A Recipe for Creativity: An Ingredient-Embedding Approach

Idea Creativity and Measurements of Idea Features

Prior research in creativity has identified idea novelty and appropriateness as the key dimensions of creative ideas (Amabile 1983; Burroughs, Moreau, and Mick 2008). Novelty refers to the extent to which an idea departs from existing and conventional ideas (Dahl and Moreau 2002; Moreau and Dahl 2005; Poetz and Schreier 2012). However, it must be “balanced with fit and appropriateness” by building on existing ideas (Runco 1988, p. 4), because simply being bizarre doesn't make something creative (Lubart 1994). At same time, simply imitating existing, tried-and-true ideas may score high on appropriateness, but doing so cannot be considered creative since it lacks novelty. Researchers typically have relied on consumer or expert evaluations to assess ideas separately on these two dimensions and have either considered these dimensions individually (e.g., Moreau and Dahl 2005; Yang et al. 2012) or combined them into a composite score (e.g., Dahl, Chattopadhyay, and Gorn 1999). However, such approaches are not scalable and require experts to evaluate individual ideas.

With the growing popularity of unstructured data in marketing applications (Balducci and Marinova 2018; Berger et al. 2020), researchers have developed natural language processing and network analysis–based approaches to automatically evaluate idea by examining the components of an idea (see Table 1). These idea components could be words or word parts in ideas, patents, names (Berger et al. 2012; Kelly et al. 2021; Toubia and Netzer 2017; Wei, Hong, and Tellis 2022), cited journals in academic papers (Uzzi et al. 2013), topics in song lyrics (Berger and Packard 2018), or cast or crew members in a movie (Packard et al. 2016). Such decompositions of ideas into smaller components to study the patterns in existing ideas aligns well with the cognitive view of idea generation. According to this view, new ideas are not created in a vacuum; rather, creative thinking involves searching through a conceptual space in the ideation domain and combining remotely associated concepts in original ways (Mehta and Dahl 2019). This conceptual space consists of a broad network of ideas, individual concepts, and their associations (Boden 1998).

OPEN IN VIEWER

Table 1.

Current Study and Literature on Creativity and Idea Generation.

Article	Creative Domain	Decomposition into Idea Components	Association Among Idea Components	Idea Novelty	Assessment Type for Idea Features	Pretrial Evaluation	Trial	Posttrial Evaluation
This article	Product ideas, food recipes	Yes	Overall internal coherence	Yes	Automated	—	Yes	Yes
Toubia and Netzer (2017)	Product ideas	Yes	Pairwise	Yes	Automated	Yes	Yes	—
Uzzi et al. (2013)	Research articles	Yes	Pairwise	Yes	Automated	—	Yes	—
Packard et al. (2016)	Movies	Yes	Pairwise	Yes	Automated	—	Yes	—
Berger and Packard (2018)	Songs	Yes	—	Yes	Automated	—	Yes	—
Wei, Hong, and Tellis (2022)	Project ideas	Yes	—	Yes	Automated	—	Yes	—
Kelly et al. (2021)	Patents	Yes	—	Yes	Automated	—	Yes	—
Berger et al. (2012)	Baby names	Yes	—	Yes	Automated	—	Yes	—
Goldenberg, Mazursky, and Solomon (1999)	Product ideas	Yes	—	Yes	Human ratings	Yes	—	—
Luo and Toubia (2015)	Product ideas	—	—	Yes	Human ratings	Yes	Yes	—
Kornish and Ulrich (2011), Poetz and Schreier (2012), Stephen, Zubcsek, and Goldenberg (2016)	Product ideas	—	—	Yes	Human ratings	Yes	—	—
Dahl, Chattopadhyay, and Gorn (1999), Moreau and Dahl (2005), Sellier and Dahl (2011), Yang et al. (2012)	Product design	—	—	Yes	Human ratings	Yes	—	—

Some of these studies evaluate the novelty or familiarity of an idea based on the similarity of its components to other ideas (Berger and Packard 2018; Berger et al. 2012; Kelly et al. 2021; Wei, Hong, and Tellis 2022). Others focus on the pairwise association among idea components. Uzzi et al. (2013) examine novelty and conventionality of scientific paper references based on similarity to and divergence from established citation patterns, comparing the frequency of each cocited journal pair across all papers with the frequency distribution created in a randomized citation network. Toubia and Netzer (2017) evaluate an idea's structural similarity to the prototypical idea by representing ideas with the distribution of their word-pair co-occurrences. Packard et al. (2016) assess the network centrality of idea components based on their prior co-occurrences.

A limitation of these methods is that they focus on a single dimension of the novelty–conventionality spectrum to evaluate an idea's consistency with domain norms and knowledge by quantifying the similarity or distance between an idea and other existing ideas or the “average” idea. In addition, they evaluate idea components in isolation or through pairwise relationships. While this approach may capture structural similarities among components, it can overlook the internal coherence of the components of the idea (i.e., the idea's components fitting together into a coherent, workable whole). An idea may contain individually common components or common component pairs that may combine poorly, or rare components and pairs that form a coherent whole. Capturing this holistic coherence requires moving beyond pairwise co-occurrences.

