SEO Recap covering October 19, 2026: DeepSeek's search API and AI citation manipulation studies

Built by Stephanie Chung·4 min read·3 stories

Daily summary of what matters in SEO, GEO, AEO, and AI search generated with Claude Code (beware of hallucinations)

Generative Engine Optimization (GEO) & Answer Engine Optimization (AEO)

DeepSeek launched its Search and Fetch APIs on Oct. 17, saying the engine indexes hundreds of billions of pages, handles hundreds of millions of daily requests, and filters SEO spam, "GEO manipulation," and low-quality AI-generated content, though it never defines GEO manipulation or how detection works.
Each Search API result returns scored fields practitioners can read:
- An authority tier from 0 to 4 (-1 unknown), with 3+ labeled "highly authoritative"
- A rerank score from 0 to 3 blending relevance, authority, and freshness, with 1.5+ considered high quality
- A 100-200 character display snippet and a 600-800 character model-facing summary
The API accepts site: queries limited to one domain and filters by language and date. Search and Fetch each cost $0.75 per 1,000 successful requests, so a single site: search checking up to 50 of your pages runs under $0.01.
The launch pages never name a crawler for the search index or identify Fetch's user agent, leaving unclear whether a robots.txt rule blocking DeepSeekBot (listed as a training-data crawler) affects search inclusion. Billing for server-side calls starts Oct. 31 at 9:00 am PT.

Fudan University researchers published posts about a fictional "Cactus Gravity Protocol" on 9 sites the tools already cite often, and all 4 English-web tools (ChatGPT, Perplexity, Grok, Google AI) cited it within 7 days. 8 of 10 tools tested did.
Citation rates varied sharply by tool. Grok linked to a post in 30 of 32 answers, Google AI in 15 of 33, Perplexity in 13 of 35, and ChatGPT in 7 of 20. Planted details were repeated far less often (Grok 53%, ChatGPT 10%, Google AI 6%, Perplexity none).
Placement beat volume. The post on a frequently cited source always produced the first citation, while 40 posts on a self-hosted WordPress site drew zero references in 7 days, and a $14 GEO vendor package delivered 13 posts in ~28 minutes that drew a Doubao citation within an hour.
None of the fictional topics was a real brand or product, so the pattern may not hold for brand searches. An opinion topic and a made-up product mostly failed to get picked up in the same window.
The tests ran in March and April, before Google clarified on May 15 that its spam policies cover attempts to manipulate generative AI responses in Search.

A model can memorize a biography word for word yet still fail to answer where the person works, a gap documented in Physics of Language Models Part 3.1, so finishing a product description is a different test than answering a buyer's question about it.
Varying wording, shuffling sentence order, and repeating full names raised factual-question accuracy from 9.7% to 96.2% in a 124-million-parameter GPT-2 model, but these were synthetic facts in a small-model experiment, not commercial brand recommendations.
Transfer effects complicate content-driven recall claims. Accuracy on an unchanged group's questions rose from 4.4% to 86.8% purely from training on varied material about other people, meaning a model update could improve brand recall without anyone rewriting the brand's pages.
Retrieval outperformed memorization in Fine-Tuning or Retrieval, where Mistral-7B scored 50.4% on original docs, 58.8% on paraphrased material, and 87.5% when given relevant passages at answer time.
Publishers control the reading material but providers control the training curriculum, so a publishing recipe that promises better recall still has to prove the work caused the result rather than borrowing a paper's unrelated finding.

Published on October 20, 2026