The Impact of Generative AI on Personalized Content Marketing in E-Commerce

May combine two claims.

Consumers may feel that brands using GenAI are cutting corners or replacing people with machines, leading to reduced emotional engagement and brand loyalty.

“Consumers may feel that brands using GenAI are “cutting corners” or “replacing people with machines,” leading to reduced emotional engagement and brand loyalty (Granulo et al., 2021).”

• cites Granulo et al., 2021 sourcePartly supported

  The Abstract (page unavailable) says firms can replace human labor with technology and that consumers prefer human over robotic labor for products with higher symbolic value. The paper does not establish that consumers see brands using GenAI as “cutting corners,” or that such use reduces emotional engagement or brand loyalty; its findings concern production mode and product preference. (LLM, conf=0.87)

  Experimental — accuracy not yet measured

Classified as probably not a substantial claim.

The study used G*Power 3.1 power analysis; for an independent-samples t-test with d = 0.30, alpha = 0.05, and power = 0.80, the required sample size was about 176 per condition.

“G*Power 3.1 was used to perform a power analysis (Faul et al., 2009).In a scenario involving an independent-samples t-test of two means, with a small-to-medium effect size (d = 0.30), alpha = 0.05 and power = 0.80, the necessary sample size is about 176 per condition.”

• cites Faul et al., 2009 sourceSupported

May combine two claims.

LLMs produce contextually fluent content that can refer to specific browsing history and adjust tone, leading to greater sense of relevance and less creepiness.

“Such contextual fluency can lead to a greater sense of relevance and less creepiness of overly specific and mechanically delivered advice (Aguirre et al., 2015).”

• cites Aguirre et al., 2015 sourcePartly supported

  The Implications for Theory section (page unavailable) says customers find personalized ads more relevant and useful when they know their data was collected. The paper does not study LLMs, tone adjustment, or mechanically delivered advice; its Psychological Ownership and Vulnerability section instead says highly personalized ads can increase discomfort when data collection is covert. It therefore does not support contextual fluency as a way to reduce creepiness. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

May combine two claims. The quote only partly states this claim.

When AI generation was disclosed, perceived authenticity dropped to M = 3.86, significantly lower than GenAI no-disclosure (t(238) = 7.31, p < 0.001, Cohen's d = -0.64).

“Perceived brand authenticity decreases strongly when consumers explicitly receive information that the content is AIgenerated (d = -0.64).”

The quote only partly states this claim.

Participants in the GenAI condition spent significantly more time on site (M = 124.6 seconds, SD = 35.1) than rule-based participants (M = 108.3 seconds, SD = 32.7).

“The mean time on site for GenAI participants was 124.6 seconds (SD = 35.1),compared to 108.3 seconds (SD = 32.7)for rule-based participants.”

• AssumptionRests on an assumption: The simulated e-commerce shop (UrbanHarbor) is a valid representation of real e-commerce shopping behavior, so findings generalize to real e-commerce.
• AssumptionRests on an assumption: The AI-generated email from GPT-4 is representative of GenAI personalization content.

May combine two claims.

Managers consistently reported that negative consumer reactions were infrequent but disproportionately intense and often triggered by disclosing AI use.

“Managers consistently reported that negative consumer reactions were infrequent, but when they occurred, they were disproportionately intense and often triggered by disclosing AI use.”

• AssumptionRests on an assumption: The six interviewed marketing managers and their three brands are representative of e-commerce brands using GenAI, so their reports can be generalized.

May combine two claims.

The impact of GenAI on brand perception, although statistically significant, was low (d = 0.15).

“The impact on the brand perception, although statistically significant, was low (d = 0.15).”

May combine two claims.

The difference in brand perception was statistically significant (t(1138) = 2.58, p = 0.010) with a small effect size (Cohen's d = 0.15).

“The difference was statistically significant (t(1138) = 2.58, p = 0.010), though the effect size was small (Cohen's d = 0.15).”

May combine two claims.

