CoffeariumBeta
  1. Plant
  2. Farm
  3. Process
  4. Trade
  5. Roast
  6. Brew
  7. Cup
Science · Cupping

Sensory science methods

How sensory scientists test coffee beyond cupping – difference tests such as the triangle test, trained descriptive panels, consumer liking scales, and the coding, serving order and statistics that make results trustworthy.

Across the chain: Science

Source-checkedby Coffearium verification pass· Last source-checked 15 Oct 2026· Prepared with AI assistance· 17 sources (10 tier A)· 1 contested· Updated 15 Oct 2026
Reading depth
Contents
  • Sensory evaluation uses people as the measuring instrument and picks the test by the question – are two coffees different, what are they like, or are they liked?1
  • In a triangle test (ISO 4120) each assessor gets three coded samples, two of them alike, and must pick the odd one; a guess is right one time in three.2,5
  • The minimum number of correct answers comes from published binomial tables – for example 8 of 12 assessors, or 13 of 24, at the 5% significance level.4
  • Descriptive analysis uses a trained panel to rate the intensity of agreed attributes; affective tests ask consumers how much they like a coffee.1,13
  • SCA Standards 103-2024 and 104-2024 leave the selection and training of tasters, and the statistical analysis of results, outside their scope.10,11

What sensory science measures#

Sensory evaluation uses people as the measuring instrument. According to a 2017 textbook chapter on coffee, its aim is to understand products in order to optimise them and to drive consumer choice.1 It borrows from physiology, psychology and statistics.1

The method follows from the question. The same chapter names four kinds of question:1

  • Are two coffees different? Difference tests, also called discrimination tests.
  • What does a coffee taste like, and how strongly? Descriptive tests.
  • Does a coffee match a standard? Grading and quality-control tests.
  • How much do people like it? Affective, or hedonic, tests.

Each test is a small experiment. The tester decides what to change, what to measure and which conditions can be controlled. The most accurate tests include a control as a baseline.1 Standardised tests are usually preferred because they come with set statistical procedures.1 How the senses themselves work is covered in How we taste coffee.

Difference tests#

The triangle test. Each assessor receives three coded samples. Two are the same and one is different, and the assessor must name the odd one.5,1 It is a forced-choice procedure: an answer is required even when the assessor is unsure.2 The International Organization for Standardization (ISO) describes it in ISO 4120. The test works whether the difference lies in one attribute or in several.2

A · One triangle-test tray12345B · The six serving orders6
  1. Two alike. Two cups hold the same coffee.5
  2. Blind code. Each cup carries a three-digit code instead of a name, so the assessor cannot tell the samples apart by label.1,15
  3. Odd sample. One cup holds the other coffee. The assessor must say which cup it is, even when unsure.2,5
  4. Water. Water for rinsing the mouth between samples; one coffee study served filtered water at room temperature.12
  5. Tasting order. In the same study the three cups stood on one tray and were tasted from left to right.12
  6. Six orders. Two products can be set out in six ways – AAB, ABA, BAA, BBA, BAB and ABB – which are randomised across assessors.5
A triangle-test tray and the six possible serving orders of two products. Our own drawing. Schematic, not to scale.

Why so many assessors? Someone who only guesses still picks the odd cup one time in three.5,12 One correct answer therefore proves little. The test is repeated with many assessors, and a published table gives the number of correct answers needed.1,4 With 12 assessors, 8 must be right for a difference at the 5% level; with 24 assessors, 13.4

The duo-trio test. Here the assessor first tastes a known reference, then two coded samples, and picks the one that matches the reference.6 ISO 10399 calls it statistically less efficient than the triangle test, but easier for assessors.3 A guess is right one time in two, so more correct answers are needed for the same confidence (our own reasoning from the procedure; see the deep dive).

