” (Dubourdieu, 2014). Much research has attempted to characterise the analytical and sensorial aspects of premium Bordeaux red wines. However, on the French market, 88 % of red Bordeaux wines are sold for less than 10 € per bottle with an average of 5.66 € in 2020 (CIVB, 2021). Pineau et al. (2010) demonstrated that red Bordeaux wines tend to have their own sensory space, with distinctive blackberry and jammy-fruit notes. However, the entry-level wines used in their study tended to be less typical than other Bordeaux wines, suggesting differences in aromatic profiles (Pineau et al., 2010). These differences between value wines and fine wines could be due to variations in the winemaking process, particularly with regard to the modalities and duration of ageing in oak wood barrels, as well as to the management of viticultural parameters and the consideration of natural environmental factors related to the terroir. Yet few studies have been carried out to characterise the organoleptic profiles of entry-level red Bordeaux wines.

Several techniques may be used to produce wine sensory profiles depending on the objectives of the work, as well as on the time and sample setup required. The conventional profiling method, i.e., quantitative descriptive analysis (QDA), is the most classical approach to describe in detail the appearance, aroma, taste, and mouthfeel of wine, as evidenced by the abundant scientific literature on this issue (Campo et al., 2010; Guinard & Cliff, 1987; Heymann & Noble, 1989; Lattey et al., 2010; Pelonnier-Magimel et al., 2020). QDA relies mainly on comparison between samples, making evaluation context-dependent (Murray et al., 2001). Customarily, a relatively small number of descriptors (between 5 and 16) is quantified by a panel on an intensity rating scale. The two most common problems with QDA are the judges’ different understanding of the terms and the difficulty of using the intensity scales, especially for complex products, such as wines (Lawless, 1999). These problems can be partly resolved after extensive training, but this can be time-consuming, as pointed out in the study by Lund et al. (2009), where 14 panellists were trained for 70 hours to select 8 wines out of 52 by using QDA.

Several alternatives to QDA have been proposed. These methods, which are based on similarity or verbal agreement, can be used by trained, semi-trained, or untrained assessors (Valentin et al., 2012). Similarity/dissimilarity approaches such as sorting and projective mapping allow a set of products to be analysed in a single session without specific training, and they have already been used to assess the organoleptic properties of wines (Bécue-Bertaut & Lê, 2011; Campo et al., 2008; Gammacurta et al., 2014; Piombino et al., 2004). Usually, the judges are asked to analyse several samples presented simultaneously and to group them according to their similarity. The judges may be required to describe the groups briefly in their own words or by using a pre-established list of words. One of the disadvantages of these methods, however, is that all the wines are evaluated simultaneously, i.e., in a single session. The number of wine samples is usually between 9 and 20 (Courcoux et al., 2015).

The verbal free choice profiling and flash profile methods are interesting alternatives for limiting panellist training and considering inter-individual differences (Williams & Langron, 1984). The taster’s evaluation of the product is based on a list of individual attributes generated during an initial session. However, as explained by Valentin et al. (2012), these methods based on inter-sample ranking are not suitable for assessing a large number of samples.

Another verbal-based method is the check-all-that-apply (CATA) method. Using a list that is provided, panellists select all the descriptors they consider appropriate to describe a product (Ares et al., 2015a). Samples are presented one by one, so it is easier to evaluate a wide range of samples. CATA was first used with consumers to understand their preferences, but it also proved relevant to allow trained or semi-trained subjects (Alexi et al., 2018; Monteiro Veríssimo et al., 2021; Nougarède et al., 2023) and wine professionals (Brand et al., 2020; Panzeri et al., 2020; Vidal et al., 2018) to characterise products.

Studying and defining the characteristics of the wines of a particular appellation or variety, or wine-growing region, requires the selection of a representative wine sample. In general, around 30 wines are selected for this type of study (Cadot et al., 2012; Iosofidis et al., 2023; Schüttler et al., 2015). However, given the size of the Bordeaux production area, the diversity of its viticultural and winemaking practices, and the use of wine blends from six different grape varieties, a much larger number of samples is required for assessing entry-level red Bordeaux wines. To be representative, this number was set at 143 samples by the research partners in the present study.

