Brands ‘live’ by the associations that consumers have of them; the associations stored in long-term memory define how a brand is perceived. Brand associations come in different forms: Firstly, there are the more concrete and tangible associations that refer to the products as well services by a brand (e.g., physical and functional attributes, possible uses), but importantly, there also are the more abstract and intangible associations of the brand concept (e.g., personality, emotions, time and place, typical or target users). Varied methodologies are available, qualitative and quantitative, for studying and mapping the brand associations held by consumers, set to explore, reveal and measure them at different levels.

The methodology reviewed here was developed by Daria Dzyabura and Renana Peres (2021) [*]. Their approach is more than interesting: the methodology they present, its implementation and examples from their detailed results, are innovative, illuminating, and also promising for marketing practice. Three pillars are most salient for highlighting: (1) The methodology extends an originally qualitative renowned approach for revealing associations into quantitative, scalable measurement and mapping of brand associations; (2) The methodology is based on visual elicitation of brand associations, using photo images, in following the qualitative research approach adopted as its foundation; (3) The advanced analyses applied entail AI-based learning algorithms, clustering, and a probabilistic analytic process and model construction. The methodology is named Brand Visual Elicitation Platform, B-VEP in short.

The methodology yields a rich spectrum of varied associations. The brand associations, organised as topics, cover objects, emotions, sceneries, activities, experiences and more abstract concepts. The associations seem to be less concerned with the physical or functional attributes of products and driven towards more creative and emotional brand conceptions. This outcome is attributed primarily to the elicitation of associations through visual photographic images rather than verbal (text) expressions. It corresponds to the ability of humans to process images and pictures, and to the notion that humans often conceive their thoughts and feelings in visual images, and possibly store them as ‘pictures’.

The inspiration comes mainly from the qualitative research methodology originated by Gerald Zaltman, the Zaltman Metaphor Elicitation Technique (ZMET). The B-VEP methodology is by no means a substitute to ZMET but rather an optional route to execute a similar approach to visual-driven thinking and ideation on a larger scale than is feasible in the qualitative technique. Critically, B-VEP cannot reach the level of depth of conceptions, and brand metaphors, that can be developed through interviewing and probing on the basis of collages of images offered by participants in application of ZMET. On the other hand, B-VEP enables the generation of a spectrum of free-thought brand associations that usually cannot be reached through more traditional verbal-driven elicitation techniques.

In the elicitation process of B-VEP, respondents are presented with a large collection of photo images (sourced from Flickr, ~100,000 photos). They can browse the photos on the right-side panel of the screen, choose images reflecting any aspects of the brand that come to their minds, and pass the chosen images to the left-side panel of the screen. Thus, respondents create (online) collages of images they relate or link with the focal brand studied in an unaided manner. Dzyabura and Peres report that they collected in their research 4,743 brand collages from 1,851 respondents (3,937 collages finally approved), for 303 national US brands from nine categories (beauty, beverages, cars, clothing, department stores, food & dining, health products & services, home design & decoration, and household cleaning products). This resulted in 15.6 collages on average per brand, with 11.45 photos on average per collage (taking 8 minutes on average to generate a collage — time was also used as a quality screening criterion).

It should be noted that in ZMET, participants collect, during a preparation period, their own set of images — photos, drawings or in other pictorial styles — which they may create themselves (e.g., taking photos) or pick-up from print publications, online websites, and other sources. The use of a unified pool of images, where the researchers could also apply descriptors (based on tag words added by Flickr users) to collate the appropriate image pool, appears necessary for carrying-out the quantitative, scalable analyses. Yet, the image pool is large and diverse enough to leave respondents much room to match images with their associations. Additionally, respondents in B-VEP are given the option to search relevant photos by key words to aid them in selecting photos. On the one hand, it raises concern that respondents might again draw associations as verbal cues rather than visually from the images. On the other hand, it could be difficult for respondents to find photos reflecting their thoughts and feelings merely by freely browsing a very large photo collection. Hence, the use of a search utility seems reasonably justifiable as an aid to retrieve images that respondents fail to come across by browsing.

It seems that Dzyabura and Peres were concerned by the less desirable implications that could arise from searching photos by keywords. They performed special checks for making persuasive arguments to assure that the collage is primarily visual-driven, and searching is applied as an initial aid to browsing.