Difference tests have a cost. Tasting many sets of samples tires the palate, and tired assessors judge less well.1 They are still worth running first: describing two coffees in detail is pointless if a panel cannot tell them apart.1

Describing a coffee with a trained panel#

Descriptive analysis answers what a coffee tastes like and how strongly. It uses a panel trained with attribute references.1 According to the textbook chapter, quantitative descriptive analysis (QDA) is probably the most commonly used form.1 The panel first agrees on its words, either from an existing vocabulary such as the World Coffee Research Sensory Lexicon or from its own sessions.1 It then rates the intensity of each attribute in every sample.1 ISO 13299 gives general guidance for building such a sensory profile.7

In one coffee study, 12 volunteers were trained in five sessions over 2 weeks. They rated 26 attributes on a 15 cm line scale.13 Training of this kind takes time, so full descriptive analysis is kept for research, product standards and products where precision matters.1 ISO 8586 sets out how sensory assessors are selected and trained.8

A quicker method is check-all-that-apply (CATA). Tasters tick every word on a list that fits the sample, without rating intensity.1,9 It needs much less training than a descriptive panel, but many more participants.9,1 The words used to describe coffee are covered in Flavour language.

Measuring liking#

Affective tests ask how much people like a coffee. They are usually run with consumers who might buy the product.1

  • Hedonic scale. Consumers rate each sample on a 9-point scale from “dislike extremely” to “like extremely”.1,13
  • Preference and ranking. Consumers choose the preferred sample of two or three, or rank several samples from most to least preferred.1
  • Just-about-right (JAR) scales. Consumers say whether a property, such as strength, is too weak, too strong or just about right.1

Liking and difference are separate questions. In a 2025 study, 42 of 60 coffee-interested participants picked the odd sample in a triangle test of freshly roasted against matured coffee.12 Of those 42, however, 22 preferred the matured coffee and 20 the fresh one – no significant preference.12

Fair testing: codes, order and rinsing#

According to the SCA, tasters can be easily biased if they are given extrinsic information about a coffee.9 To limit bias in tasting, samples are labelled with three-digit codes, served in a random or balanced order, repeated, and judged by each assessor alone.1

  • Blind coding. A code hides the sample’s identity. Three-digit codes are the textbook rule and appear in coffee studies.1,15
  • Order. Serving order can itself bias judgements, so orders are randomised or balanced across assessors.1,5
  • Carry-over. A lingering taste from one sample can change the next. Designs balanced for carry-over aim to control this.17
  • Rinsing. Water, and sometimes unsalted crackers, are served between samples.12,13
  • Temperature and timing. Coffee flavour changes as it cools, so preparation and tasting are timed.1
  • Seeing. One study lit its tasting booths with red light to hide colour differences between coffees.13

How it differs from cupping#

Cupping is the coffee trade’s own tasting method, described in Cupping. It relies on industry experts who judge quality on a quality scale. A review names this as what sets it apart from descriptive analysis.16 SCA Standards 103-2024 and 104-2024 leave the selection, training and qualifications of tasters outside their scope.10,11 Both also leave out the statistics needed to interpret their data with confidence.10,11

Since 2024, the SCA’s Coffee Value Assessment (CVA) borrows from sensory science. The SCA specifies that intensity is rated on 15-point scales and that descriptors are ticked on CATA lists.10 According to the SCA, 15-point scales are the usual choice in traditional descriptive methods.9 Its affective part uses a 9-point scale, the format the SCA calls the most common affective measure in food science.9 The uniformity box on the affective form is a small difference test.11 According to the SCA, a cupping group aligning its scores is not doing what sensory science calls calibration. Calibration means training a panel on set references for one specific test.9

Reference values#

What each method asks and how it is scored, as given in the sources. Minimum numbers of correct answers are those of the FAO table for the triangle test.4

