DeepSeek has opened a public Search and Fetch API, giving SEOs a direct look at how an AI search engine ranks and scores websites. The pitch is strong: every result carries an authority score, and the system claims to filter SEO spam, GEO manipulation and low-quality AI content. A small independent test suggests the scores do not always match that promise.
For teams optimizing for ChatGPT, Perplexity or Google AI Overviews, this is a rare chance to inspect retrieval directly instead of guessing from outputs.
The scores are easy to inspect
One tester ran 110 calls from Dubai on Oct 17, 2026 and spent roughly $0.08. Across 20 SEO and GEO queries at 50 results per query, the 1,000 URLs returned rerank scores from 0.75 to 2.70, with a 1.63 median. About 68% of results scored at least 1.5, the documented high-quality threshold.
Search and Fetch cost $0.75 per 1,000 successful requests, with failed requests, empty results and timeouts not billed. A temporary exemption for server-side calls made through Claude Code or Codex runs until 1 November at 00:00 Beijing time.
Authority looked like a host score, not a page score
Across 4,322 results, the lowest authority tiers never appeared. In site: probes across 16 hosts, every URL from the same host carried the same authority tier. Even Reddit split by subdomain: qh.reddit.com scored 1 while www.reddit.com scored 2.
That host-level behavior had a clear consequence. Spam PDFs on government and university hosts received authority 4, while the tester’s own site, suganthan.com, received authority 1. Tool queries were often won by PyPI, npm, the Chrome Web Store, WordPress.org and GitHub, while original content about the tools sat below the 1.5 rerank threshold.
The spam filter let questionable pages through
DeepSeek says it filters SEO spam, GEO manipulation and low-quality AI content. Yet spam PDFs on trusted domains passed through, and self-published “frameworks” on zenodo.org received authority 4 and ranked first and third for the term “generative engine optimization”.
The pattern matters because authority appears to attach to the host. A trusted domain can carry parasite pages that inherit the same high score, which makes the authority tier a weak signal of individual page quality.
What to audit now
- Query your actual target keywords against the DeepSeek Search API and record rank, rerank score and authority.
- Run a site:yourdomain.com search to check the authority tier for your root domain and subdomains.
- Set localization explicitly so tests reflect the market you care about, not the caller’s IP.
- Add machine-readable dates to pages that need to appear in date-filtered searches; undated pages disappeared when filters were applied.
- Watch for a Fetch fingerprint: Firefox/149.0, Accept-Language zh-CN and Huawei Cloud AS136907 from Hong Kong.
The Fetch API also presented itself as Firefox, loaded CSS but not JavaScript or images, and returned content from URLs disallowed by robots.txt, including Wikipedia’s Special:Search and google.com/search. Do not block the Firefox string or the whole ASN blindly; both can catch legitimate traffic from other tenants.
For SEOs and AI search optimizers, this is a useful moment: the weighting is visible enough to test. Start with your own authority tier, then compare it against the domains actually winning your queries.
Source: Search Engine Journal
