DeepSeek

What it is

DeepSeek builds and publishes large language models, with two main lines: V3 style general models and R1 style reasoning models that produce visible chains of thought before answering. Model weights have been released publicly under permissive terms, so developers can run them on their own infrastructure rather than relying only on a hosted API. The company also offers a consumer chat assistant and a developer API.

Why it matters

DeepSeek matters to search and marketing teams for two reasons. First, it is another surface where people ask questions and receive synthesised answers, so brands can appear or be absent in those responses. Second, low cost open weight reasoning models change the economics of AI features, making it viable to run classification, summarisation and content workflows in house at a scale that would be expensive with frontier commercial APIs.

How it works

Teams use DeepSeek models either through the hosted API and chat interface or by self hosting the open weights on their own or a cloud provider's GPUs. In practice, marketers test brand and category prompts in the assistant to see how the brand is described and which sources are used, while engineering teams benchmark the models against incumbents on cost, latency and output quality for specific tasks. Data residency and governance reviews are common before any production use, given the models and hosted services originate outside the UK and EU.

When it applies

It is relevant when auditing brand visibility across multiple AI assistants rather than just the best known ones, and when evaluating model options for cost sensitive or privacy sensitive workloads that benefit from self hosting.

Examples

• A B2B marketer runs the same twenty buying intent prompts through DeepSeek, ChatGPT and Gemini to compare how the brand and its competitors are described.
• A publisher self hosts an open weight DeepSeek model to tag and summarise a large archive without sending content to a third party API.
• An engineering team benchmarks a DeepSeek reasoning model against its current provider on a structured data extraction task, comparing accuracy and cost per thousand records.

How it is measured

• Share of tested prompts where the brand is mentioned in DeepSeek responses
• Accuracy of brand and product claims in those responses, checked against source pages
• Cost and latency per task versus incumbent models on the same benchmark set
• Referral sessions and assisted conversions attributed to AI assistant traffic where identifiable