Do AI-Generated Video Ads Work for Pakistan's Small Businesses? A Research Proposal

AI-generated video ad tools are spreading rapidly, allowing small merchants to produce voiced product ads at extremely low cost within hours. But whether these ads actually work — the existing experimental evidence comes almost entirely from Western or English-language contexts, and the conclusions contradict each other.

Recently, Muhammad Usman, an AI major at Superior University in Lahore, released a research proposal attempting to test this question in the context of Pakistani small and micro enterprises and the Urdu language.

An important caveat: the experiment has not actually been run yet. The author explicitly marks this as a research proposal inviting collaboration and feedback, not completed empirical results.

The Contradictory Landscape of Existing Evidence

The proposal reviews several key studies, and their conclusions are inconsistent:

• MIT/Harvard's WhatsApp experiment (21,000 consumers) reported a 9.4 percentage point increase in click-through rate for AI-generated ads.
• A June 2026 HKBU field experiment (150+ ads) found that AI-generated video ads had higher click-through and retention rates but lower conversion rates.
• A 2026 Syracuse/Ipsos study of 3,000 U.S. consumers showed that AI ads performed 14% worse on short-term effectiveness and 17% worse on brand equity.
• Another study found that when disclosure of AI involvement was made to viewers, click-through rates dropped by about 31.5%.

These findings conflict with one another, and none of them has examined the scenario of Urdu-language AI video ads, targeting Pakistani SMEs, distributed via WhatsApp — which is precisely the channel on which a large number of small businesses in Pakistan depend.

Click-through rate, conversion rate, and brand equity are three indicators at different levels in advertising effectiveness evaluation. Click-through rate measures the efficiency of capturing attention and belongs to the shallowest level of behavioral response; conversion rate reflects actual purchase or inquiry behavior and is directly tied to revenue; brand equity refers to long-term intangible value such as consumer awareness, associations, and trust toward a brand. The three often diverge — an ad may achieve high clicks due to novelty but fail to drive purchases due to a lack of persuasiveness, or even damage brand image through an overly promotional feel. This is precisely the root of the contradictions among the studies above: the MIT/Harvard experiment used click-through rate as its main metric and reached a positive conclusion, while the Syracuse/Ipsos study incorporated brand equity and found negative effects. Understanding this distinction is a key prerequisite for judging whether AI ads are effective.

AI disclosure refers to explicitly informing viewers within an ad that the content was generated by artificial intelligence. This practice stems from advertising transparency ethics and the compliance requirements of some platforms, but its effects are disputed. Existing research shows that after disclosing AI involvement, click-through rates drop by about 31.5%. A possible mechanism is "algorithm aversion" — once viewers realize the content was not created by a human, suspicion about authenticity and sincerity is triggered, reducing willingness to engage. However, disclosure may also increase trust, especially among audiences that are open to AI. Therefore, the disclosure effect itself is a variable that needs to be measured separately, rather than a fixed negative factor. Understanding the separability of disclosure and generation effects helps explain why different studies evaluate AI advertising so differently.

Four Core Research Questions

The proposal puts forward four questions to be tested:

RQ1: In a WhatsApp broadcast list distribution scenario, do Urdu-language AI-generated video ads outperform manually produced ads in click-through rate and retention?

RQ2: Will the conversion rate gap found in Western studies reappear for low-priced everyday products among Pakistani SMEs?

RQ3: Does disclosing the AI-generated nature of an ad reduce engagement among Pakistani audiences?

RQ4: Is synthetic Urdu voiceover a trust bottleneck — that is, does audience perception depend on the voice or the visuals?

The fourth question deserves a closer look. Most discussions focus on the overall label of "AI advertising," but the proposal breaks it down into a specific variable: viewers' discomfort may stem primarily from AI voice rather than visuals. If true, this means the optimization priorities for localized AI ads would differ from common assumptions.