The mean privacy concern score in the GenAI condition was 4.67 (SD = 1.34), significantly above the scale midpoint of 4 (t(239) = 7.74, p < 0.001, d = 0.50).

“The mean privacy concern score (M = 4.67, SD = 1.34) was significantly above the scale midpoint (t(239) = 7.74, p < 0.001, d = 0.50), indicating that consumers, on average, agree that they are concerned about privacy in the context of GenAI personalization.”

The quote only partly states this claim.

The difference between observed detection accuracy (53.3%) and chance (50%) was not statistically significant (p = 0.12).

“The difference was not statistically significant (p = 0.12).”

The quote only partly states this claim.

Purchase intentions and trust decrease by the same levels as perceived authenticity when AI is disclosed.

“Purchase intentions and trust decrease by the same levels.”

May combine two claims.

Participants performed at chance level, failing to reliably identify AI-generated content.

“Participants performed at chance level, failing to reliably identify AIgenerated content.”

• AssumptionRests on an assumption: The human-written emails used in the detection task are representative of human-crafted marketing content and comparable to the AI-generated emails except for authorship.
• AssumptionRests on an assumption: The AI-generated email from GPT-4 is representative of GenAI personalization content.

May combine two claims.

All six managers described a hybrid workflow where AI generates first drafts and humans review, edit, and approve; fully automated zero-touch GenAI personalization was not practiced by any of the brands.

“All six managers described a hybrid workflow: AI generates first drafts; humans review, edit, and approve.Fully automated, zero-touch GenAI personalization was not practiced by any of the brands.”

• AssumptionRests on an assumption: The six interviewed marketing managers and their three brands are representative of e-commerce brands using GenAI, so their reports can be generalized.

Probably not a substantial claim: may not pass S3 (load-bearing).

Merely indicating that a message is AI-generated made the message seem less trustworthy.

“They discovered that merely indicating that a message is AI-generated made the message seem less trustworthy, with or without the actual use of AI.”

• cites Jakesch et al. (2019) sourceNot checked — no full text of the source

May combine two claims.

Random assignment to conditions, same simulated browsing histories, and controlled experimental environment supported internal validity.

“Random assignment to conditions, same simulated browsing histories, and controlled experimental environment supported internal validity.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The study received ethical approval from concerned experts.

“This study received ethical approval from concerned experts of the concerned field.”

May combine two claims.

No significant differences between conditions were observed on any demographic variable (all p > 0.10), confirming successful randomization.

“No significant differences between conditions were observed on any demographic variable (all p > 0.10), confirming successful randomization.”

May combine two claims.

LLMs can generate personalized email subject lines, email body copy, product descriptions, chatbot responses, and brand voice directives.

“In e-commerce marketing, LLMs have the ability to generate personalized email subject lines, email body copy, product descriptions, chatbot responses, and even brand voice directives (Peres et al., 2023).”

• cites Peres et al., 2023 sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

For the A/B experiment, participants who failed the attention check were excluded prior to analysis (n = 41 excluded, final N = 1,159).

“For the A/B experiment, any participant who failed the attention check (e.g., selecting the same response for all items) was excluded prior to analysis (n = 41 excluded, final N = 1,159).”

Classified as probably not a substantial claim.

The human-written emails used in the AI detection task were verified as human-written via author correspondence with the brand's marketing team.

“The other was an actual email of a commercial brand (e.g., Everlane, Uniqlo) that was verified to have been human-written (via author correspondence with the marketing team of the brand).”

Probably not a substantial claim: may not pass S3 (load-bearing).

The order of the two emails (left/right) in the detection task was randomized.

“The order (left/right) was randomized.”

May combine two claims.

External validity was enhanced by recruiting a diverse U.S. sample of real online shoppers and by triangulating with practitioner interviews.

“External validity was enhanced by recruiting a diverse U.S. sample of real online shoppers and by triangulating with practitioner interviews.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

No personally identifiable information was gathered, participant IDs were anonymized before analysis, and interview recordings were kept on encrypted university servers to be removed after three years.

