Do AI-Generated Video Advertisements Work for Pakistan's Small Businesses?

Do AI-Generated Video Advertisements Work for Pakistan’s Small Businesses?

A Field-Study Proposal for Urdu-Language AI Video Ads in WhatsApp Commerce

Muhammad Usman | BS Artificial Intelligence, Superior University Lahore | Research Proposal | October 2026

Abstract

Generative-AI video advertisements are now cheap enough for small businesses to produce, and a growing body of experimental research measures their effectiveness — but every major study to date has been conducted in Western or English-language developing-market contexts. No published study examines AI-generated video ads in Urdu, for Pakistani small and medium enterprises (SMEs), distributed through WhatsApp, the channel on which much of Pakistan’s small-business commerce actually runs. This paper (i) surveys the existing experimental literature on AI-generated ad effectiveness, (ii) identifies the Pakistan/Urdu/WhatsApp gap as an unaddressed research frontier, and (iii) proposes a concrete, low-cost randomised field experiment — executable by an undergraduate researcher with cooperating Lahore businesses — comparing AI-generated versus human-produced Urdu video ads on click-through, retention, conversion, and brand perception.

1. Introduction

In 2024–2026, text-to-video and image-to-video models collapsed the cost of producing a polished video advertisement from thousands of dollars to a few dollars of compute. The question is no longer can AI make the ad, but does the AI-made ad work? A first wave of experimental studies has begun to answer this. Yet those answers come from contexts far removed from Pakistan: US consumers, English copy, Instagram and web placements. Pakistan’s SME commerce runs substantially through WhatsApp broadcast lists and status updates, in Urdu (often code-switched with English), for audiences whose media literacy and trust cues differ from Western panels. Whether the headline findings of the existing literature survive that translation is unknown — and that is the gap this proposal addresses.

2. Literature Review: What the Experiments Say So Far

2.1 Personalisation and engagement gains

Kumar and Kapoor (MIT Initiative on the Digital Economy / Harvard Business School) ran an online experiment with over 21,000 consumers, delivering AI-generated personalised video ads through a WhatsApp marketing provider. AI personalised videos achieved click-through rates 9.4 percentage points higher than personalised image ads and 6.5 points higher than generic video. A separate large-scale study reported average CTR of 28% versus 15% for traditional formats, with stronger emotional-engagement ratings (4.3 vs 2.7).

2.2 The conversion gap

A June 2026 randomised field experiment by Li and Li (Hong Kong Baptist University) on more than 150 unique video advertisements found that AI-generated ads achieved higher CTR, stronger viewer retention, and fewer negative reactions than human-created ads — but performed worse on conversions.

2.3 Brand-equity and disclosure penalties

Peruta and colleagues (Syracuse University, with Ipsos, 2026) tested 10 human-created and 10 AI-generated ads on 3,000 US consumers. Only 25% of viewers could identify the AI-made ads with confidence — yet AI ads were on average 14% weaker on short-term effectiveness and 17% weaker on long-term brand equity. Related industry research reports that when consumers know AI was involved, CTR drops by roughly 31.5%.

2.4 What remains consistent

Across studies: (a) AI-generated ads reliably win on attention metrics (CTR, retention, novelty), and (b) they lag on persuasion metrics (conversion, brand equity, perceived authenticity).

3. The Research Gap

Every study above shares three contextual features: Western or English-language audiences, formal ad placements, and established-brand products. None examines:

  • Urdu-language AI video ads — including Urdu–English code-switching — and whether synthesised Urdu voiceover carries the same persuasive weight as human speech.
  • Pakistani SMEs as advertisers — businesses with no brand equity to lose, for whom a Rs. 500 AI ad versus a Rs. 50,000 shoot is the actual decision.
  • WhatsApp as the distribution channel — broadcast lists and Status, where the ad arrives as a personal message.
  • Disclosure effects in a Pakistani audience — whether the ~31.5% CTR penalty for known-AI content replicates where AI-media literacy is lower.

To the author’s knowledge, no peer-reviewed or preprint study has been published on AI-generated video-ad effectiveness in any of these dimensions.

