On October 17, Chinese AI company DeepSeek released its in-house search engine as an API for developers. Each search result comes with an authority rating for the page and a score indicating its relevance to the query. DeepSeek says the system filters out SEO spam, GEO manipulation, and low-quality AI-generated content.
The dawn of search results with scores
The newly released Search API is essentially a search-engine component: send it keywords, and it returns web search results. DeepSeek says the engine already powers search on its website and APIs, and handles an index of hundreds of billions of pages as well as hundreds of millions of search requests from around the world each day.
What stands out is the information included with each result. Every page gets an authority rating from 0 to 4, or “-1” if the rating cannot be determined. A score of 3 or higher marks a source as having “high authority.” Each result also gets a relevance score from 0 to 3, taking into account not only how well it matches the query but also its authority and freshness. A score of 1.5 or higher is considered an indicator of high quality.
In other words, search results come with a built-in “trust indicator.” The design effectively delegates some of the work of selecting source material for AI-generated answers to the search engine.
Data designed for AI to read
The returned data is not designed solely for display on a human-facing screen. Short snippets for display are typically around 100–200 characters, while summaries for models to read are around 600–800 characters. The API can also fetch the full page on request, narrow results to a domain using site:, and filter by language or publication date.
Search and Fetch each cost $0.75 per 1,000 successful requests. Searches that return no results are not billed, and the price is the same whether you request 50 results or just one. That works out to less than one cent per site: search.
For those using the DeepSeek API through Claude Code, Codex, or DeepSeek Harness, server-side search and page fetching are already handled through those APIs and are currently free. Billing is scheduled to begin at 16:00 UTC on October 31.
A pledge to filter out GEO manipulation
Another notable point is what DeepSeek says it excludes. Its changelog says it filters SEO spam, GEO manipulation, and low-quality AI-generated content. GEO refers to efforts to shape content so that a company’s material is more likely to be cited in AI-generated answers. It is, in a sense, the AI-era version of SEO, which aims to rank higher in search results.
It is symbolically significant that a search engine has explicitly said it monitors optimization for AI. Google also said on May 15 that its spam policies apply to generative AI answers in Google Search. As AI-generated answers become a new arena for visibility, efforts to manipulate them—and the push to stop that manipulation—are coming into the open.
However, the public materials do not explain exactly what GEO manipulation entails, how it is detected, or how authority ratings and relevance scores are determined. The existence of these evaluation criteria is now clear, but we will have to wait for further disclosures to learn how they work.
What site operators can see
For site operators, the practical benefit is being able to check how their site appears in DeepSeek search. A site: search can retrieve up to 50 pages from a site and show the authority and relevance scores for each. At less than one cent per search, it is also affordable to try.
On the other hand, the API does not reveal why an expected result is missing. Is it outside the index, caught by a filter, or excluded for some other reason? Neither the user guide nor the API reference describes a way to determine the cause.
The unanswered question of which crawlers are involved
Another point of interest is how crawling is handled. The Fetch API returns a single page in Markdown format, and the guide says it follows the target site’s availability and access policies. An EU-facing overview of the training data for DeepSeek-V4 says crawlers are designed to respect standard protocols such as robots.txt, but does not specify which crawlers it means.
Neither the launch guide, API reference, changelog, nor announcement names the crawler used for the search index or identifies the user agent used by Fetch. The Known Agents directory, which compiles information on crawlers, lists DeepSeekBot as being used for AI training data. The public pages do not make clear whether blocking DeepSeekBot in robots.txt would also affect inclusion in the search index.
Between convenience and opacity
The release adds another way to measure how a site appears in AI search. But without knowing what the scores are based on, it is difficult in practice to decide what to do just by looking at the numbers.
What bears watching is how DeepSeek’s documentation fills in the definition of GEO manipulation, how its levels are determined, and the name of its indexing crawler. The API’s practical usability and the transparency of its evaluation criteria may prove to be the two factors that determine whether these become established benchmarks for AI search.
Comments