Parameter Value Source
Triangle test Three coded samples, two alike; identify the odd one; forced choice ISO 4120; SSP2,5
Triangle test, chance level 1/3 (33.3%) SSP; Zimmermann et al.5,12
Triangle test, serving orders Six (AAB, ABA, BAA, BBA, BAB, ABB), randomised across assessors SSP5
Triangle test, limits Products must be homogeneous; ISO 4120 calls the test less suitable for products with strong carry-over or lingering flavours ISO 41202
Minimum correct, α = 0.05 n = 6: 5 · 9: 6 · 12: 8 · 15: 9 · 18: 10 · 21: 12 · 24: 13 · 27: 14 · 30: 15 · 33: 17 · 36: 18 · 42: 20 · 48: 22 · 54: 25 · 60: 27 FAO4
Minimum correct, α = 0.01 n = 6: 6 · 9: 7 · 12: 9 · 15: 10 · 18: 12 · 21: 13 · 24: 15 · 27: 16 · 30: 17 · 33: 18 · 36: 20 · 42: 22 · 48: 25 · 54: 27 · 60: 30 FAO4
Minimum correct, α = 0.001 n = 9: 8 · 12: 10 · 15: 12 · 18: 13 · 21: 15 · 24: 16 · 27: 18 · 30: 19 · 33: 21 · 36: 22 · 42: 25 · 48: 27 · 54: 30 · 60: 33 FAO4
Duo-trio test Identified reference, then two coded samples; pick the match SSP6
Duo-trio, reference schemes Constant reference (one product familiar, e.g. regular production) or balanced reference ISO 103993
Duo-trio vs triangle Statistically less efficient, easier for assessors ISO 103993
Descriptive panel, one study 12 panellists; five sessions over 2 weeks; 26 attributes; 15 cm unstructured line scale; eight coffees in triplicate Frost et al.13
Hedonic scale 9 points, “dislike extremely” to “like extremely” Thomas et al.; Frost et al.1,13
JAR scale 5 points, “much too low” to “much too high” (one study); 5-point JAR questions (another) Guinard et al.; Frost et al.15,13
Sample codes Three digits Thomas et al.; Guinard et al.1,15
CVA descriptive scale 15 points, total intensity of each cupping section SCA 103-2024; CVA document10,9
CVA affective scale 9 points, impression of quality CVA document9

We recalculated every minimum in the table from the binomial distribution with a chance level of 1/3; all values agree with the FAO table.4 The table is printed in an FAO technical paper on fish quality, but the statistics do not depend on the product.

The statistics of a triangle test#

Under the assumption of “no difference”, every assessor has a one-in-three chance of naming the odd sample.5 The number of correct answers among n independent assessors then follows a binomial distribution with p = 1/3. The probability of x or more correct answers by chance alone is

P(X ≥ x) = Σ from k = x to n of C(n, k) · (1/3)^k · (2/3)^(n − k)

The minimum in the table is the smallest x for which this probability does not exceed the chosen significance level α. If the observed count reaches the tabled value, the assumption of “no difference” is rejected.4 This is our own statement of the calculation; it reproduces the FAO values.4

010203040506005101520253035Number of assessors (n)Minimum correct answersα = 0.0014 α = 0.014 α = 0.054
Data table
SeriesNumber of assessors (n)Minimum correct answersSource
α = 0.05654
α = 0.05964
α = 0.051284
α = 0.051594
α = 0.0518104
α = 0.0521124
α = 0.0524134
α = 0.0527144
α = 0.0530154
α = 0.0533174
α = 0.0536184
α = 0.0542204
α = 0.0548224
α = 0.0554254
α = 0.0560274
α = 0.01664
α = 0.01974
α = 0.011294
α = 0.0115104
α = 0.0118124
α = 0.0121134
α = 0.0124154
α = 0.0127164
α = 0.0130174
α = 0.0133184
α = 0.0136204
α = 0.0142224
α = 0.0148254
α = 0.0154274
α = 0.0160304
α = 0.001984
α = 0.00112104
α = 0.00115124
α = 0.00118134
α = 0.00121154
α = 0.00124164
α = 0.00127184
α = 0.00130194
α = 0.00133214
α = 0.00136224
α = 0.00142254
α = 0.00148274
α = 0.00154304
α = 0.00160334
Minimum number of correct answers in a triangle test for a significant difference, by number of assessors, at three significance levels; values from the FAO table, a selection of its rows.4 For six or fewer assessors the table gives no value at α = 0.001.

The lines rise more slowly than the number of assessors. With 12 assessors, 8 correct answers (67%) are needed at α = 0.05; with 60, 27 (45%).4 A larger panel can therefore detect a smaller difference. The table also has no value at α = 0.001 for six assessors: even six correct answers out of six are too likely by chance.4

Other ways to compute it. The Society of Sensory Professionals (SSP) gives a chi-square calculation as one way to evaluate a triangle test.5 A coffee study judged its triangles from Z-scores and P values.13 Minimums stated in published studies can differ from the exact table; see the contested item below.