• Keywords used by respondents for search are likely to produce subsets of several dozens to hundreds of photos (according to tag or label words in Flickr); therefore, the search results help in directing a respondent to relevant candidate images, but leaving sufficient choice space to focus on an appropriate image more closely reflecting the image one has in mind (e.g., “family” retrieves 318 photos, “nature” 3,621 photos, “happy” 210 photos, “laptop” 46 photo images). It is noted that respondents tend to repeat using a limited set of words more frequently (e.g., the top 30 search words, 0.5% of unique words, account for 17% of total searches, “family” being in the lead). Eventually, in 17.5% of the collages the search tool was not employed.

In the first stage of analysis, the researchers had to interpret the content of images selected in collages: identify objects, and possibly infer whatever more meanings the content might imply. Dzyabura and Peres utilised an image tagging tool, based on a pre-trained learning model commercially available from AI firm Clarifai, in order to assign descriptive tags (labels) to images. The model (a form of deep convolutional neural networks) is trained on millions of photos with more than 10k concepts in the learning process of matching images with words (also see image recognition by Clarifai). Importantly, it allows to classify or index images according to semantic tags (as opposed to basic visual-graphic features), ascribing to objects, scenery, actions, emotions, adjectives, and other visual elements. This capability is essential for later linking the visual content meaningfully to brand characteristics.

Each photo is assigned the top 20 tags with highest confidence scores. However, ‘reviewing’ the tags for collages of images could be just too tedious and not very productive — remember that this was only an intermediary stage. The next task was to better organise the tags (labels), by grouping them into meaningful, sensible subsets, that would also enable more efficiently to characterise and comprehend the brand-related collages. Dzyabura and Peres describe the task in the second stage as “extracting associations from collage tags“. The practical implication is constructing a set of association topics. A topic specifies or signifies a certain kind of association (e.g., “Flowers (romantic)”, “Glamour”, “Wedding”, “Family”, “Fashion”), wherein the association is explicated by a subset of characteristic tags. An icon was chosen to visually and concisely represent each association (topic). The collages will subsequently be described summarily by the associations (i.e., at the topic level).

• The process followed for executing this task was probabilistic and multi-step. The modelling method, known as latent Dirichlet allocation (LDA), was adopted from the area of text-mining documents, only that in this case each ‘document’ is relatively small — each collage, containing 11-12 photos on average (with 20 tags per photo), is defined as a document. Hence, the researchers applied a modified version of the methodology labeled ‘guided LDA’. It will suffice to explain here that each topic is defined as a distribution over words in the vocabulary and each document (collage) as a distribution over topics. An unsupervised learning model generated tag-word structures, submitted as input to a clustering technique (k-means) for grouping tag items into topic clusters. Important aspects of this method are that: (1) tags are assigned with probability to topics; (2) tags can belong in more than one topic; and (3) it sets the ground for relating topics to collages.

Overall, 150 association topics were identified. Research assistants interpreted the meanings of subsets of tags (most probable to be included in a topic) for naming them and attaching icons. For example, the association topic Cityscape has tags (with their probabilities) such as downtown (9.4%), cityscape (9.4%), skyline (7.9%), skyscraper (7.4%), tower (5.0%), and bridge (4.4%). As another example, the Running association is composed of tags such as athlete (7.8%), runner (6.4%), race (5.9%), fitness (5.7%), and marathon (5.7%). Thereby association topics are getting form and meaning, ready to move forward to describe collages, and thereof respective brands, by those associations.

The concluding task was to construct a ‘profile’ of associations for each brand studied, based on the collages created for the brand. Each of the 150 topics can occur in a collage with a probability; these probabilities of association topics are averaged across collages for a brand, whereby the associations most probable to occur (Top 5) are selected to describe or characterise a brand.

Dzyabura and Peres display in their article a mapping of brand associations for three deodorant brands in the beauty category (AXE, Degree, Secret). Figure 1 in the article provides also the tags included in each association to help in comprehending and appreciating each association (topic). This form of mapping is illuminating, instructive and even intriguing. It can tell marketers about the perceived strengths of each brand, what makes it different from competing products, who could be the most appropriate target users for the brand, and more.