We address this gap by operationalizing the internal coherence or fit of idea components via representation learning. Our representation learning–based approach allows us to model the context-dependent relationship of idea components in a way that has not been done in previous work on ideation. It helps us capture the complex and holistic interactions between idea components beyond mere co-occurrence patterns, even those not previously observed in the data. This is especially important in combinatorial ideation domains such as food recipes, where a rare combination of a particular pairing does not necessarily imply its inappropriateness. For example, while existing methods might dismiss the pairing of rare ingredients (e.g., Brazil nuts and dates) as inappropriate, our approach infers their potential fit by learning broader relational patterns from data, such as the close relationship between Brazil nuts and other nuts that often appeared in recipes with dates and helps us identify promising and unconventional combinations that might otherwise be overlooked. In our empirical applications, we contrast the pairwise measure of associations with the gestalt representation learning–based measure.

Additionally, we complement the notion of fit among idea components with a more traditional measure of novelty of the idea components using the frequency in which they appear in similar corpora (e.g., the frequency of an idea component among other ideas in the same domain). These two measures, fit and frequency, allow us to separately assess the impact of these two dimensions on downstream outcome. Indeed, we find diverging effects, where component fit leads to both higher trial and higher evaluation posttrial, but component novelty is associated with lower trial but higher posttrial evaluation.

Our proposed approach to capture component fit is consistent with gestalt theory, which proposes that people comprehend objects holistically as unified wholes rather than at an atomistic level as collections of separate elements; therefore, the whole is perceived as more than the sum of its parts (e.g., Koffka 1962; Radford and Bloch 2011; Wertheimer 1938). In the context of creative ideas, this approach suggests that a person evaluates the idea as a complete entity rather than analyzing its components separately. While people may have certain preferences toward individual components, the gestalt of all of the idea's components combined together in a holistic way acts as another factor in their evaluation of a creative idea. Similarly, in the context of food recipes, which is one of the main applications in this article, culinary experts often emphasize the fit among ingredients for creative and successful recipes (Page 2017). Wolfgang Puck of Spago draws similarities between cooking and art and states that even though the flavors are limited, “it's how you combine them that sets you apart”; Daniel Patterson of Coi and LocoL emphasizes that “just because two components are amazing doesn't mean that combining them will work”; and bartender Audrey Saunders of Pegu Club tells aspiring chefs to “make sure (your) ingredients dance well together” (Page 2017, pp. 27–28).

Our approach operationalizes both idea novelty and fit based on the idea's components. This approach can be applied to other creative domains where ideas can be decomposed into smaller, identifiable, and partially substitutable components such as attributes in a product, scents in a perfume, or people in a team. While decompositions are popular in studying various ideas (see Table 1), not all domains are suitable for this practice. For example, some creative work, such as instrumental music, poetry, abstract art, or jokes, may not be easily deconstructed into identifiable smaller components. For such domains, more subjective measures of creativity by domain experts (e.g., Consensual Assessment Technique; Amabile 1982) have been widely used.

Creative Cognition Through Representation Learning

The idea of using representation learning for idea generation is consistent with the creative cognition process. Innovation often emerges from reapplying structured knowledge patterns to solve novel problems (Altshuller 1984) or inspire new product ideas (Goldenberg, Mazursky, and Solomon 1999). A larger and more diverse domain-relevant knowledge base enables generation of more ideas through combining, recycling, and recombining elements (Amabile 1988, 1996; Fleming and Szigety 2006; Hargadon and Sutton 1997). The Geneplore model of idea generation proposed by Finke, Ward, and Smith (1992) presents idea generation as an iteratively performed two-stage process. In the generative phase, people select the pre-inventive structures to combine in a new idea. In the exploratory phase, they evaluate the fitness of this combination. These pre-inventive structures may include representations or mental models of specific idea components and category exemplars (Burroughs, Moreau, and Mick 2008; Ward 2001).

The idea of mental models or representations of pre-inventive structures closely resembles concepts from a very different stream of literature in machine learning and text analysis: representation learning—specifically, word embeddings. Word embeddings are a family of methods for analyzing language and capturing semantic meaning of words based on the context in which they appear. Since the introduction of word embeddings (Bengio et al. 2003), many variants and extensions have been developed (e.g., Mikolov, Chen, et al. 2013; Mikolov, Sutskever, et al. 2013; Rudolph et al. 2016). In this line of research, relationships between concepts (e.g., words in idea texts, ingredients in recipes) are learned by analyzing patterns in large datasets. Unlike traditional methods that study these concepts in isolation or in pairs, word embeddings capture meaning based on the whole context. The key idea is that words that frequently appear in similar contexts tend to have similar meanings or play a similar role in the language (Firth 1957; Harris 1954).