How many discriminate? A significant result does not say that every assessor perceived the difference. One 2025 study reported 42 correct answers out of 60 (70%) and estimated the proportion of true discriminators at 55.0% (95% confidence interval 41.3–68.7%).12 The estimate corresponds to correcting 70% for guessing: (0.70 − 1/3) ÷ (2/3) = 0.55 (our arithmetic).

The duo-trio test. Here a guess is right half the time, because the assessor chooses between two samples, one of which matches the reference.6 By the same binomial calculation with p = 1/2 (our own), 24 assessors need 17 correct answers at α = 0.05, against 13 in a triangle test. This is the practical meaning of ISO 10399’s statement that the duo-trio test is statistically less efficient.3 ISO 10399:2017 was withdrawn in May 2026 and replaced by ISO 10399:2026 (as of October 2026).3

Difference or similarity. Both ISO 4120 and ISO 10399 cover testing for a perceptible difference and testing for similarity, for example after a change of ingredients, processing, packaging, handling or storage.2,3 Both standards also name the selection, training and monitoring of assessors as a use.2,3 The settings and tables for similarity testing are in the standards’ full texts, which were not read for this entry.

Descriptive analysis in detail#

Variants of quantitative descriptive analysis include QDA and the Spectrum method.1 In QDA the panel first fixes a consistent vocabulary – taken from an existing lexicon or developed over several sessions – then rates each attribute’s intensity; the result is a flavour profile of the sample.1 Results are often compared with principal component analysis, which places samples and attributes on one map.1 ISO 13299 lists uses of sensory profiling that include developing or changing a product, defining a product or trading standard by its sensory attributes, shelf-life studies, and relating perceived attributes to instrumental data or to consumer acceptability.7

Panel selection and training. ISO 8586:2023 specifies criteria for selecting, and procedures for training, trained and expert sensory assessors. It supplements the general guidance of ISO 6658.8 In a 2019 coffee study, the descriptive panel spent its first two training sessions generating attributes and the next three agreeing terms and references.13 Panellists then rated eight coffees in triplicate over four sessions.13 According to the SCA, a trained panel in sensory science is calibrated to use a set of sensory references accurately in a specific test environment and protocol.9

Rapid methods. Several faster alternatives to classic descriptive analysis exist; they can be carried out by experts, trained assessors or untrained consumers.1 In CATA, products are presented one at a time in a Latin square or randomised design. The counts of how often each word was ticked for each product form a frequency table, analysed by correspondence analysis.1 Its main drawback is that the list must be fixed in advance.1 Coffee studies have tested CATA results with Cochran’s Q test.13,15 Other rapid methods include sorting tasks, which have been applied to milk coffee drinks, and flash profiling.1

Affective tests in detail#

The 9-point hedonic scale was the liking scale in a 2019 study of 85 consumers, alongside 5-point JAR questions and a CATA question.13 JAR data are commonly read with penalty analysis. It shows how much liking, or here rated quality, drops when a property is judged too low or too high.15 In a 2026 study of 56 coffee experts and 12 coffees, roast and colour judged too dark were the most damaging to quality.15

The CVA uses the 9-point format, but its labels speak of impression of quality rather than liking.9,11 According to the SCA, its own research showed that cuppers could act as reliable descriptive panellists, but also suggested that personal preferences divide them into distinct clusters.9 On that account, preference can differ between tasters, even among professionals.

Designs against order and carry-over effects#

Each sample can change the perception of the next one, for example through a lingering taste. Sensory experiments therefore often use designs balanced for carry-over effects.17 A 2004 statistical report found that designs only nearly balanced for carry-over still gave conservative variance estimates, even with large carry-over effects.17

In practice coffee studies combine several tools. One descriptive study used a randomised Williams Latin square block design to control carry-over, and randomised the order of its triangle sets by session.13 A 2026 expert study printed a randomised serving order on each ballot.15 In a 2025 triangle test, assessors who tasted together were excluded from the analysis.12