• Degree, for example, is perceived very much as sports-related, suitable in particular during practice of Running, Fitness, and Ball Sport (& Sports overall), which involve action. A top association with Degree is Water (e.g., wet, clean, rain, purity, bubble, droplet), but in this case might hint at battling sweat, refreshing and staying clean. Axe has different connotations, associated primarily with Fashion, Flowers-Romantic, Urban Youth (e.g., face, hair, ‘finelooking’, adolescent, guy), and conspicuously Astronomy (e.g., space, galaxy, exploration, science); another top association is Bodybuilding — the brand appears more social, masculine and inspirational (or rather aspirational).
• A separate table of results (Table 2) in the article lists the five most frequently occurring associations for each of the beauty brands studied. One may consider, for instance, that the association Water for brands such as Crest, Colgate, Dial Soap or Head & Shoulders could have a different implication for consumers than in regard to the brand Degree. It is also interesting to notice that for a number of brands their ‘first’ occurring association is Glamour (e.g., Always, Avon, Clinique, Maybelline, Revlon). Attention-grabbing associations reflect impressions of brands beyond their products (e.g., Fruits, Streams, Cat, Sailing, Erotic, Ocean, Frosty).
• Still, consumers cannot avoid selecting photo images that reflect ‘generic’ descriptors as their associations with a brand (e.g., product type, category or industry); they show up in summaries of associations at the category level (Table 3).

The researchers offer results from a few more analyses and suggest further applications to demonstrate the possible practical implications of the mapping of associations for brand management.

• For demonstrating the richness of associations elicited by images, the researchers compared them with associations generated in another study in which respondents wrote text-free paragraphs for describing focal brands. It was found that respondents had difficulty to provide detailed descriptions, and they often resembled what one would find in product reviews. Dzyabura and Peres concluded that the descriptions were not distinctive, yielding terms (representing associations) that did not seem unique to brands (e.g., too general on product usage, functional, evaluative).
• In the main research, respondents were asked to explain in words the collages they created; terms extracted from these verbal explications included more intangible attributes that relate more to brand associations (e.g., “cool”, “happy”, “fresh”, “confident”, “strong”) — it is not as rich as the image-extracted topics, but already advantageous to verbal-only elicitation, since the explanations were based on the collages of images chosen by respondents (i.e., visually-driven).
• Creating a Prototypical Brand Collage — Designers and managers concerned with brand communications may index photos from different sources (outside the image pool in this research) to create a repository of photos that most closely reflect the brand associations revealed by B-VEP (i.e., by comparing the indexes of new photos to the topics and corresponding tags from B-VEP).
• Brand managers may further improve the visual representation of a brand by establishing its correspondence to brand personality and brand equity characteristics (tested by the researchers on the basis of measures of respondents’ ratings on those brand attributes). Dzyabura and Peres show examples of brand associations that have the most positive and most negative correlations with personality and equity attributes (i.e., most and least fit to support those attributes).
• Uniqueness and Commonalities — The researchers show how, on the one hand, managers may identify associations (topics) that are most and least likely to belong in a focal brand (through its collages) relative to competing brands in the same category, offering points of strength and differentiation; on the other hand, managers may identify those brand associations in a given category that are most similar with associations of brands in other categories as a basis for potential cooperation and collaboration (e.g., co-branding).

Use of images has been demonstrated to successfully disrupt well-rehearsed narratives, revealing hidden, unarticulated ideas (p. 63)

The mapping of brand associations generated through the B-VEP methodology by Dzyabura and Peres clarifies the meanings of brand perceptions and their significance to consumers. It further presents opportunities for additional courses of action in brand management, creative and strategic. These possibilities are based primarily on the advantages of visual elicitation of associations with collages of photographic images (see quote above), enabled by applications of some advanced modelling and analytic techniques. The associations captured may encompass a rich and varied span of objects, occupations, sceneries and natural elements, abstract constructs (e.g., emotions, intangibles), actions, and more. That truly looks inspiring and promising for brand research & marketing.

Ron Ventura, Ph.D. (Marketing)

Reference:

[*] Visual Elicitation of Brand Perception; Daria Dzyabura and Renana Peres, 2021; Journal of Marketing, Vol. 85 (4), pp. 44-66