4. Research Questions and Hypotheses

  • RQ1: Do Urdu-language AI-generated video ads achieve higher CTR and retention than human-produced video ads for the same Pakistani SME products on WhatsApp? (H1: yes.)
  • RQ2: Do they convert worse, as in Li & Li (2026)? (H2: gap replicates but smaller for low-price products.)
  • RQ3: Does disclosing AI involvement reduce engagement among Pakistani viewers? (H3: penalty replicates but weaker than ~31.5%.)
  • RQ4: Does synthesised Urdu voiceover underperform human Urdu voiceover on trust, holding visuals constant? (H4: yes — voice is the authenticity bottleneck.)

5. Proposed Methodology

5.1 Design

A randomised field experiment with 8–12 cooperating SMEs in Lahore. For each business, two 15–30 second vertical video ads for the same product and offer: one AI-generated (generative pipeline with synthesised Urdu voiceover) and one human-produced (smartphone-shot, human Urdu voiceover), matched on script, duration, and call-to-action. Each business’s WhatsApp broadcast list is randomly split; halves receive one version. A subset of AI-ad recipients additionally receives an AI-disclosure label.

5.2 Metrics

  • Attention: view-through rate, average watch time, CTR on the order/catalogue link.
  • Conversion: inquiries and completed orders per 1,000 recipients, 7-day attribution.
  • Perception: post-exposure survey (n ≈ 200) on trust, quality, purchase intent; AI-identification check.
  • Cost: production cost and turnaround per ad.

5.3 Analysis

Paired comparisons per business, aggregated with mixed-effects models; pre-registered hypotheses RQ1–RQ4; α = 0.05 with confidence intervals throughout.

5.4 Ethics

Businesses give informed consent; customers receive genuine offers at real prices; the disclosure arm satisfies transparency; all data anonymised and reported in aggregate.

6. Expected Contributions

  • First empirical evidence on AI video-ad effectiveness in a South-Asian, Urdu-language, WhatsApp-commerce setting.
  • Practical guidance for Pakistan’s SMEs: when AI ads are cost-effective, where the conversion gap bites, whether disclosure labels help or hurt.
  • A replicable low-cost protocol other undergraduate researchers can reuse, including bilingual survey instruments.
  • Voiceover insight: isolating whether synthesised Urdu voice — rather than AI visuals — is the binding constraint on trust.

7. Limitations

WhatsApp broadcast delivery is subject to platform limits and list-quality variation; Lahore SME results may not generalise to rural markets or other languages; novelty effects may decay — a follow-up wave after 6–12 months is advisable. Long-term brand equity requires longitudinal design beyond this proposal’s scope.

8. Conclusion

The experimental literature has established that AI-generated video ads win attention and lose persuasion — in the West, in English, on formal placements. Whether that trade-off holds for an Urdu-speaking shopkeeper’s customers receiving a video on WhatsApp is not a footnote; for Pakistan’s digital economy, it is the main question. This proposal lays out a feasible, ethical, undergraduate-executable path to answering it.

References

  1. M. Kumar and A. Kapoor, “Frontiers: Generative AI and Personalized Video Advertisements,” MIT Initiative on the Digital Economy Research Brief, 2025.
  2. H. Li and Q. Li, randomised field experiment on 150+ video ads, conference paper, June 2026.
  3. A. Peruta et al. (with Ipsos), human-made vs AI-generated ads, 3,000 US consumers, 2026.
  4. “The Impact of Personalized AI-Generated Video Ads on Consumer Click-Through Rates,” SCIRP, 2025.
  5. OM Media Trials / Zefr, brand reactions to AI-generated content adjacency, ~5,000 US/Canada respondents, 2026.
  6. Industry synthesis on visual generative-AI ad effectiveness and disclosure effects, 2026.

Methodological note: this is a research proposal, not a completed experiment. Cited effect sizes are as reported in the referenced studies’ public abstracts and coverage; the proposed experiment has not yet been run.

Comments

Popular posts from this blog

The Kalera Biradari: A Historical and Ethnographic Inquiry into a Punjabi Muslim Jat Gotra

Switch-Point Prosody: Why AI Voices Break Where Pakistani Ads Switch Languages