In this framework, idea components are represented as vectors that capture their unique features and functions, learned from many different ideas. The probability of a component appearing in an idea depends on how its mental representation (i.e., its vector representation) interacts with the mental representations of the other components in that idea. As a result, words such as “running” and “walking” that frequently appear in similar contexts (e.g., with words like “activity,” “pace,” “distance”) will have more similar vector representations and appear closer in the learned vector space (embedding space), whereas components that complement each other and “fit together” will have a higher probability of appearing together.

The existing methods that evaluate idea components in isolation or through pairwise relationships rely on static information in retrieved exemplars. Therefore, they mostly focus on a single dimension of creativity, as ideas closely resembling the existing ideas are less novel (more appropriate) and ideas completely diverging from existing ideas are more novel (less appropriate). Our approach helps us access the broader knowledge structures by deconstructing the information contained in existing ideas into more flexible and abstract representations (Ward 1995) and enables us to separate idea appropriateness from idea novelty.

Researchers have recently explored the ability of large language models (LLMs) to support creative processes. While the use of LLMs in idea generation may enhance efficiency, productivity, and idea quality, LLM-generated ideas have been found to be less novel and diverse than human-generated ideas (e.g., Boussioux et al. 2024; Meincke et al. 2024). Our preference for the embedding approach primarily stems from how it helps us operationalize gestalt fit among idea components and distinguish fit from novelty in studying their relationship with idea popularity and evaluation. While LLMs can articulate their reasoning, their internal computations remain opaque.

The Downstream Consequences of Creative Ideas

The impact of idea creativity on idea success has been captured using different measures such as adoption or popularity rankings (Berger and Packard 2018; Berger et al. 2012), funding performance or revenue (Packard et al. 2016; Wei, Hong, and Tellis 2022), number of citations (Kelly et al. 2021; Uzzi et al. 2013), potential product success (Kornish and Ulrich 2011), customer appeal (Dahl, Chattopadhyay, and Gorn 1999), and pleasure (Giora 2003). However, these measures often focus on either pretrial evaluation or trial and rely mostly on stated preferences rather than actual behavior. These studies consistently show that high-quality ideas emerge by rearranging elements from existing knowledge bases in novel ways (Dahl and Moreau 2002; Ward 1994), and novel ideas with high appropriateness tend to lead to more positive outcomes while extremely novel or overly common ideas perform less well (Dahl, Chattopadhyay, and Gorn 1999; Giora 2003; Toubia and Netzer 2017). However, to the best of our knowledge, no paper has explored simultaneously the effect of creativity on both trial and posttrial evaluations (see Table 1).

We expect novelty and fit to have distinct effects on trial and posttrial evaluation due to systematic differences in pre- and posttrial evaluations (e.g., Hamilton and Thompson 2007; Wright and Lynch 1995). Since more common and recognizable components may be easier to process or perceived as lower risk at pretrial (Wright and Lynch 1995), consumers may exhibit an initial hesitation toward trying out ideas with high novelty. In contrast, high fit among idea components reduces the perceived risk of a bad outcome by signaling that the combination is coherent and appropriate. On the flip side, ideas that lack novelty and closely imitate existing ideas may not be “new” enough to spark interest for trial. We investigate the pretrial evaluation across both empirical applications using outcome measures such as perceived popularity, purchase intent, or actual consumer trials and find consistent results.

Posttrial evaluation, in contrast, reflects the consumption experience rather than anticipations or expectations. Studying posttrial evaluation requires posttrial data, which is less commonly available and hence has been explored less frequently in previous research (see Table 1). In this case, novelty may work in an idea's favor, since consumers update their beliefs and attitudes after trial (Kempf and Smith 1998; Oliver 1980), and novel components, especially when they fit well together, may be evaluated more favorably. High fit among components should continue to be important at the posttrial evaluation since coherent combinations tend to deliver a more satisfying or integrated experience. Considering the posttrial evaluation is a consistent predictor of repeat purchase (Smith and Swinyard 1983), idea creativity, specifically the novelty dimension, is expected to have a differential impact on idea success based on how we measure it. This is indeed what we find in our second empirical application when we investigate the impact of novelty and fit among recipe ingredients on both recipe trial and ratings.

Overall, our work builds on and extends the extensive work on ideation and creativity by introducing a novel methodological application: the use of representation learning through embeddings to measure component fit in creative ideas and explore it along a traditional measure of idea novelty. We further explore the downstream consequences of our proposed approach for creativity with two outcome measures, trial and posttrial evaluation, which have not been jointly studied in the literature, and we explore their possibly diverging downstream consequences.

Embedding Approach

The core idea underlying the embedding approach in text analysis is to represent each word with two types of vectors: one that captures the meaning of the word itself (the embedding vector) and another that captures its role in the context (the context vector). Both are multidimensional vectors of the same length, which are much smaller than the total number of words in the vocabulary. This approach attempts to maximize the likelihood of the observed text by assigning similar vectors to words that tend to appear in similar contexts. Models of this kind form the foundation of modern LLMs, which more recently became popular for idea-generation tasks.