Synthetic speech refers to AI voiceover generated through text-to-speech or voice cloning technology. Compared with visuals, the "unnaturalness" of speech is often easier for viewers to notice, because humans are extremely sensitive to intonation, pauses, and emotional fluctuations. In cross-language scenarios, Urdu, as a low-resource language, usually has synthetic speech naturalness lower than that of mainstream languages such as English, which may cause viewers to perceive a "machine feel" or distrust. RQ4 breaks "voice" and "visuals" into independent variables precisely to test which modality contains the trust bottleneck. If the bottleneck is indeed in speech, then the optimization focus for localized AI ads should shift toward voiceover quality rather than visual effects. This distinction has hardly been systematically tested in existing literature.

Proposed Method: A Randomized Field Experiment

The proposal designs a replicable experimental protocol:

• Design: A randomized field experiment targeting 8–12 Lahore SMEs.
• Stimuli: Paired AI-generated and manually produced 15–30 second vertical ads for the same product, holding product, offer, and price constant.
• Distribution: Randomized WhatsApp broadcast groups, with an additional AI disclosure treatment group.
• Metrics: Attention (completion/retention), conversion (inquiries, orders), perception questionnaires, and cost per ad.
• Analysis: Preregistration, informed consent from all participants, and anonymized aggregate reporting.

The combination of a paired design and a disclosure treatment group effectively treats "AI generation effect" and "AI disclosure effect" as separable variables to be measured. This is uncommon in existing literature — most studies measure only one of them.

Preregistration means that before collecting data, researchers publicly register their research hypotheses, experimental design, primary metrics, and analysis plan on a verifiable platform (such as OSF or AsPredicted). Its core purpose is to prevent "p-hacking" and post hoc selective reporting — that is, researchers adjusting hypotheses after seeing the data or reporting only significant results, thereby exaggerating effects. Against the backdrop of frequent controversies over AI advertising effectiveness, preregistration can significantly enhance the credibility of conclusions. The proposal explicitly requires preregistration, informed consent, and anonymized aggregation, indicating that the author is attempting to respond to the contradictory conclusions in existing literature with rigorous experimental norms, making the results replicable and comparable across contexts.

Expected Contributions

The proposal lists three expected contributions:

1. The first empirical evidence on the effectiveness of AI video advertising for Pakistani SMEs, in Urdu, in a WhatsApp context.
2. Practical guidance for whether Pakistani small merchants should adopt AI video advertising.
3. A replicable, low-cost experimental protocol for other researchers and practitioners to use.

Why This Gap Is Worth Filling

From the existing literature, the effects of AI advertising are not monotonically "good" or "bad": click-through rates may rise, but conversion and brand perception may fall; disclosing AI identity may further harm engagement. Whether these effects are stable across different markets, languages, channels, and product price points currently has no answer.

The particularity of Pakistani SMEs lies in the channel (WhatsApp rather than Meta/Google ad systems), the language (Urdu rather than English), and the product structure (mainly low-priced everyday consumer goods). The combination of these three makes directly transferring Western conclusions clearly risky.

The author is looking for collaborators — researchers, local Lahore SME owners, or AI practitioners interested in this question. Contact: su92-bsaim-f25-025@superior.edu.pk

It must be emphasized again: this is a research proposal, and the experiment has not yet been run. Its value lies in clearly defining the problem, method, and hypotheses to be tested, rather than providing conclusions.

One-line Summary

A Pakistani researcher proposes the first randomized field experiment on AI video ads for Urdu-language SMEs, not yet run.

Paragraph Summary

AI-generated video ad tools are spreading rapidly, but existing experimental evidence comes almost entirely from Western contexts and is contradictory: MIT/Harvard reports a 9.4 percentage point increase in click-through rate, HKBU finds lower conversion rates, and Syracuse/Ipsos shows 14% weaker short-term effects and 17% weaker brand equity, while disclosing AI involvement reduces click-through rates by about 31.5%. Muhammad Usman of Superior University in Lahore has released a research proposal to conduct a randomized field experiment among 8–12 Lahore SMEs, distributing paired Urdu AI and manually produced ads via WhatsApp broadcast, with an AI disclosure treatment group, and separately examining whether synthetic voice constitutes a trust bottleneck. The experiment has not yet been run, and the author is seeking collaborators.