Coffee studies that used these methods#

Study Question Method Assessors Result
Zimmermann et al. 2025 Can consumers tell fresh from matured roasted coffee, and which do they prefer? Triangle test (ISO 4120), then a preference choice among those who answered correctly; booths; 60 ± 5 °C; 140 mL in 180 mL cups 60 untrained, coffee-interested 42 of 60 correct (p < 0.001); preference 22 to 20, not significant12
Frost et al. 2019 Does the shape of a drip-brewer basket change the coffee? Six triangle tests; QDA; consumer test with 9-point hedonic, JAR and CATA; about 65 °C 45 (triangles); 12 trained (QDA); 85 consumers At the fine grind 23 of 45 discriminated flat-bottomed from semi-conical baskets (P = 0.024, as reported)13,16
Batali et al. 2020 Does brew temperature matter at fixed strength and extraction? Sensory profiling of drip brews at 87, 90 and 93 °C Not stated in the abstract Strength and extraction changed the profile; temperature had no appreciable impact14
Guinard et al. 2026 Can an overall rating plus JAR and CATA replace the 100-point score? Overall quality rating, 5-point JAR, CATA; three-digit codes; randomised order 56 experts, 12 coffees Roast and colour judged too dark were the most damaging to quality15

Cupping and sensory science side by side#

Aspect Cupping (SCA CVA) Sensory science methods
Who tastes Industry experts judging quality16 Assessors chosen for the test: trained panels for description, consumers for liking1
Selection and training Outside the scope of SCA 103 and 10410,11 Specified in ISO 85868
Statistics Outside the scope of SCA 103 and 10410,11 Part of every standardised test1
Difference testing Uniformity box, a small discriminative test11 Triangle, duo-trio and paired tests1
Description Total intensity of seven sections on 15-point scales, plus CATA10 Intensity of each agreed attribute (QDA), or CATA1
Preference 9-point impression of quality, own or market preference9 Consumer liking on hedonic scales, preference, JAR1
“Calibration” Better called alignment, according to the SCA9 Training on references for a specific test9

SCA Standard 104-2024 also lists taster selection, training and qualifications as outside its scope.11

Open questions#

  • Exact versus approximate tables. Coffee studies compute triangle-test significance in different ways, by tables, Z-scores or their own calculation.12,13 Their stated minimums can differ from the exact binomial table.4,12
  • Cuppers as panellists. The SCA reports that cuppers can act as consistent descriptive panellists, citing its own research.9 How cupping descriptors compare with a trained panel’s is discussed in Cupping.
  • Similarity testing in coffee. No coffee study opened for this entry used the similarity mode of ISO 4120 or ISO 10399, for example to show that a cheaper process gives the same cup.

Key studies#

Zimmermann et al., 2025 – Fresh versus matured roasted coffee. Sixty coffee-interested participants tasted freshly roasted and matured coffee in a triangle test following ISO 4120, then chose the one they preferred. Most (42 of 60) picked the odd sample, but preference split almost evenly – a clear example of difference and liking as separate questions.

European Food Research and Technology 251(11), 3839–3845 · doi:10.1007/s00217-025-04873-0

Frost et al., 2019 – Basket geometry. Combined three methods in one study: triangle tests with 45 assessors, a trained descriptive panel of 12 with 26 attributes, and 85 consumers rating liking on a 9-point hedonic scale. The paper documents panel training, a Williams Latin square to control carry-over, and red light to hide colour differences.

Journal of Food Science · doi:10.1111/1750-3841.14696

Batali et al., 2020 – Brew temperature. Drip-brewed coffees were matched for strength and extraction yield at 87, 90 and 93 °C and profiled. Strength and extraction changed the sensory profile; brew temperature on its own had no appreciable effect within that range.

Scientific Reports 10 · doi:10.1038/s41598-020-73341-4

Guinard et al., 2026 – Rating plus JAR and CATA. Fifty-six coffee experts assessed 12 coffees with an overall quality rating, five 5-point JAR scales and a CATA list, served blind under three-digit codes in randomised order. Penalty analysis showed which deviations from “just about right” cost the most quality.