To construct a measure of internal coherence, we model the context-dependent relationship among idea components. We use the embedding approach to represent each idea component conditional on other components that appear in the same idea (Rudolph et al. 2016). In embedding model, each component n is represented by an embedding and a context vector (ρ_n, α_n) of length t, where t defines the number of dimensions in the embedding space. The choice of t is determined by the researcher, often based on model fit criteria. The interaction among idea components is captured by the inner product ρ n T g ( α , c nr ) where g ( α , c nr ) is the context representation of component n in idea r. If this inner product is high, it is more likely to observe component n with the rest of the components in idea r.

Since we are interested in internal coherence, we assume that the conditional distribution of component n in idea r given its context c_nr, stated as x nr | x c nr , follows a Bernoulli distribution. This modeling choice helps us focus on whether the component is included in the idea rather than its prevalence within the idea (i.e., number of word mentions or ingredient quantities).

x nr | x c nr ∼ Bernoulli ( σ ( ρ n T g ( α , c nr ) ) ) , where σ ( . ) is the logistic function .

Each data point x_nr = 1 in our N × R binary data matrix indicates the presence of component n in idea r, where N denotes the number of unique components across R ideas. We represent the context c_nr by averaging the context vectors of all the other components in idea r: g ( α , c nr ) = 1 | c nr | ∑ m ∈ c nr α m . Averaging gives us a normalized representation of the context and ensures comparability across ideas with different number of components. Alternative context aggregation methods, such as summing the context vectors, would inflate the internal coherence scores for ideas with a higher number of components and confound our results.

The model jointly estimates the embeddings ρ and context vectors α by maximizing the objective function, which contains conditional log-likelihoods of observed components in ideas and regularization terms that resemble log-priors for the embeddings and context vectors:

Idea Creativity Measures

Category-Aware Extension in Embedding Approach

In the domain of food recipes, each recipe is a collection of food ingredients, and the fit among ingredients is critical for creative and successful recipes (Page 2017). When we model the ingredients’ relationship, we focus only on whether the ingredient is used in the recipe and ignore the exact quantity (e.g., 1 tablespoon, 2 cups). Our reasoning for ignoring the amount of the ingredient is threefold. First, the existence of an ingredient in the recipe is a more important factor than the amount of the ingredient, especially for consumers to decide if they want to make or eat that recipe. Second, the contribution of the ingredient in the recipe may not always be attributed to the amount used in the recipe. For example, 1 tablespoon of ground cinnamon may be more potent than 1 cup of rice flour in a recipe. Third, the conversion across different units for quantities is not straightforward (e.g., 1 cup sugar = 200 grams, 1 cup oatmeal = 100 grams). Accordingly, the conditional distribution of observing an ingredient given other ingredients in the recipe can be modeled as Bernoulli, as described previously.

In our recipe dataset, recipes belong to an inherently different set of categories (e.g., dessert vs. main dishes). These categories introduce another source of contextual information that can be leveraged in understanding ingredient interactions. While it may be tempting to estimate a separate embedding model for each recipe category, beyond data limitations of several relatively small recipe categories, we may want to leverage ingredient fit from one recipe category (e.g., mint and lemon in cocktails) to other categories (e.g., mint and lemon in salads), as ingredients that work well together in one category may also have a higher fit in another category. Indeed, the creativity literature and cooking experts emphasize the transfer of knowledge across categories as a key driver of creativity (e.g., Gabora and Carbert 2015; Page 2017).

To allow for possible category-specific ingredient interactions but leverage possible cross-category knowledge transfer, we extend our context definition in the embedding model specification to incorporate the category information (see Web Appendix F.2 for a detailed specification). Specifically, we introduce a category-specific context vector θ_d of length t (i.e., the same dimension as the ingredient vectors) for each of the 14 main recipe categories (e.g., Main Dish, Drinks, Salads). Then, we redefine the context representation of recipe r in category d(r) by reweighing the contribution of other recipe ingredients in the context using the Hadamard product: g ( α , d ( r ) , c nr ) = 1 | c nr | ∑ m ∈ c nr θ d ⊙ α m . This approach accounts for nuances across categories, as the interaction with the category vector may amplify certain features and downplay others in the context representations even if the recipe ingredients are identical. We retain shared knowledge across recipes through global representations for ingredients while still adapting to category-specific differences. ⁵ We choose the dimensions of the embedding space as t = 140 based on model fit criteria (see Web Appendix F.3 for details).

A visualization of the ingredient vectors in Figure 2 shows high face validity. Many ingredients with similar features and/or culinary roles appear closer together in the embedding space capturing many intuitive and meaningful relationships among ingredients (see Web Appendix F.4 for a detailed explanation).

OPEN IN VIEWER

Figure 2.

Two-Dimensional Projection of the Embedding Space (t = 140) for Ingredients That Appear in at Least 150 Recipes Using t-SNE.

