A model vendor building a search engine is not new. A model vendor building one and advertising that it filters AI-generated pages is.
On October 17 DeepSeek opened DeepSeek Search as a public API. The index is described as running into the hundreds of billions of pages, with multilingual coverage and a freshness requirement: results have to be current, not merely relevant. The pitch to developers is ordinary. The screening rules underneath are not.
What was released
Two things landed the same day. The Search API itself, callable on its own without touching a model, and a matching Fetch endpoint that returns page text. Both are metered per call at $0.75 per thousand, the same rate as the web tools that later wrapped them. Because the endpoints are stateless, they slot into pipelines that already have their own orchestration: retrieval for a RAG store, a fact-check step, a competitor monitor.
Then there is the part the announcement spends most of its words on. DeepSeek says the ranking layer actively filters three categories: conventional SEO spam, GEO manipulation, and low-quality AIGC.
Why the third filter is awkward
GEO — generative engine optimisation — is the practice of shaping pages so that answer engines quote them. It is, functionally, SEO retargeted at models. A search engine owned by a model company has an obvious incentive to suppress it: every manipulated page that reaches the context window is a wrong answer attributed to the model.
Low-quality AIGC is the harder call. DeepSeek's own models generate text at a price that makes bulk publishing close to free, and its API is now sold to anyone. A filter that demotes machine-written pages protects answer quality and simultaneously constrains the cheapest use of the company's own product. Both things are true at once.

What this means if you publish
Three practical readings, none of them comfortable.
First, the era of padding a page to rank is ending on both sides of the fence. Google has spent two years demoting scaled content; now a model vendor with a search index is doing the same thing for retrieval. If your page exists to fill a keyword gap, it is now failing two independent tests instead of one.
Second, specificity is the only durable signal. A page that reports a number nobody else has — a price you checked, a benchmark you ran, an error message you reproduced — is expensive to fake and cheap to verify. That is what survives a filter trained to spot the opposite.
Third, do not optimise for the filter, optimise for the citation. The measurable goal is not "rank" but "get quoted correctly". Being quoted requires a claim stated plainly enough to lift, in a sentence that still makes sense out of context. Pages built entirely of hedged paragraphs never get quoted, however clean they look to a classifier.
[IMG2]
The caveats
No vendor publishes its filter thresholds, and DeepSeek has not published evaluation numbers for the spam classifier. Treat the marketing language as a direction, not a guarantee. The pricing is the verifiable part: $0.75 per thousand searches and per thousand fetches, against roughly $10 per thousand for comparable search tooling from OpenAI and Anthropic, is a real gap and the likeliest reason developers will try it.
There is also a structural question nobody has answered. If the index is trained to distrust machine-written text, and the same company sells the machine, the filter's real job may be less about quality than about keeping the answer layer clean while the generation layer stays cheap. That is a reasonable business position. It is worth knowing which side of it your site sits on.
The practical test is small. Take your three best pages, run the questions they are meant to answer through the Search API, and read what comes back. If a competitor's page is quoted and yours is not, you have learned more in ten minutes than in a month of keyword research.