DeepSeek's new search API scores your host, not your pages

own domain returned tier 1 on every URL…

Summary

The authority tier in DeepSeek's new Search API behaves as a host score: across 16 hosts tested, every URL on a host came back with the same tier.

Baseline the tier for each host that carries your content before changing anything on a page.

DeepSeek opened its search index to the public through a Search and Fetch API on Oct. 17, 2026, and an early hands-on test of its scoring found that the authority tier attached to every result tracks the host rather than the page. Suganthan Mohanadasan, co-founder of Snippet Digital, an AI SEO agency, ran 110 calls against the API and published his results in Search Engine Journal.

The host evidence came from 16 site: probes. On each host, every URL that came back carried the same authority tier. Reddit’s subdomains show the unit is the host rather than the registrable domain: qh.reddit.com scored 1 while www.reddit.com scored 2. DeepSeek’s documentation defines the field as a tier running from 0 to 4 and describes it as a property of the page, the reading in this site’s report that the API rates pages 0 to 4 for authority. Mohanadasan’s probes do not support that reading.

A high tier is also not a verdict on the page sitting at that URL. DeepSeek says its engine filters SEO spam, generative engine manipulation and low-quality AI content, and Mohanadasan still found webform-uploaded PDFs titled “Free Robux Generator in 2025” on oig.dc.gov, the host of a US government inspector general, carrying the top tier. Uploads on business.columbia.edu pushing CBD gummies scored the same. A former MP’s domain serving “Minimum Deposit Casinos UK” ranked second for “best casino sites uk” at tier 3. Host inheritance is the simplest explanation for all three: the junk sits on a host with a high tier, so it gets that tier.

Inheritance cuts the other way for a small site that is indexed and on-topic. Mohanadasan’s own domain returned tier 1 on every URL the API surfaced, from the homepage through to blog posts. On a query matching one of his own articles, it landed at position 48 with a rerank score of 1.14, under DeepSeek’s documented threshold for high quality. Package registries and app stores took the top of those queries instead, among them PyPI, npm, Apify, the Chrome Web Store and WordPress.org. How much of that gap the tier accounts for is not visible from outside the API. The likelier lever is a listing on a host that already scores well, not a more on-topic page.

The Fetch half of the API is the part that shows up in server logs, and it does not identify itself as DeepSeek. Mohanadasan’s test fetches arrived from five IPs with a plain desktop Firefox user agent, no bot token and no From header.

SignalWhat the fetcher sent
User agentMozilla/5.0 (X11; Linux x86_64; rv:149.0) Gecko/20100101 Firefox/149.0
Accept-Languagezh-CN,zh;q=0.9
Networkfive IPs, all AS136907, Huawei Cloud, Hong Kong
Renderingloaded CSS, ran no JavaScript, loaded no images

Fetch also returned full content from three URLs disallowed for User-agent: *, among them httpbin.org/deny and google.com/search, and Mohanadasan could not confirm that robots.txt was requested at all. Fetch runs when an API caller asks for one specific URL, which puts it in the same category as ChatGPT-User and Claude-User rather than bulk training traffic like GPTBot. Robots.txt files already treat training and retrieval bots inconsistently, as a fetch test across 122 well-known sites showed, and a fetcher presenting as Firefox sits outside that control entirely.

What to do

Testing is cheap. Search and Fetch each cost $0.75 per 1,000 successful requests. Failed calls and empty result sets are not billed, and the whole test came to roughly $0.08. Server-side Search and Fetch calls made through the DeepSeek API in Claude Code or Codex stay free until Nov. 1 at 00:00 Beijing time.

  • Run one site: probe per host you publish on, including subdomains, a separate docs or help host, and third-party profiles. Record the tier against the host, since that is the number that will move together.
  • Set localization explicitly on every call. Left unset it follows the caller’s IP. Switching the calling IP between the UK, US and China changed the rerank score on up to 41 of the 50 results, though only by about 0.05 on average.
  • Run the product and tool queries you care about and note which hosts own them. If registries and marketplaces hold the top, a listing there is the faster win than another page.
  • Grep logs for Firefox/149.0 alongside the zh-CN Accept-Language and AS136907, and require all three before attributing a visit to DeepSeek. Any one of them on its own has a real browser behind it often enough to mislead.

Watch out for

Date filters removed every undated page from results, so anything without a machine-readable date drops out of a filtered query. The bigger exposure is publishing on a host you do not control: a university host, a government portal, Medium or Zenodo passes its tier to your content, and to the spam uploaded next to it. A tier 4 result is not evidence that DeepSeek checked the page.