# Memes.ai: proposed creative-pattern experiment Protocol version 0.1 · September 12, 2026 · Planning document; not preregistered or launched. ## Research question Does explicitly supplying a recognizable cultural pattern improve advertising outcomes compared with generic AI creative produced from the same business brief? What changes in production effort and concept diversity? ## Hypotheses The primary null hypothesis is no difference between arms C and B in the preselected business-conversion rate under the experiment’s assignment. Use a two-sided test: a pattern can help, have no effect, or hurt. C versus A is secondary. Predeclare multiplicity handling if making confirmatory claims about additional comparisons. ## Experimental arms ### A. Conventional creative A professional product-led brief. Use the same brand, offer, format, and production allowance as the other arms. Build a credible control. ### B. Generic AI creative The same brief, without supplied cultural patterns. Match the generator version, production allowance, and review process to arm C wherever feasible. ### C. Pattern-informed creative The same brief, with an explicit pattern and brand connection. Use a recorded pattern family and multiple distinct premises. Preserve the common offer and destination. ## Design and analysis commitments **Primary comparison.** C versus B tests the added value of explicitly supplying a cultural pattern. C versus A is a secondary comparison of creative strategies. **Assignment.** Use a randomized platform experiment when supported; keep eligible audiences mutually exclusive. Document the platform’s delivery and optimization rules. **Primary outcome.** Preselect a verified business conversion and report its rate using the declared assignment denominator. Include acquisition cost as an operational outcome. **Production controls.** Match the brief, offer, destination, number of candidate concepts, review rubric, and production allowance. Record model versions, edits, and unavoidable differences. **Analysis.** Fix the sample-size calculation, minimum meaningful effect, stopping rule, and attribution window before launch. Publish uncertainty and all tested variants. **Publication status.** Protocol proposed. Brands, budget, sample size, dates, and experimental assignment are not yet committed. No performance results are available. Randomization reduces audience confounding but does not force an ad platform to deliver identical exposure. Record eligibility, actual delivery, spend allocation, and optimization. When assignment-level outcomes are available, analyze the assigned groups. When only campaign aggregates are available, clearly limit the estimand to the experiment and reporting system that produced them. Use several independently developed concepts in each arm. Specify the number before production. Evaluate candidates with the same brand-safety, factual-accuracy, and offer-relevance rubric; log every rejected or revised candidate. Do not give the pattern arm extra selection attempts without reporting that difference. ## Complete before launch - Participating brands, sectors, countries, eligibility and exclusion rules: [pending] - Business conversion event and verification source: [pending] - Unit of assignment and primary analysis denominator: [pending] - Baseline outcome rate, minimum meaningful effect, power and significance threshold: [pending] - Sample-size calculation and adjustment for clustering/repeated exposure: [pending] - Total budget, allocation, production allowance and number of concepts: [pending] - Start/end dates, minimum observation window and stopping rule: [pending] - Attribution window, refunds/cancellations and outcome-maturation rules: [pending] - Analysis owner, preregistration location, and release date: [pending] ## Data and reporting The accompanying CSV is an empty collection template. A blank field means unknown, not zero. Keep direct identifiers and customer-level records out of public releases. Preserve pseudonymous joins internally so verified business outcomes can be reconciled with campaign exports. Report denominators, absolute differences, relative differences where meaningful, and uncertainty intervals. State missingness, exclusions, failed variants, assignment imbalance, protocol deviations, and the time window. Keep raw spend in its original currency unless a documented exchange-rate method is applied. Report brand-level effects before choosing a pooled estimator. The weighting strategy must match the target question: the average advertiser and the average delivered exposure are different estimands. Avoid claiming industry-specific effects from underpowered subgroups. Separate production results from market results. Minutes saved and a greater number of distinct concepts can be useful outcomes without establishing incremental sales. A creative experiment compares creative strategies; a no-ad holdout tests the incremental value of advertising. ## Separate landing-page experiment To compare /whitepaper with the API docs, use an ad promise that both destinations fulfill, such as “See how meme advertising works.” Randomize the destination while holding the campaign objective, audience, offer, and ad constant. Confirm cross-domain attribution to the same verified signup or purchase event. Judge by qualified signups or paid customers after an equal maturation window; report reading and preview generation as diagnostics. The research-reaction ad says “Read the whitepaper” and belongs on /whitepaper. Comparing that ad/page combination with historical API-docs traffic estimates neither a clean page effect nor a causal improvement. ## Supporting methodology Gordon et al. (2019), A Comparison of Approaches to Advertising Measurement: Evidence from Big Field Experiments at Facebook. https://pubsonline.informs.org/doi/10.1287/mksc.2018.1135 This protocol is a proposal. It does not authorize campaign spending, describe a registered experiment, or supply measured results.