“No personally identifiable information (PII) was gathered.Anonymized before analysis, prolific participant IDs were anonymized.Recordings of the interviews were kept in encrypted university servers which will be removed after three years.”

May combine two claims.

The study identifies a detection-disclosure paradox in which consumers cannot recognize AI content but remain suspicious when it is disclosed.

“It identifies a “detection-disclosure paradox”-where consumers cannot recognize AI content but remain suspicious when it is disclosed”

Probably not a substantial claim: may not pass S3 (load-bearing).

All participants signed a consent form explaining purpose, procedures, risks, and right to abandon.

“All participants signed a consent form that was explained to them and signed digitally with the purpose, procedures, risks, and their right to abandon the procedure at any time without being fined.”

May combine two claims.

Consumers perceive AI as more competent but less warm compared to humans.

“Castelo et al. (2019) suggest a single framework: consumers perceive AI as more competent but less warm compared to humans.”

• cites Castelo et al. (2019) sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

The manipulation concern scale had Cronbach's alpha = 0.84.

“Concern in Manipulation (RQ2 exploratory): A 3-item scale designed in this paper.Sample item: I believe that AI-personalized content attempts to influence my buying behaviour.α = 0.84.”

Classified as probably not a substantial claim.

The study planned N = 1,200 overall (600 per condition) to account for attrition, attention check failures, and smaller effects.

“We planned N = 1,200 overall (600 per condition) to consider any possible attrition, attention check failures, and to capture smaller effects.”

Classified as probably not a substantial claim.

Participants were randomly assigned to the rule-based control and GenAI treatment conditions via Prolific's randomizer.

“The rule-based (control) and GenAI (treatment) condition were also randomly assigned to the participants via the randomizer of Prolific.”

Probably not a substantial claim: may not pass S4 (based).

No study has directly compared GenAI and rule-based personalization in a controlled e-commerce setting measuring both performance and perception outcomes.

“But no study has directly compared the two approaches in a controlled e-commerce setting measuring both performance and perception outcomes.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Algorithm aversion is especially acute in subjective, creative, or morally charged areas.

“Later studies have demonstrated that the issue of algorithm aversion is especially acute in subjective, creative, or morally charged areas (Castelo et al., 2019;Bigman and Gray, 2018).”

• cites Castelo et al., 2019 sourceNot checked — no full text of the source
• cites Bigman and Gray, 2018 sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

No study has examined detection accuracy for personalized marketing content in an e-commerce context, nor the downstream effects of detection on trust and purchase intentions.

“No study has examined detection accuracy for personalized marketing content in an e-commerce context, nor the downstream effects of detection on trust and purchase intentions.”

May combine two claims.

GenAI models such as LLMs and diffusion models generate new content dynamically based on context, enabling hyper-personalized content at scale.

“Contrary to rule-based systems, which look up and combine existing material, GenAI models, such as large language models (LLMs) like GPT-4 and image generators like DALL•E 3, generate new content dynamically based on the context.”

• cites Jarek & Mazurek, 2023Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In debriefing, the simulated e-commerce shop (UrbanHarbor) was revealed as a simulation, and participants were told they would not actually make purchases and browsing history simulation did not collect actual browsing.

“In the debriefing the simulated e-commerce shop (UrbanHarbor) was revealed as a simulation.The participants were made aware that they would not actually make any purchases and that the browsing history simulation did not entail collection of their actual browsing.”

May combine two claims.

Rule-based personalization systems are template-based and cannot generate original content tailored to individual users.

“To begin with, they are template-based: the information they provide is assembled but not created.”

May combine two claims.

Diffusion models enable personalized banner images, product visualizations, and dynamic ad creatives that change based on user attributes.

“In e-commerce, diffusion models enable personalized banner images, product visualizations (e.g., a sofa in the user's preferred colour and room setting), and dynamic ad creatives that change based on user attributes (Jarek & Mazurek, 2023).”

• cites Jarek & Mazurek, 2023Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

The privacy concern scale had Cronbach's alpha = 0.87.