Recipe Creativity Measures

Consistent with the category-aware embedding model specification, we use the modified context representation g(α, d(r), c_nr) in our computation of category-specific ingredient fit. This approach allows the interaction between ingredients to be scaled by the category vector and the ingredient fit measure to vary based on the recipe category even if the recipe ingredients are identical.

Recipe categorization further helps us refine our ingredient novelty measure. As discussed previously, novelty is defined as the extent to which a recipe deviates from what is commonly observed within the recipe domain. Since ingredient usage can substantially vary even within broad recipe categories (e.g., Main Dish, Desserts), we use subcategory information in defining our reference set to operationalize ingredient novelty. Specifically, we measure ingredient novelty by the frequency of that ingredient across all recipes in a specific recipe subcategory. For example, brown sugar is frequently used in the Desserts > Crisps and Crumbles subcategory (66.2%), but only 14.9% of Desserts > Cakes recipes include brown sugar. Similarly, ground black pepper is used in more than half of Main Dish > Meatloaf recipes (53.2%), but it is not as popular in Main Dish > Pizza subcategory (13.5%).

In addition to the selection of the ingredients in the recipe, ingredients may be combined in conventional or novel ways to prepare a recipe. For instance, grilling may be a usual preparation technique for chicken recipes, but it is an unconventional method for dessert recipes. Therefore, we also define a measure of preparation novelty similar to the ingredient novelty and operationalize it as the frequency of preparation methods within a recipe subcategory. To that end, we automatically text-mine the preparation instructions using the Python spaCy package (Honnibal and Montani 2017). We use part-of-speech tagging to identify the verbs used in the preparation instructions. We obtain a total of 1,193 verbs in the lemmatized form across recipes. We then calculate preparation verb frequency within each recipe subcategory. Finally, we aggregate the ingredient fit, ingredient frequency, and verb frequency measures at the recipe level by computing the mean and standard deviation.

Analysis

In this section, we investigate the relationship between ingredient novelty and fit and recipe success, controlling for a host of recipe and recipe creator characteristics. We focus on two distinct recipe success measures: recipe trial and posttrial evaluation. We measure recipe trial by the number of consumers who have tried the recipe based on Allrecipes. Users self-report this measure by clicking “I made it” button on the recipe page and then receive a prompt to rate and review the recipe. Due to the long tail of the trial distribution and to account for recipes with no trial (9.66% of the data), we use log(trial + 1) as a measure of recipe trial. The posttrial evaluation of recipes is reflected in the distribution of the recipe's star ratings. Due to the large proportion of positive reviews (Schoenmueller, Netzer, and Stahl 2020), we use the percentage of 5-star ratings as a measure of posttrial evaluation. Since we do not observe the distribution of ratings for 35.44% of the recipes, we define a dummy variable to indicate recipes with no detailed rating information when we control for this variable in our analysis of recipe trial.

In addition to their main effect, we explore the nonlinear relationship for the mean ingredient fit, the mean ingredient frequency, and the mean verb frequency. Recipes posted earlier tend to have a higher number of trials, and previous research has shown that ratings decline over time (Godes and Silva 2012). Since the data on when the recipe was posted are not directly available, we use the days since the first review as a proxy for the recipe age and define a dummy variable to indicate recipes with no reviews (13.9% of the recipes). Recipe complexity may also affect trials, ratings, and reviews, as users may be inclined to make simpler, quicker, or more elaborate recipes. Kornish and Jones (2021) show that it is important to control for the number of concepts (ingredients) in the idea when assessing the novelty of the idea. Therefore, we control for the number of ingredients, preparation time, total time, and number of words in the preparation instructions. We use imputation for missing preparation and total time variables (9.48% and 9.12% of the data, respectively). Moreover, we control for the recipe creator’s number of followers and the number of personal recipes the recipe creator has posted on Allrecipes, as more connected and more experienced creators may receive more trials and favorable ratings and reviews. Because users may consult the website more frequently for some subcategories or evaluate some subcategories more favorably, we include recipe subcategory fixed effects.

Each observation corresponds to a recipe in our dataset. After removing recipes with missing data on the dependent and independent variables, the final dataset consists of 53,879 recipes for recipe trial analysis and 34,877 recipes for posttrial analysis. We provide the details in our data augmentation approach and the descriptive statistics for our key variables in Web Appendix G. As Table 4 shows, we find that for both trial and posttrial evaluation (percentage of 5-star ratings), consumers prefer recipes in which the ingredients fit well together. This finding is consistent with the typical view in the food industry that one of the most important factors in a successful recipe is how well the ingredients fit together (Page 2017). However, when it comes to novelty of the ingredients themselves and the preparation methods, we find a differential effect between trial and posttrial evaluation.

OPEN IN VIEWER

Table 4.

The Relationship Between the Ingredient Fit and Novelty Measures and Recipe Success Measures.