Only the CATA method was considered potentially suitable with such a large number of samples. A panel of professionals tasted the wines as they are more capable of describing complex products such as wine (Gawel, 1997; Lawless, 1984), thanks to their perceptual and cognitive experience (Hughson & Boakes, 2001). Zamora and Guirao (2004) demonstrated that professional panels are less consensual than a trained panel but appear to be more discriminating about descriptors. Moreover, experts are used to tasting many samples in the same session. In tasting competitions, the maximum number of samples is 45 over the day, as defined by the International Organization of Vine and Wine (OIV) regulations. In France, during quality control tastings for PDO, professionals can evaluate up to 40 samples in 3 hours.

The aim of this study was therefore to assess whether CATA analysis could be used to describe a large sample of wines and identify their typical organoleptic profiles. To validate the relevance of this method for assessing a large number of wines, panel repeatability was tested, and the descriptions obtained for 20 wines were compared with profiles obtained by conventional descriptive analysis.

Materials and methods

1. Wines and basic chemical composition

One hundred forty-three red Bordeaux wines, all readily available on the French market, were selected (Table 1). All wines were from the 2018 (76 %) or 2019 (24 %) vintages and from different appellations in the Bordeaux region, with a price ranging between 3 € and 8 €, including tax. Wines were donated or purchased from individual and cooperative wineries, trading houses, or distributors. Bottled wines were stored in a cellar and allowed to equilibrate at room temperature for 24 hours before tasting. At least five bottles were used for the CATA questions (bottles identified by the panel to be corked were immediately replaced), and another one was used for QDA.

The standard chemical parameters of wines, like ethanol content level (% v/v), residual sugars, pH, titratable acidity, volatile acidity, and SO₂ content, were analysed using the official methods or those recommended by the OIV.

2. Check-all-that-apply (CATA) questions

2.1. Tasters

The expert cohort for this study consisted of 48 wine industry professionals (69 % male; age 46 ± 9 years (mean ± s.d.)), including wine merchants, wine researchers, consultant oenologists, winegrowers, trade association representatives, agronomists, and coopers (Table S1). Their experience and qualifications met the criteria of “expert” set out by Parr et al. (2004). All participants were volunteers. They provided written informed consent and completed a demographic questionnaire. The experimenter explained that the information they provided would be used solely for scientific purposes and treated as strictly confidential. Anonymity of the results was guaranteed. They were informed of the presence of spittoons and breathalysers. They were not paid for their participation.

2.2. Procedure

Assessing 143 wines in a single session poses the issue of panellist sensory fatigue. Thus, the wines were randomly divided into three sessions of 48, 48 and 47 wines (Table 1). Each session comprised five sub-sessions. The panellists attended a sub-session of each session in random order and therefore analysed all the wines. For technical reasons, it was not possible to serve the wines in a random order to each judge. Thus, for each session, a total of 10 different presentation orders were tested. The sessions took place in the sensory laboratories of the Institute of Vine and Wine Science (ISVV, Villenave-d’Ornon, France) or at the Planète Bordeaux site (Beychac-et-Caillau, France) according to international guidelines (ISO 8589:2010). Wines were labelled with three-digit random numbers and served (20 mL) in standard ISO wine glasses (ISO 3591:1977). They were presented at each session one by one in random order. Still water and bread were available for rinsing between samples, if the panellists requested them.

The panel was instructed first to evaluate the overall perceived quality of each wine sample on a 1-to-10-point intensity scale, then to tick in the accompanying list the words that applied to the wine. The CATA list was composed of 62 descriptors divided into 12 groups (Table 2): visual examination (six terms), global aroma (eight terms), fruit nuances (six terms), green nuances (four terms), woody nuances (six terms), spicy nuances (four terms), flavours (five terms), off-flavours (five terms), taste (three terms), mouthfeel (seven terms), tannin characterisation (five terms), and aftertaste length (three terms). Flavour is defined as the aromatic part perceived during retronasal evaluation. This classification into 12 groups facilitated the taster’s work by reducing the cognitive charge. The off-flavours group also contained an open-ended comment question that was not used in the processing of CATA data, but only to identify and discard wines with off-flavours different from those proposed. All descriptors were selected according to several studies on red wines (Chira et al., 2012; Gawel et al., 2000; Pelonnier-Magimel et al., 2020) and to evaluations by researchers who are experienced in tasting and in descriptive sensory analysis. The 12 groups were presented in a different randomised order to each judge and between each wine to avoid fatigue (Ares et al., 2015b; Meyners & Castura, 2016). Tasters could select as descriptors as needed.

