AI methodology

Product research: ranking trending products against the Winning Score framework. AI weighs the inputs; the inputs themselves are deterministic marketplace and search signals.

Listing generation: titles, descriptions, bullets, alt text, and ad scripts. Each marketplace has a separately fine-tuned model that learned from millions of high-performing listings on that channel.

Pricing recommendations: AI suggests price points based on competitor monitoring; final decisions stay within the operator-set guardrails.

Order routing, inventory sync, fulfillment, and payment processing are deterministic code, not AI. We don't use language models to make decisions about money or stock.

We use a mix of fine-tuned open-source models (Llama 3.1, Qwen 2.5) and commercial APIs (OpenAI GPT-4 class, Anthropic Claude 3.5). Listing models are fine-tuned on 14M Amazon listings, 8M Shopify product pages, and 4M Etsy listings — all licensed or publicly accessible data.

Customer catalog data is never used to train models that other customers use. Each customer's tone-of-voice fine-tune is private to their workspace.

Every model change is evaluated against a held-out test set of 1,200 real listings before release. Metrics: keyword coverage, readability, conversion lift (measured via opt-in A/B testing with consenting customers), and hallucination rate.

A model only ships if it beats the current production version on at least three of four metrics with no regression on the fourth.

AI-generated copy can occasionally hallucinate product specs that don't exist. Every listing should be reviewed before publishing. Sellrith's UI stages every generation in a diff view to make this review fast.

Pricing suggestions are not financial advice. Operators are responsible for the final price and margin decisions.