Dependent Variable:	Recipe Trial Log(Trial + 1)						Recipe Evaluation % of 5-Star Ratings
Model Specification:	Linear			Quadratic			Linear			Quadratic
	Model 1A			Model 1B			Model 2A			Model 2B
	Coef.	SE	p	Coef.	SE	p	Coef.	SE	p	Coef.	SE	p
Category-Aware Internal Coherence
Ingredient fit mean (mean-centered)	5.6628	.2177	<.001	5.5714	.2191	<.001	.2632	.0527	<.001	.2625	.0528	<.001
Ingredient fit mean (mean-centered)2				16.0012	3.5285	<.001				.0660	.8498	.938
Ingredient fit SD	−5.0594	.3181	<.001	−5.1962	.3224	<.001	−.2952	.0771	<.001	−.2942	.0781	<.001
Novelty of Ingredients (Reverse-Coded)
Ingredient frequency mean (mean-centered)	1.8285	.0909	<.001	2.6840	.1218	<.001	−.1269	.0211	<.001	−.1371	.0292	<.001
Ingredient frequency mean (mean-centered)2				−5.3347	.4635	<.001				.0785	.1089	.471
Ingredient frequency SD	−1.0328	.1167	<.001	−2.0766	.1528	<.001	−.0157	.0276	.568	.0044	.0365	.905
Novelty of Preparation (Reverse-Coded)
Verb frequency mean (mean-centered)	.3641	.0587	<.001	.3193	.0615	<.001	−.0964	.0140	<.001	−.1175	.0148	<.001
Verb frequency mean (mean-centered)2				.2720	.3065	.375				.3458	.0735	<.001
Verb frequency SD	.3683	.1083	.001	.3946	.1173	.001	.0060	.0261	.818	.0573	.0283	.043
Idea Complexity
No. of ingredients (mean-centered)	.0022	.0018	.232	−.0003	.0020	.875	.0037	.0004	<.001	.0037	.0005	<.001
No. of ingredients (mean centered)2				−.0001	.0003	.838				.0001	.0001	.270
No. of words in instruction (in 100 cases)	.0101	.0113	.371	.0110	.0113	.327	.0176	.0028	<.001	.0176	.0028	<.001
Prep time (in 100 min)	−.0067	.0030	.026	−.0064	.0030	.034	.0005	.0007	.477	.0005	.0007	.484
Total time (in 100 min)	−.0001	.0003	.868	−.0001	.0003	.876	.0003	.0001	.021	.0003	.0001	.030
Ideator Variables
Ideator's follower count (in 100,000 cases)	.0007	.0026	.797	.0009	.0026	.721	.0030	.0008	<.001	.0030	.0008	<.001
Ideator's recipe count (in 100 cases)	.0690	.0033	<.001	.0697	.0033	<.001	.0050	.0008	<.001	.0049	.0008	<.001
Other Control Variables
Dummy for no recipe age	−1.4623	.0174	<.001	−1.4585	.0174	<.001	.1545	.0083	<.001	.1547	.0083	<.001
Recipe age (in 100 days) × (1 − dummy for no recipe age)	.0252	.0003	<.001	.0252	.0003	<.001	−.0022	.0001	<.001	−.0022	.0001	<.001
Dummy for no rating	.2703	.0196	<.001	.2705	.0195	<.001
Percentage of 5-star ratings × (1 − dummy for no rating)	2.1279	.0259	<.001	2.1229	.0258	<.001
Log(number of trials + 1)							.0654	.0009	<.001	.0655	.0009	<.001
No. of observations	53,879			53,879			34,877			34,877
R2	.632			.633			.187			.187
Adj. R2	.631			.632			.184			.184

Notes: Prep time and total time variables are imputed. Subcategory fixed effects are included. For brevity, we do not report the estimates for subcategory fixed effects in this table.

On the one hand, consumers are more likely to try recipes with common or typical ingredients and preparation methods, though the marginal impact on trial is diminishing for extremely common ingredients (Figure 3). On the other hand, when it comes to posttrial evaluation, consumers prefer recipes with more novel ingredients and either extremely novel or extremely common preparation methods. Note that these effects hold after controlling for the complexity of the recipe as measured by the number of ingredients, the length of the instructions, preparation time, and total time. Furthermore, our findings for recipe trial are consistent with the results of ideation dataset analyses on idea adoption and perceived popularity.

OPEN IN VIEWER

Figure 3.

Relationship Between Ingredient Fit, Frequency, and Recipe Trial.

Thus, we conclude that when it comes to food recipes, the effect of novelty may be more nuanced than commonly found in the creativity literature, especially in domains where consumers participate in the production of the final output. First, novelty may not be preferable when it comes to recipe trial. However, when we look at posttrial evaluation of these consumers, they prefer recipes that have unique ingredients. Furthermore, distinguishing between novelty of the ingredients themselves and the fit of ingredients, we find that conventional rather than unusual ingredient combinations (high ingredient fit) are preferred for both trial and evaluation. In other words, balanced creativity (i.e., novelty of the components and coherence of the whole) is positively related with posttrial evaluation. To the best of our knowledge, this is the first article to separate these two important behavioral outcomes of novelty—namely, trial and posttrial evaluation—which are likely to be the antecedents for repeat consumption. Such separation is difficult in studies that that measure stated preference for outcome measures, as in many studies in the creativity literature (e.g., Dahl, Chattopadhyay, and Gorn 1999; Luo and Toubia 2015; Toubia and Netzer 2017).