The procedure of the sensory analysis and the instructions given to the assessors were unchanged between sessions. First, the panellists were informed that they would evaluate red Bordeaux wines, sold to consumers between 3 € and 8 €, including tax, from the 2018 or 2019 vintages, all appellations and origins combined. Then, a warm-up took place in which two wines were presented blind, one by one. These wines were evaluated in a previous pilot session with the same CATA question protocol by three ISVV researchers to determine their profile. The first wine was evaluated by the panel, followed by a discussion on its evaluation and comparison with the profile previously determined. The second wine was then evaluated, and perceptions were discussed and compared. The panellists were informed that the same two wines would be presented in the same order during the warm-up session at the beginning of all sessions. The aim was (1) to familiarise the panel with the CATA questions, and (2) to place them in the same conditions at each session to enhance reliability (Plemmons & Resurreccion, 1998). The two wines selected for the warm-up were among the 143 wines and were also evaluated (wine codes 342 and 539, Table 1). However, this information was not given to the panellists. Finally, all the wines of the sub-session were evaluated. To evaluate the repeatability of the panel, two wines identified with two different three-digit codes were presented twice in random order (Table 1 and Table S2). Each sub-session lasted about 3 hours, with a forced break of at least 20 minutes taken after the sensory evaluation of 25 wines. The CATA study was carried out from 4 to 17 December 2020 and 7 to 13 January 2021, the entry-level wines being intended for consumption within a few years of bottling. The data were collected using a Google Form on a laptop, tablet, or smartphone.

3. Conventional quantitative descriptive analysis

Following the CATA study, 20 of the 143 wines were analysed by a semi-trained panel. A wine selected from the 143 wines and characterised by an average level of quality was also evaluated by the panel and used as a standard wine. Wines used in this part of the study are indicated in Table 3.

3.1. Tasters and training

Participants were 17 ISVV students or staff members (2 men and 15 women, median age: 24 years) who were selected on the basis of their interest and availability. All of them had extensive wine tasting experience, but none of them had previously taken part in descriptive analysis panels, and none had participated in the CATA study. They were not paid for their participation.

The panellists received descriptive sensory training once or twice a week over a four-month period. They were provided with a list of 12 terms established by compiling terms from CATA and other lists used for describing Bordeaux wines (Chira et al., 2012; Pelonnier-Magimel et al., 2020; Tempère et al., 2019). During training, various reference standards representative of aroma and taste terms were presented. Commercially available standards were taken from Sigma-Aldrich (Saint-Quentin-Fallavier, France) and Acros Organics (Geel, Belgium), with purity > 99 % and a food-grade quality. Standards not commercially available were prepared with natural products (Table S3). The sensory properties of the “chemical standards” remained stable throughout the study. The “natural standards” were prepared before each training day to guarantee good aroma quality. The training period was composed of three phases. First, the panellists became familiar with the descriptors and smelled or tasted different standard references in the model solution (12 % ethanol v/v; 5 g/L tartaric acid) and spiked red wine. The references were presented blind, and the panellists were asked to associate them with their descriptor. Second, they were familiarised with the rating of each descriptor individually by tasting red wine spiked with different standard contents. Finally, they were trained to detect and rate descriptors in red wines spiked with several standards at different concentrations. At the end of each session, the panellists were asked to self-correct and to repeat the exercise if deemed necessary. When needed, additional individual sessions were added.

3.2. Conventional QDA

The sensory tests were conducted in the ISVV sensory analysis room (ISO 8589:2010). Still water and crackers were provided as palate cleansers, but cleaning was not enforced. Wines (20 mL) were served at room temperature in ISO glasses (ISO 3591:1977), coded with three digits. The panel received the same information on the wines as the CATA panel, i.e., they would evaluate red Bordeaux wines sold to consumers between 3 € and 8 €, including tax, from the 2018 or 2019 vintage, all appellations and origins combined. Samples were presented one by one. The order of sample presentation was balanced across judges according to the Williams Latin square design approach. A forced break of at least 10 minutes was taken after 10 wines. Twelve descriptors (two for colour, five for aroma, and five for taste and oral sensation) were selected according to the results obtained from the CATA analysis study. The panel gave intensity ratings for the following descriptors: colour intensity, fresh fruits, cooked fruits, green pepper/capsicum, sawdust, woody, sweetness, acidity, bitterness, and astringency on a 10-cm unstructured scale ranging from “not very intense” on the left to “very intense” on the right. Visual aspect was evaluated on a 10-cm unstructured scale ranging from “purple” to “tawny” with the help of a colour chart. Tannic quality was rated on a 10-cm unstructured scale ranging from “silky/velvety” to “dry/harsh”. Finally, panellists rated wine quality on a 10-cm unstructured scale ranging from 0 to 10. One week before the 20-wine analytical session, tasters were asked to evaluate the standard wine in the same conditions. Standard wine and score results (corresponding to the mean responses of panellists) were available during the entire analytical session as a reference. The judges were allowed to taste the standard wine as many times as they wanted throughout the session. The score served as a standard to compare. This approach has been described as improving consensus between panellists (Plemmons & Resurreccion, 1998). The data were collected using FIZZ software (Fizz Acquisition 2.70, Biosystem SAS, 21560 Couternon, France).