“Privacy Concern (H4): Measured on the scale of 4 items of Internet users information privacy concerns (IUIPC) of Malhotra et al. (2004).Sample item: “I fear that the AI that personalizes this content is aware of too much about my personal information (1 = strongly disagree, 7 = strongly agree).α = 0.87.”

• cites Malhotra et al. (2004) sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

The loss of human touch scale had Cronbach's alpha = 0.81.

“Loss of Human touch (RQ2 exploratory): A 3-item scale.Sample item: This brand is more of a robot than a person.α = 0.81.”

May combine two claims.

There were no statistically significant age, gender, income, or past online purchase frequency differences between conditions (all p > 0.10), confirming successful randomization.

“There were no statistically significant age, gender, income, or past online purchase frequency differences between conditions (all p > 0.10), which confirmed successful randomization.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The perceived authenticity scale had Cronbach's alpha = 0.89.

“Perceived Authenticity (H3): This will be measured with a 5-item brand authenticity instrument (Morhart et al., 2015) modified to include the email content instead of brand in general.Sample question: The marketing content that I have just watched was authentic (1 = strongly disagree, 7 = strongly agree).α = 0.89.”

• cites Morhart et al., 2015 sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Two authors independently coded two transcripts to establish inter-coder reliability (Cohen's κ = 0.84); discrepancies were resolved through discussion.

“Two authors independently coded two transcripts to establish inter-coder reliability (Cohen's κ = 0.84); discrepancies were resolved through discussion.”

Probably not a substantial claim: may not pass S3 (load-bearing).

State-of-the-art LLMs hallucinate in 3-15% of generated outputs, depending on the domain.

“Early evidence suggests that even state-of-the-art LLMs hallucinate in 3-15% of generated outputs, depending on the domain (Ji et al., 2023).”

• cites Ji et al., 2023Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S4 (based).

Empirical data on consumer perceptions of ethical risks in GenAI personalization are scarce.

“While scholars have theorized about risks such as privacy concerns, manipulation, and loss of human touch, empirical data on consumer perceptions of these risks in the context of GenAI personalization are scarce.”

Classified as probably not a substantial claim. May combine two claims.

GenAI is not inherently manipulative but poses new risks by its ability to produce content at scale customized to specific psychological profiles.

“Although GenAI is not inherently manipulative, it poses new risks by its ability to produce content at scale that is customized to specific psychological profiles.”

Probably not a substantial claim: may not pass S3 (load-bearing).

GenAI is particularly effective for utilitarian product categories.

“It is particularly effective for utilitarian product categories, including electronics, office supplies, and household goods, where efficiency and scalability are priorities.”

May combine two claims.

Nondisclosure can be functionally invisible and maintain performance benefits, but denies consumers informed choice and can be against new regulations.

“Nondisclosure can be functionally invisible and maintain performance benefits, but denies consumers informed choice and can be against new regulations.”

May combine two claims.

Visual specificity of DALL•E images can align with user aesthetic preferences and augment mental simulation, a recognized stimulus of purchase intention.

“This visual particularity can augment mental simulation (i.e., visualizing oneself with the product), a recognized stimulus of purchase intention (Elder and Krishna, 2012).”

• cites Elder and Krishna, 2012 sourceNot checked — no full text of the source

Classified as probably not a substantial claim.

The experiment was conducted at only one point in time.

“The time we conducted our experiment was only at one point.”

May combine two claims.

GenAI content can produce immediate behavioural outcomes (clicks, purchases) but does not significantly alter overall brand perceptions when AI generation is hidden.

“It indicates that although GenAI content can result in immediate behavioural outcomes (clicks, purchases), it does not significantly alter, either positively or negatively, the overall brand perceptions when hidden.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The study used GPT-4 (text) and DALL•E 3 (image).

“We made GPT-4 (text) and DALL•E 3 (image).”

May combine two claims.

Consumers are unable to reliably recognize AI-generated content but do not trust it when revealed.