Internal Coherence Versus Pairwise Co-Occurrences

The embedding-based measure of internal coherence captures a gestalt form of ingredient fit, in contrast to commonly used pairwise ingredient associations in the literature (e.g., Toubia and Netzer 2017; Uzzi et al. 2013). In the first empirical application, we saw that the embedding-based measure results in higher model fit relative to the pairwise measure. To further demonstrate the distinction between the two measures, we repeat that analysis by replacing the embedding-based ingredient fit measure in Table 4 with the Jaccard index (see Web Appendix D.3 for full results). First, similar to the first empirical application, we observe that the relationship between the overall co-occurrence of recipe ingredients and recipe trial has an inverted U-shape, and that a higher overall ingredient co-occurrence is associated with lower recipe evaluation. Moreover, the model based on ingredient co-occurrences exhibits a lower explanatory power than internal coherence model for both recipe success measures. These results illustrate that the embedding-based ingredient fit measure is empirically distinct from the co-occurrence-based Jaccard measure and provides a better model fit.

To provide intuition for this distinction, we present in Table 5 two representative examples that illustrate how the embedding-based ingredient fit and co-occurrence-based Jaccard index can yield opposing assessments of ingredient complementarity. Recipe 1 combines turnip, bacon, and brown sugar in a vegetable side dish recipe and ranks approximately in the 95^th percentile of the overall embedding-based ingredient fit but only the 3^rd percentile of the co-occurrence-based overall Jaccard index measure. This ingredient combination is not obvious based on the co-occurrence patterns, as brown sugar frequently appears with baking ingredients (e.g., all-purpose flour, vanilla extract), bacon is typically paired with savory ingredients (e.g., cheddar cheese, onion, potato) and turnip commonly appears with other root vegetables (i.e., parsnip, rutabaga). However, this trio has a high overall ingredient fit, since brown sugar binds bacon and turnip together into a coherent and complementary combination.

OPEN IN VIEWER

Table 5.

Recipe Examples with Contrasting Assessments for Category-Aware Ingredient Fit and Jaccard Index.

	Recipe 1	Recipe 2
Recipe subcategory	Side Dish > Vegetable	Meat and Poultry > Chicken
Ingredients	turnip, bacon, brown sugar	chicken, stuffing, butter
Category-aware ingredient fit	∼95th percentile	∼94th percentile
Jaccard index	∼3rd percentile	∼4th percentile

Notes: Percentiles are computed empirically as the proportion of recipes in the Allrecipe dataset with scores less than or equal to the focal recipe.

Recipe 2 combines chicken, stuffing, and butter in a chicken recipe within the Meat and Poultry category. It approximately ranks in the 94^th percentile of the overall embedding-based ingredient fit but only the 4^th percentile of the overall Jaccard index measure. In the recipe dataset, butter commonly appears with baking ingredients (e.g., all-purpose flour, egg), chicken is often paired with soup ingredients (e.g., chicken broth, celery, carrot), and stuffing appears most frequently in Thanksgiving dishes (i.e., turkey, turkey gravy). Even though the fit among these ingredients is very high, the co-occurrence-based measure underestimates their complementarity because of the versatility of the ingredients.

These examples highlight the key distinction between the two approaches. The embedding model learns the underlying relationship among ingredients and can infer the ingredient fit even in the absence of direct co-occurrences. This allows us to discover ingredient combinations that are coherent even when they are rarely observed in a given dataset. In contrast, inferences with the Jaccard measure completely rely on observed co-occurrences; therefore, they are sensitive to the size, sparsity, and diversity of the dataset. This often results in underestimating the compatibility of infrequently observed ingredients and ingredient combinations. Further investigation of the full set of 54,951 recipes reveals that ingredient frequency predicts the difference between the fit and Jaccard measure (b = −.128, p < .001), such that recipes with less common ingredients show a higher fit relative to their pairwise co-occurrence value.

To provide further evidence for construct validity of our ingredient fit measure, we conducted an online panel study on MTurk. We randomly sampled 500 recipes from Main Dish and adjacent categories such as Meat and Poultry; Seafood; and Soups, Stews, and Chili. Following standard practice (e.g., Kornish and Ulrich 2011; Toubia and Netzer 2017), we asked each study participant to evaluate a subset of these recipes one at a time. We employed a two-step screening process and an attention check question before each recipe evaluation to ensure the quality of evaluations. Overall, we collected a total of 2,978 recipe evaluations from 610 participants across the 500 recipes. Each participant evaluated at most five randomly assigned recipes, and each recipe received at least five evaluations (median = 6, max = 9). For each recipe, participants rated perceived ingredient fit (i.e., “How well do you think the ingredients in this recipe fit together?”) and perceived popularity (i.e., “How popular do you think this recipe would be?”) on a five-point Likert scale. Each observation corresponds to a respondent's evaluation of a recipe. The details about the participant screening process, stimulus design, and descriptive statistics for our collected and computed measures are available in Web Appendix D.4.