4. Statistical analyses

Data obtained from the CATA analysis were analysed with the CATA functionality of XLSTAT software (XLSTAT 2021.1.1, Addinsoft, Paris, France), which created contingency tables containing the CATA data by counting the number of citations for each attribute across the judges for every wine sample. Contingency tables were then analysed with Cochran’s Q test, followed by multiple pairwise comparisons using the critical difference (Sheskin) procedure. The chi-square test was used to evaluate attribute independence. Attributes identified as significant were then subjected to correspondence analysis (CA). Penalty-lift analysis was used to identify the main drivers of quality perceived by the experts. Average quality scores were calculated by considering participants and samples for which the attribute was, or was not, selected (Meyners et al., 2013). Differences between the two arithmetical mean quality values were calculated, and their significance was evaluated by the unpaired t-test assuming equal variance. Hierarchical cluster analysis (HCA) was performed to create wine groups through the visual inspection of the dendrograms obtained. HCA identifies groups of samples with principal sensory characteristics, using input samples coordinates in the 55 dimensions of the CA and considering Euclidean distance, Ward criteria agglomeration, and automatic truncation. The attributes best defining the resulting clusters were obtained by comparing the frequency value in a given cluster and the frequency of the same attribute for all wine samples according to a hypergeometric law. Data obtained from quality scoring were subjected to two-way analysis of variance (ANOVA).

Individual panellists’ repeatability was assessed by using two indices adapted from the following: the index (Ri) proposed by Campo et al. (2008), and the global index (RIi) proposed by Jaeger et al. (2013). The index proposed by Campo et al. (2008) considers the repeatability of the assessors only on the terms that are used to describe the samples, not taking into account those terms that are not selected (i.e., “not applicable” CATA terms), unlike the equation proposed by Jaeger et al. (2013).

To compare results obtained with CATA and conventional QDA, multivariate analysis was performed as described by Campo et al. (2010). Data from QDA were analysed with one-way ANOVA and PCA. The quality score of QDA variables was projected as illustrative quantitative variables in the PCA map. Multiple factor analysis (MFA) was performed using the results obtained with the PCA for conventional QDA and CA for the CATA test. Quality scores from both sensory analyses were used as supplementary data. The RV coefficient was used to calculate the degree of similarity between the two profile datasets. Finally, HCA with the Euclidean distance and the Ward criteria was applied to the factorial coordinates of the wines in the spaces defined by PCA and CA. The attributes best defining the resulting clusters were identified by computing their probability of characterising a cluster. In the case of PCA, this probability is obtained by comparing the mean of one attribute for a given cluster and the mean of the same attribute for all wine samples. Concerning CA, the probability is obtained by comparing the frequency value in a given cluster and the frequency of the same attribute for all wine samples according to a hypergeometric law. All analyses were carried out with XLSTAT software (version 2021.1.1, Addinsoft, Paris, France).

Results and Discussion

1. Chemical analyses

The 143 Bordeaux wines underwent standard chemical composition analysis, which is presented in Table S4. The wines ranged in pH from 3.42 to 3.80. All the wines were technically dry and contained between 12.4 % and 15.5 % (±1 %) vol. ethanol with an average of 13.7 % v/v. TA values ranged from 4.12 g/L to 2.96 g/L (±6 %), and VA did not exceed 0.67 g/L H₂SO₄. Both free and total SO₂ were measured and ranged between 0–33 mg/L and 2–139 mg/L, respectively. Lower contents in organic (44 ± 25 mg/L total SO₂) and sulphite-free (8 ± 3 mg/L total SO₂) wines, well below their legal limits (set at 100 mg/L and 30 mg/L, respectively). Total SO₂ content averages 69 ± 22 mg/L in conventional wines, far below the legal limit set at 150 mg/L for these wines.