Foods 15(4), 678 · doi:10.3390/foods15040678

Contested
  • Minimum correct answers for 60 assessors: a 2025 coffee study, citing ISO 4120, states that 26 correct answers were required at α = 0.05 and 30 at α = 0.00112; the FAO table gives 27 and 334, as does our own check with the exact binomial distribution. The table in ISO 4120 itself was not read for this entry. The difference does not change the study’s result: 42 of 60 assessors answered correctly.12
Share or cite this entry

Sources

  1. BThomas, E.; Puget, S.; Valentin, D.; Songer, P. (2017). Sensory Evaluation – Profiling and Preferences. In: Folmer, B. (ed.), The Craft and Science of Coffee, Academic Press, ch. 18, pp. 419–456. doi:10.1016/B978-0-12-803520-7.00018-9 · other copies
  2. AInternational Organization for Standardization (2021). ISO 4120:2021 Sensory analysis — Methodology — Triangle test. ISO.
  3. AInternational Organization for Standardization (2017). ISO 10399:2017 Sensory analysis — Methodology — Duo-trio test. ISO.
  4. AHuss, H. H. (1995). Triangle test for difference: critical number (minimum) of correct answers. In: Quality and quality changes in fresh fish, Appendix A. FAO Fisheries Technical Paper 348, Food and Agriculture Organization of the United Nations, Rome.
  5. BSociety of Sensory Professionals Triangle Test. Society of Sensory Professionals, Sensory Wiki.
  6. BSociety of Sensory Professionals Duo-Trio Test. Society of Sensory Professionals, Sensory Wiki.
  7. AInternational Organization for Standardization (2016). ISO 13299:2016 Sensory analysis — Methodology — General guidance for establishing a sensory profile. ISO.
  8. AInternational Organization for Standardization (2023). ISO 8586:2023 Sensory analysis — Selection and training of sensory assessors. ISO.
  9. A/BSpecialty Coffee Association (writing team Fernández-Alduenda, M.; Ionescu, K. E.; Giuliano, P.; Rugolo, J.; Delrue, R.) (2024). A System to Assess Coffee Value: Understanding the Specialty Coffee Association’s Coffee Value Assessment (June 2024). Specialty Coffee Association.
  10. A/BSpecialty Coffee Association (2024). SCA Standard 103-2024 Coffee Value Assessment: Descriptive Assessment. Specialty Coffee Association.
  11. A/BSpecialty Coffee Association (2024). SCA Standard 104-2024 Coffee Value Assessment: Affective Assessment. Specialty Coffee Association.
  12. AZimmermann, Y. C.; Schwarz, S.; Lachenmeier, D. W. (2025). Post-roast maturation and coffee quality: a sensory perspective on freshness. European Food Research and Technology 251(11), 3839–3845. doi:10.1007/s00217-025-04873-0 · other copies
  13. AFrost, S. C.; Ristenpart, W. D.; Guinard, J.-X. (2019). Effect of Basket Geometry on the Sensory Quality and Consumer Acceptance of Drip Brewed Coffee. Journal of Food Science. doi:10.1111/1750-3841.14696 · other copies
  14. ABatali, M. E.; Ristenpart, W. D.; Guinard, J.-X. (2020). Brew temperature, at fixed brew strength and extraction, has little impact on the sensory profile of drip brew coffee. Scientific Reports 10. doi:10.1038/s41598-020-73341-4 · other copies
  15. AGuinard, J.-X.; Lim, L. X.; Elliott, B.; Cotter, A. (2026). Validation of the Coffee Cuality™ Method for the Expert Assessment of Coffee Sensory Quality. Foods 15(4), 678. doi:10.3390/foods15040678 · other copies
  16. ASeninde, D. R.; Chambers, E., IV (2020). Coffee Flavor: A Review. Beverages 6(3), 44. doi:10.3390/beverages6030044 · other copies
  17. BKunert, J.; Sailer, O. (2004). On Nearly Balanced Designs for Sensory Trials. Technical report, SFB 475, Universität Dortmund. doi:10.17877/DE290R-15240 · other copies

Tier A · authoritative · B · trade and expert · C · general. Supporting quotes for every source are in the data version of this entry. If a publisher’s site refuses access, try the doi link or “other copies”.

Suggest a change

On Sensory science methods. Tell us what you would change and why. A source that supports it helps most.