“Third, consumers are unable to reliably recognize marketing material generated by AI (accuracy = 53.3,not significantly higher than that of chance) but do not trust it when revealed.”

Classified as probably not a substantial claim. May combine two claims.

The study's dependent variables were short-term behavioural (CTR, conversion, time-on-site) and instant perceptions (authenticity, trust).

“Short-term behavioural (CTR, conversion, time-on-site) and instant perceptions (authenticity, trust) were our dependent variables.”

May combine two claims.

Disclosure honours consumer autonomy but decreases trust and intention to purchase.

“Disclosure honours consumer autonomy but decreases trust and intention to purchase.”

May combine two claims.

GenAI-generated content is novel and individually created, which may draw attention by being unexpected and exclusive.

“Conversely, GenAI-generated content is novel in nature.The emails, subject lines, banner images are created individually to that user.The lack of repetition that is templated might draw attention by being unexpected and exclusive.”

May combine two claims. The quote only partly states this claim.

Disclosing AI generation reduces authenticity, trust, and purchase intention.

“Disclose AI generation Reduced authenticity, lower trust, lower purchase intention (as shown in our post-detection trust analysis: trust dropped from 4.85 to 3.94, d = -0.65)”

May combine two claims. The quote only partly states this claim.

The performance advantages of GenAI personalization are at the cost of high consumer perception (lower authenticity, trust, purchase intention when disclosed).

“Second, though, these performance advantages are at the cost of high consumer perception.”

May combine two claims.

GenAI is more effective than rule-based approaches on behavioural metrics (CTR, conversion), and imposes authenticity penalties on disclosure.

“The fact that we find GenAI to be more effective than rule-based approaches on behavioural metrics (CTR, conversion), and imposes authenticity penalties on disclosure, offers a fine-grained, empirically-based perspective on the costs and benefits of GenAI adoption.”

May combine two claims.

Consumers are not reliably able to detect AI-generated marketing content, but when informed that content is AI-generated, trust and purchase intentions decrease significantly.

“We report a paradoxical result: consumers are not reliably able to detect AI-generated marketing content (detection accuracy at chance), but when informed that content is AI-generated, the level of trust and purchase intentions decreases significantly.”

Classified as probably not a substantial claim.

Respondents who had recently purchased online within the last 30 days were chosen to enhance ecological validity.

“Mitigation: we also chose respondents who had recently purchased online within the last 30 days to enhance the ecological validity.”

May combine two claims.

The experiment used an emulated e-commerce storefront (UrbanHarbor), fake products, and no actual financial exchange.

“In our experiment we have used an emulated ecommerce storefront (UrbanHarbor) and fake products and no actual financial exchange.”

May combine two claims.

The qualitative sample consists of six managers representing three brands, providing in-depth information but not statistically generalisable.

“Our qualitative sample, which consists of six managers representing three brands, will provide some in-depth information, but it will not be statistically generalisable.”

• AssumptionRests on an assumption: Not all brands share the same experiences, as they may differ in size or product lines.

May combine two claims.

Compulsory disclosure can decrease brand performance without consumer protection because consumers cannot identify AI content anyway.

“Compulsory disclosure (the new policy consensus) can have a negative impact on the brand performance without consumer protection-as consumers cannot identify AI content in any case, disclosure can only decrease trust without allowing them to make an informed choice.”

May combine two claims.

Consumers are not consistently opposed to GenAI personalization but prefer to have control over when and how it is applied.

“It posits that consumers are not consistently opposed to GenAI personalization but will prefer to have it under control as to when and how it is applied.”

May combine two claims.

Consumers have a background assumption that marketing content is created by humans unless otherwise indicated; authenticity is harmed when that assumption is broken.

“Second, it implies the possible existence of a background assumption that the consumers might have that marketing content is created by humans unless otherwise indicated.Authenticity is harmed when we break that assumption.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Informed consent was obtained from all individual participants included in the study.

“Informed consent was obtained from all individual participants included in the study.”

Probably not a substantial claim: may not pass S3 (load-bearing).

All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.

“All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.”