2. CATA question analysis and distribution of wines according to their sensory profiles

2.1. Frequency of use of sensory terms and differences among samples

Of the 143 wines analysed, 10 were identified by the judges as having a “bretty” off-flavour (data not shown). Given that the objective of this study is to highlight the different aromatic profiles of entry-level Bordeaux wines, these wines were removed from the statistical analysis so as not to skew the results. Cochran’s Q test was carried out to identify significant differences between samples for each of the sensory terms (Table 2). Among the 62 descriptors proposed to the panel to characterise wines, the term “pepper” was the only one not to be used discriminately. Panellists selected an average of 19–35 % of the terms to describe samples using the CATA questions. The terms used by less than 20 % of the panel to describe at least one wine, such as “dull”, “delicate/elegant”, “dusty”, “bretty”, and “corked”, were also left out from the study, so the final CA map was run over the responses of a total of 48 panellists for 133 wines with 56 terms.

2.2. Panel repeatability

The overall panel repeatability was explored using the projection of wine replicates on the first three dimensions of the CA map (66.9 % inertia; Table S2). As can be seen, most of the wine replicates have similar coordinates, indicating that they are close to each other on the map. Since a different bottle was used for each repetition, the panel was globally repeatable when describing wines. Concerning individual repeatability, the maximum value of the Ri index (which varies from 0 to 1) was 0.85, which corresponds to 85 % of common terms between the two replicates for a subject. The minimum value was 0.32 (median = 0.62). Campo et al. (2008) suggested that the Ri of a trained assessor should be higher than 0.20. In the present research, all the 48 non-trained panellists had an Ri index above this threshold. As expected, values for the global repeatability index (RIi) were higher than those generated by the index proposed by Campo et al. (2008) (Ri). The value of RIi ranged from 0.72 to 0.92 (median = 0.82), suggesting that panellists reliably ticked or did not tick 72–94 % of the CATA terms to describe the products. These results are in accordance with those reported by Jaeger et al. (2013).

2.3. CA and hierarchical clustering

The projection of wines and terms onto a bidimensional CA map (59.43 % inertia) can be seen in Figure 1. The first principal component, PC1 (39.80 %), clearly distinguished wine samples on the right of the bi-plot, whose aromatic opposition could be summarised by “negative” (e.g., “tawny”, vegetal notes, off-flavours, watery mouthfeel with harsh and dry tannins) versus woody terms, complex aroma, sweetness, soft and full, well balanced, full-bodied, and dense mouthfeel with velvety tannins and a strong aftertaste, on the left. The second principal component, PC2 (19.63 %), distinguished wine samples in the top half of the bi-plot perceived as “overripe”, possessing developed or evolved and tired aromas, having notes and flavours of wood and cooked fruits, and for some of them, presenting oxidised notes from those perceived as purple, with fruity and floral notes, and reduction odour, in the bottom half of the CA map. In addition to performing a CATA on the 133 wines, the panellists were also asked to attribute a quality note to each of the wines. Wines scored between 3.25 and 7.29 out of 10, with an average of 5.63. As can be seen on the map, the quality score was correlated on PC1 with woody notes, complex aromas, sweetness, and “positive” mouthfeel terms.