First, we assess the construct validity by regressing the perceived ingredient fit on both overall internal coherence measure (ingredient fit mean) and co-occurrence based benchmark measure (Jaccard index mean). Ingredient fit significantly predicts perceived ingredient fit (b = 1.767, p = .025) whereas the Jaccard index does not (p = .14). This analysis provides evidence for construct validity as our ingredient fit measure indeed captures how recipe ingredients complement each other over and beyond simple ingredient co-occurrence patterns.

Next, we examine the underlying mechanism of internal coherence. As expected, our ingredient fit measure is positively related to the perceived popularity of the recipe (b = 2.133, p = .003) controlling for the number of ingredients. We also find that the perceived ingredient fit fully mediates this relationship. Our mediation analysis (Hayes 2022, PROCESS Model 4 with 5,000 bootstrap samples) shows a significant indirect effect of overall ingredient fit on perceived popularity via perceived ingredient fit (b = 1.638, 95% CI = [.72, 2.59]), while the direct effect of overall ingredient fit on perceived popularity is not significant (b = .495, p = .375). These results indicate that our internal coherence measure impacts recipe popularity by shaping people's perception of ingredient compatibility.

Taken together, we can conclude that ingredient fit is a useful and distinct measure that captures ingredient compatibility and internal coherence of a recipe beyond simple co-occurrences observed from the data. It is a significant predictor of how consumers perceive the compatibility of recipe ingredients and affects recipe popularity through this perception.

Robustness Checks

Evaluating App's Recommendations

To show how our model of ingredient fit and frequency can generate recommendations that improve recipes on trial or rating evaluations, we analyze a holdout sample of 498 recipes. These recipes are chosen from a total of 1,072 recipes that were excluded from our regression analysis due to missing variables, based on two criteria: each recipe has (1) at least one trial and (2) complete information on the distribution of star ratings.

For each recipe in this holdout sample, we use Gentool to generate ingredient recommendations. Table 6 shows the impact of Gentool's top ingredient recommendation on recipe trial and evaluation. Across all success measures, Gentool improves a vast majority of the recipes (98.8%–100%); however, the type of the top recommended modification differs by objective. The top Gentool recommendation for improving trial is to remove an ingredient (75.4%), whereas the top Gentool recommendation for improving evaluation is to add an ingredient. On average, we observe that top Gentool recommendation increases the number of trials by 31.65%, corresponding to 73.79 additional trials. In terms of recipe evaluation, the top Gentool recommendation increases the percentage of 5-star ratings by 1.17 percentage points (1.88 additional 5-star ratings), reduces the percentage of 1-star ratings by .33 percentage points (.53 fewer 1-star ratings) and improves the average rating by .03 points on average. Due to skewness in number of trials and ratings at the extremes, we also report the median changes on these measures in Table 6. When extrapolated to the full Allrecipes platform, these improvements imply that a single ingredient modification per recipe could generate approximately 4.26 million additional trials, 108,500 more 5-star ratings, or 30,500 fewer 1-star ratings on average. Taken together, these findings demonstrate that small, targeted ingredient adjustments can produce economically and behaviorally meaningful improvements at scale.

OPEN IN VIEWER

Table 6.

Improvement on Recipe Trial and Posttrial Evaluation with Top Gentool Recommendations.

		Recipe Trial	Recipe Evaluation
		No. of Trials	No. of 5-Star Ratings	No. of 1-Star Ratings	Avg. Star Rating
Improvement rate	Recipes improved (%)	98.8%	99.8%	100.0%	99.8%
Top Gentool recommendation type	Addition	10.6%	72.2%	94.0%	75.7%
	Substitution	14.0%	15.7%	4.0%	12.5%
	Removal	75.4%	12.1%	2.0%	11.9%
Change in outcome	Avg. % change	+31.65%	+1.17	−.33	N.A.
	Mdn % change	+21.97%	+1.02	−.30	N.A.
	Avg. absolute change	+73.79	+1.88	−.53	+.03
	Mdn absolute change	+12.53	+.32	−.10	+.02

Notes: N.A. = not applicable. Top Gentool recommendation type and change in outcome are computed conditional on improvement in the related outcome measure.

In an online panel study, we also compare Gentool's recommendations to human-generated suggestions and find that Gentool's top recommendations are superior in idea quality to both expert and novice consumers’ suggestions (see Web Appendix I.3 for details). While Gentool provides a systematic, data-driven way to generate ingredient recommendations that align with consumer preferences instead of relying on intuition or trial and error, we also acknowledge several limitations related to the nature of the dataset and integration of domain knowledge that could be explored in future research (see Web Appendix I.4).