Polywood used a pragmatic mix of public signals, machine learning and generative AI to drive measurable ecommerce gains: the company reports a 22% uplift in conversion rate and a 12% rise in average order value (AOV). For marketers and ecommerce leaders, the value here is concrete—Polywood turned weather and housing data into timing signals, made model outputs queryable, and automated creative at scale to improve both ad efficiency and onsite experience.

Data, signals and an internal large language model (LLM)

Polywood’s forecasting rests on three external predictors it links to first-party purchase history: regional weather, home-sale activity and building permits. The retailer ingests national weather feeds and multiple listing service (MLS) real-estate data, then matches those feeds to historical orders to surface purchase patterns tied to housing events.

That analysis produced a counterintuitive operational finding: a seven-day weather forecast proved a stronger near-term sales indicator than a day-of, three-day or 14-day forecast. Polywood trained machine-learning models on historical weather and order data to validate that signal and then used it to time marketing—deciding when to send catalogs or serve regionally targeted ads.

Polywood also built an internal operating system around a large language model (LLM) with chatbot-like query capability. Analysts and marketers can ask the system factual questions about past sales and trends and get actionable outputs that inform retargeting rules, persona definitions and campaign timing. The company also uses Anthropic’s Claude to assist with code construction.

Personas, creative and measurable outcomes

Rather than rely primarily on psychographic segments, Polywood’s personas are anchored to home types and lifecycle events. Historical orders showed timing differences by house attributes: buyers of recently purchased homes tend to buy outdoor furniture within about six months, while buyers of homes with pools often buy within a month.

Those personas drive three practical outputs: creative selection, attribution inputs and targeting parameters. Polywood feeds its product models into a generative pipeline to render lifestyle imagery at scale across roughly 150,000 SKUs and color permutations, and it generates ad copy variants automatically. Spiegel, Polywood’s chief digital officer, attributes a 12% lift in AOV and a 22% improvement in conversion rate to the combined changes. The company also reports reduced cost-per-click after introducing AI-generated ad copy.

Polywood acknowledges attribution ambiguity—when ad copy, landing-page personalization and imagery roll out together, isolating each contribution is difficult—but the company documents net improvement across acquisition and on-site conversion metrics.

Platform migration and developer strategy

About 18 months ago Polywood migrated its ecommerce platform to Shopify. The platform change coincided with a move to what Spiegel described as “100% AI-based coding” for implementation tasks. Crucially, Polywood retained its development team rather than replacing staff; Spiegel reports developers are now roughly twice as fast at deploying features and capabilities when working with AI tools.

The company treats AI primarily as an augmentation layer for engineers, preserving institutional knowledge while accelerating delivery. That approach reduces the operational risk of outsourcing model-driven changes and keeps governance and data control in-house.

Practical takeaways for marketers and ecommerce teams

Polywood’s setup highlights several actionable practices:

  • Fuse contextual external signals with first-party transaction data. Weather and MLS feeds are most valuable when linked to customer behavior and lifecycle markers.
  • Expose model outputs through queryable tools. An LLM-powered interface that answers business questions makes the models actionable for marketers without constant engineering handoffs.
  • Test different signal horizons. Polywood’s seven-day forecast finding is testable and may not generalize—teams should validate the optimal horizon on their own data.
  • Keep developers involved. Using AI to boost developer productivity helps retain institutional knowledge and maintain control over deployment and data practices.
  • Accept attribution ambiguity but measure net lift. When multiple improvements launch together, prioritize unit economics and overall ROI over fine-grained attribution until experiments can isolate effects.

Polywood’s results are company-reported and specific to a high-AOV, seasonal category. That limits direct transferability, but the architectural pattern—signal fusion, a queryable internal LLM and integrated creative generation—is repeatable for retailers with access to comparable data and engineering resources.

Next steps for teams exploring similar work: identify high-signal external datasets you can link to transactions, pilot small ML models to validate timing hypotheses (for example, compare different weather horizons), and expose outputs through tools marketers can query. What to watch next: how Polywood governs model outputs at scale—image realism and copyright, bias and reproducibility—and whether the weather-plus-housing signal pattern proves useful across other categories and geographies.