Penalty lift analysis (Figure S2) was conducted using the results of the CATA question and the quality scores of the wines. Mouthfeel and taste were the most important parameters influencing the quality score of wines. Seven terms were appreciated by the panel and were significantly (p-value < 0.01) related to wine quality (full-bodied, well balanced, soft and full mouthfeel, sweetness, velvety tannins, tannic, and medium aftertaste), whereas watery, acidity, a low aftertaste, and dry and tight tannins depreciated wine quality. The taste balance of red wines is based on acidity, bitterness and sweetness, modulated by tannin structure and astringency (Peynaud & Blouin, 1996). The aftertaste or persistence describes the final sensation perceived in the mouth after wine is released from the mouth, involving taste and the perception of astringency. In general, there is agreement about the positive contribution of persistence to quality, or at least their association (Lattey et al., 2010; Niimi et al., 2018; Rinaldi et al., 2021; Souza Gonzaga et al., 2019; Vidal et al., 2018). These findings are in accordance with those of other authors, where quality was mostly correlated with taste and mouthfeel terms in Australian Cabernet-Sauvignon (Lattey et al., 2010), Bordeaux and Bordeaux supérieur (Szolnoki & Hoffmann, 2011) and Spanish ultra-premium red wines (Liu et al., 2023). On the contrary, watery, acidic or burning were negatively correlated with quality (Liu et al., 2023; Szolnoki & Hoffmann, 2011). After mouthfeel and taste perception, fruity flavour was the fourth most important term influencing wine quality (Figure S2). The fruity drivers for quality were ripe, dark and fresh fruits, and varietal and red fruits notes. Woody (vanilla, roasted/coffee) and fresh (minty) notes were also important. In other studies, quality was linked positively with fruity notes, characterised by red and black fruits, and woody aroma (Lattey et al., 2010; Liu et al., 2023; Souza Gonzaga et al., 2019). In particular, Szolnoki and Hoffmann (2011) reported that coffee and oak were the main aroma attributes that improved the perception of red Bordeaux supérieur wines. Finally, higher perceived quality was also linked to a brilliant dark red colour (Figure S2), as described by Niimi et al. (2018) for Australian Cabernet-Sauvignon wines.

HCA was performed to group wines with similar profiles in a multidimensional space. Four main clusters of wines with large sensory differences could be identified among the 133 wines studied, thanks to HCA. Table 3 shows the four clusters, their associated descriptors, and the quality scores for each wine. Cluster 1 contained six wines scored between 3.25 and 5.25 (average = 4.35), and mainly characterised by a tawny colour, developed and tired aroma, and weedy and oxidation notes. Surprisingly, cluster 1 wines had a significantly higher total SO₂ value (Kruskal–Wallis test, p-value < 0.05; data not shown), which should have prevented their oxidation and preserved their colour (cluster 1: 88 ± 16 mg/L; cluster 2: 68 ± 17 mg/L; cluster 3: 65 ± 27 mg/L; 55 ± 25 mg/L; mean in total SO₂ ± standard deviation). The second cluster presented a priori low-quality terms according to expert and consumer judgements (Liu et al., 2023; Sáenz-Navajas et al., 2013), such as vegetal notes, light red colour, bitterness and acidity, watery mouthfeel, and low aftertaste. The 16 wines in cluster 2 scored between 3.94 and 5.31, with an average of 4.62. Clusters 3 and 4 contained 56 and 55 wines, respectively, with a significantly higher quality score (6.01 for cluster 3 and 5.66 for cluster 4; ANOVA p-value < 0.0001). Wines in cluster 3 were characterised by a ripe and complex aroma, dark cooked fruits, and woody notes and flavours, sweetness, a pleasant mouthfeel and a strong aftertaste. As highlighted by Tempère et al. (2019), this wine profile, which was initially appreciated by consumers, seems to be subject to a phenomenon of lassitude and declining appreciation with repeated exposure. Cluster 4 contained wines associated with fruity and green terms, as for group 2, but with a darker colour, harder tannins and a medium aftertaste. CATA has permitted the profiling of 133 wines divided into four main groups. However, one of the limitations of this approach, regularly pointed out in the literature, is that the CATA format does not allow direct measurements of intensity, making discrimination of products with subtle differences more difficult (Ares et al., 2014). The fact of obtaining two groups containing more than 50 wines each is perhaps the result of this limit, unless the wines studied are not very different. Moreover, it is possible that the tasters were influenced by the visual characteristics of the wines. As demonstrated by several studies (Morrot et al., 2001; Nguyen & Durner, 2023; Valentin et al., 2016; Wang & Spence, 2019), such characteristics can affect both the description and the categorisation of wines. Beyond validating the sensory methods used on a large sample, the characterisation of these wines might have ideally been conducted using black glasses. Therefore, we sought to examine the profiling and discrimination capabilities of the CATA method for wine evaluation, compared with those of the conventional DA method. Wines were presented to a panel of five experts, including those in the CATA panel, group by group. They were asked to select representative wines from each group. The selection of 20 wines (identified by asterisks in Table 3) was obtained by a consensus between these experts.

3. Comparison of the discriminative ability of CATA versus QDA

Conventional QDA was performed on the 20 selected wines by 17 semi-trained panellists. One-way ANOVA on the normalised QDA data indicated that descriptors significantly discriminated the samples except “acidity”, “bitterness”, and “tannic quality” (Table S5). Configuration similarity for both sensory studies was evaluated in two ways: (1) by visual comparison of the 20 selected samples and attribute alignment within the MFA plots and by statistical quantification of the similarities in sample configurations with the RV coefficient, and (2) by HCA applied to the factorial coordinates of the wines in the spaces defined by CA (for CATA; Figure S3) and PCA (for QDA; Figure S4), as described by Campo et al. (2010).

MFA of CATA and QDA with statistically significant attributes indicated that the first two factors explained 47.5 % of the variation in sensory data. Overall, the quality scores obtained with CATA and QDA clustered in the left part of the biplot, demonstrating similarities between the two panels in their definition of the quality of a Bordeaux red wine. However, the squared cosine of the CATA quality score (0.79) on axis 1 was higher than the QDA (0.38). While the two panels gave similar overall quality ratings, this result highlights differences in the evaluation of certain wines, as shown by Hopfer and Heymann (2014). The visual inspection of the biplot (Figure S5) disclosed similarities between the CATA and QDA methods, with a clear separation regarding the colour on the first dimension, along with the terms sweetness and pleasant mouthfeel (soft and full, full-bodied), and between oaked and unoaked samples on the second dimension. Vegetal attributes were poorly but better represented on axes 1 and 2, highlighting some differences in the perception of vegetal nuances. All dimensions were used to determine the RV coefficient in order to compare product positioning in the MFA map. The RV coefficient was found to be 0.579 and significant with a p-value < 0.001. The closer the RV coefficient is to 1, the more similar the matrices are. This result shows that there are similarities between the two analyses, but almost as many differences. The most important point of difference relates to fruity terms better represented on different axes and not correlated with each other. However, the first principal components of PCA and CA, as well as their second factor, showed a significant Spearman correlation value (0.583, p < 0.01 and 0.570, p < 0.05, respectively), which indicates that wines were similarly ordered on these two axes.

Concerning HCA, a five-cluster partition was chosen as it allowed a more precise description of wines belonging to each cluster. The five sensory characteristics were vegetal, evolved, fruity, woody, and colourful. Table S6 presents the terms best characterising wines in each cluster. Wines clustered together according to each of the characteristics were not the same between CATA and conventional QDA. Wine 253 was identified by both panels as being evolved, with a tawny colour and not astringent, while wine 627 was characterised by an intense purple colour and pleasant mouthfeel. A relatively good agreement was found for the woody cluster, as it shares five samples: 557, 690, 741, 912, 940. Two wines were common to the cluster characterised by fruity notes (only for CATA) and non-woody aroma (024, 968), whereas only one was common to the vegetal cluster (894).

In this study, the two sensory methods broadly agreed on the main sensory attributes characterising the selected wines, i.e., colour, woody aroma, taste, and mouthfeel. However, fruity and, to a lesser extent, vegetal notes were evaluated differently. As suggested by Campo et al. (2010), the first hypothesis could be related to the lower number of descriptors available in conventional QDA, especially for fruity notes. Secondly, the lack of specific reference materials illustrating fruity and vegetal notes could have made it difficult for panellists to produce a unified interpretation of the descriptors. The descriptor green is an ill-defined concept that can refer to several ideas and has been considered a multidimensional sensory descriptor (Sáenz-Navajas et al., 2021). In addition, differences among panels might be related to their use of different frame references. For instance, CATA results showed that experts used the terms “cooked fruits” and “overripe” indifferently. In the literature, over-ripeness is mainly associated with dried fruit notes. Indeed, dried fruits such as dates or figs and prune juice are often used for panel training (Sáenz-Navajas et al., 2010), whereas the notion of “cooked fruits” is more linked to jam (Campo et al., 2010). Semi-trained panellists refer to the standards provided during the training period, while experts rely on the attributes stored in their long-term memory, whose definition may differ from that of the semi-trained panel. In addition, Lawless (1999) claimed that perceived aromas in wine only vaguely resemble those in standards, and judges, without prompting, often evaluate descriptors based on their applicability rather than their actual intensity.

Finally, the QDA was unable to discriminate the selection of wines studied in greater detail than the CATA analysis. In addition to the nuances raised above concerning panel training and descriptor interpretation, it is possible that the wines studied do not present sufficiently different aromatic profiles to be discriminated. This observation is in line with the results of Szolnoki and Hoffmann (2011), who divided Bordeaux and Bordeaux supérieur red wines into three distinct sensory groups using QDA with a trained panel. In this context, CATA analysis appears to be an interesting tool for efficiently discriminating and categorising a very large number of wines.

4. Number of experts needed to perform CATA analysis

One of the limitations of this approach for assessing a large sample of wines is the recourse to a large number of experts (n = 48). While this masks differences from one individual to another, it may act as a constraint and a difficulty in recruiting so many judges. Whether we would have obtained the same results with a smaller panel remains to be investigated. From the panel of 48 judges (rated 100 % panel), five panels corresponding to 75 % of the initial panel (i.e., 36 judges) were created by randomly selecting the judges (rated 75 %—A to 75 %—E). Similarly, five panels corresponding to 50 % of the initial panel, i.e., 24 judges (rated 50 %—A to 50 %—E) and five panels corresponding to 25 % of the initial panel, i.e., 12 judges (rated 25 %—A to 25 %—E), were also created at random. The 10 wines with previously highlighted defects were removed from this study, but no sorting of descriptors was carried out. To compare the panels, we compared (1) the overall quality scores with a Kruskal–Wallis test, (2) the p-value of the chi-squared test for independence of attributes and wines, (3) whether a descriptor was discriminant or not via the Cochran test, and (4) the cumulative variance on axes 1, 2 and 3 for the graphical representation of the results.

No significant differences were found between the overall assessment scores obtained by panels reduced to 75 % and 50 % (Table S7). On the other hand, significant differences were observed for four wines between the initial panel and the 25 % panels. The chi-square tests for independence, associated with the contingency table built on attributes and wines, were all significant with p-values < 0.0001 (data not shown), indicating that the analysis was quite robust regardless of the number of panellists (Nougarède et al., 2023). When analysing the frequency of descriptor citations using the Cochran test, three terms (delicate/elegant, minty/fresh, pepper) seemed to be influenced when comparing the initial panel with the 75 % panels. The number rose to 15 significantly different descriptors for the 50 % panels and to 26 terms for the 25 % panels (Table S8). Finally, with a panel of 48 judges, the cumulative variance between axes 1 and 2 was 56.7 % and 64.1 % with axis 3 (Table 4). With a panel of 36 judges, the variance was slightly lower (on average 53.4 % for axes 1–2 and 60.9 % with axis 3). The variance was only 49.3 % (axes 1–2) on average and 43.6 % with a panel reduced to 24 and 12 judges, respectively, indicating a loss of information. CATA analysis, therefore, requires a relatively large panel, which could have been reduced to around 40 judges in the present study. Sample sizes used in previous works varied considerably, and the appropriate sample size remains to be established. Jaeger et al. (2014) used around 60 consumers to evaluate different types of food products (crackers, chocolate, beer, mussels) using CATA, but only 14 trained panellists to evaluate chocolate (Jaeger et al., 2013), whereas Ares et al. (2015a) studied milk desserts, orange drinks, raspberry coulis, and white wine, used between 100 and 140 consumers per experimental group. However, they found that the minimum number of consumers needed ranged from 19 to 110, depending on the degree of difference between the samples. Concerning the studies of wine using CATA, they were generally performed by 10 to 60 trained or semi-trained panellists or 30 wine experts, depending on the authors (Nougarède et al., 2023; Panzeri et al., 2020; Rinaldi et al., 2021; Vidal et al., 2018). In the present study, the CATA method proved efficient to describe a large number of wines with a minimum of 40 people, in accordance with previous studies.

Conclusion

The CATA method and cluster analysis presented in this study were successful in differentiating a large number of entry-level red Bordeaux wines, showing that the aroma characteristics of the same varieties but from different appellations do differ. The tested wines could be divided into four groups: evolved, vegetal, fruity, and woody wines with differences in their quality evaluation. In view of this diversity, it appears important to use a method that allows the selection of a representative sample of this diversity. CATA proved to be a robust, easy-to-understand tool understood by wine professionals and provided accurate wine descriptions without any training session. Compared to classical descriptive sensory analysis, it made it possible to study a very large number of wines with a relatively small number of panellists in a short space of time.

Acknowledgements

The authors thank the Conseil Interprofessionnel du Vin de Bordeaux (CIVB) for providing samples and funding. They are also grateful to the wine experts who kindly participated in this study. They also thank Dr Ray Cooke for proofreading the manuscript.