ObsurfableGenerative engine optimization was born because AI answers select sources humans never click. On October 17, 2026, DeepSeek added a twist: its changelog opening the Search and Fetch APIs says the in-house engine filters SEO spam, GEO manipulation, and low-quality AI-generated content — without defining manipulation or publishing filter benchmarks.
For AEO teams, the release matters for two reasons. First, another major stack now exposes pre-answer retrieval with scored pages. Second, a search provider explicitly named GEO as a spam class — even if marketers cannot yet see why a page was dropped.
What DeepSeek shipped
DeepSeek documents that the same engine powers search on its site and through the API. Developers can:
- Search API: ranked web results (up to 50 per query), including
site:probes. - Fetch API: retrieve a page as Markdown for downstream apps.
Published pricing (per DeepSeek docs): $0.75 per 1,000 successful Search or Fetch requests.
The changelog’s GEO line is marketing until DeepSeek publishes definitions and false-positive rates. The inspectable part is the response schema.
Scores you can read today
| Field | Documented meaning |
|---|---|
authority | Tier 0–4 (−1 = unknown). ≥3 = highly authoritative sources. |
rerank_score | 0–3 blend of relevance, authority, freshness. ≥1.5 = high quality. |
There is no field for “filtered as GEO spam.” If a page is absent, the API does not explain whether ranking, policy, or crawl failure removed it.
That gap matters for compliance-minded brands: you can audit what ranks, not what was suppressed.
What early independent testing found
Search Engine Journal’s October 2026 walkthrough (Suganthan Mohanadasan) ran ~110 API calls on launch day — 93 Search, 17 Fetch — costing about $0.08.
Directional findings (preliminary, single-day sample):
Rerank scores look bucketed and sort cleanly
Across 20 SEO/GEO queries at count=50 (1,000 results):
rerank_scorerange 0.75–2.70, median 1.63- 68% scored ≥1.5
- Results sorted by
rerank_score; only 196 distinct score values across 1,000 rows — suggesting rounding or discrete buckets
Authority tiers clustered at 2–3; host-level ties
Across 4,322 results:
| Authority tier | Share | Median rerank |
|---|---|---|
| 1 | 2% | 0.95 |
| 2 | 32% | 1.45 |
| 3 | 60% | 1.83 |
| 4 | 6% | 1.98 |
site: probes on 16 hosts: every URL from a host shared the same authority tier — e.g. independent consultant sites at 1, major SEO publishers at 3, some academic/government hosts at 4.
That pattern looks more like host reputation classes than page-level judgment — similar in spirit to domain factors dominating AI citation models, though DeepSeek’s tiers are proprietary.
Fetch behavior raised crawl-policy questions
The tester reported Fetch presenting as Firefox from Hong Kong and returning content from URLs disallowed in robots.txt in at least one case — a reminder that API fetch paths may not mirror public crawler etiquette. Validate your robots.txt, CDN bot rules, and terms before assuming disallow blocks all retrieval — see Training Crawlers vs Search Crawlers.
“Spam filtering” vs observed long-tail PDFs
The launch promises filtering; the test also saw spam PDFs on high-trust hosts scoring authority 4. Policy claims and live results can diverge until DeepSeek publishes enforcement metrics.
What “GEO manipulation” might mean in practice
DeepSeek did not define the term. Reasonable hypotheses — not confirmed by DeepSeek:
| Tactic class | Why a search vendor might flag it |
|---|---|
| Mass-produced “answer pages” with thin entity swaps | Resembles programmatic SEO spam |
| Hidden text / schema stuffing aimed at rerankers | Classic spam signals |
| Parasite SEO on unrelated high-authority hosts | Aligns with Google’s October 2026 spam update themes (mass programmatic content) |
| Adversarial “stat injection” without editorial substance | Matches low-quality AI content filters |
Until DeepSeek documents features, treat the phrase as risk language, not a checklist of banned HTML patterns. The GEO Evidence Review 2026 still shows mixed evidence for specific rewrite tactics — manipulation filters, if real, likely target abuse, not clear answer-first pages.
Implications for AEO strategy
1. Retrieval APIs are becoming inspectable
You can run site:yourdomain.com queries, read authority / rerank_score, and compare to competitors before an LLM writes an answer. That complements answer-level monitoring in Obsurfable Explorer.
2. Host authority may gate page-level GEO
If authority is host-stable, mid-market sites on shared hosting or young domains may need off-site footprint before page tweaks move retrieval — consistent with Indexably-style domain dominance.
3. Do not optimize for a black-box “anti-GEO” score
Without transparency, the safe path remains accurate, extractable answers, honest schema, and diverse third-party validation — not keyword-stuffed “AI paragraphs.”
4. Separate Search API retrieval from ChatGPT / Google behavior
DeepSeek’s filter claims apply to its index. ChatGPT, Gemini, and Google AI Mode each maintain different retrieval graphs — see Why Each AI Engine Reads a Different Search Index and Google’s October inline-link push in Google AI Search Inline Links.
How Obsurfable fits
API scores describe candidate sources. Brand outcomes live in answers: who is named, who is cited, whether prose matches your facts. The App tracks mention and citation movement on fixed buyer prompts; Explorer keeps the underlying observations public.
FAQ
Will DeepSeek’s filter affect my Google or ChatGPT visibility?
Not directly. Each stack has its own index and policies.
Can I appeal if DeepSeek filters my pages?
No public appeals process was announced with the API launch.
Should I block DeepSeek crawlers?
Treat it like any AI crawler decision: weigh discovery vs training vs terms. Document choices in robots.txt and CDN settings.
Is high rerank_score enough to win brand mentions?
No. Retrieval eligibility ≠ recommendation. Track naming separately — Sentic October 2026.
Bottom line
DeepSeek’s October 2026 Search API makes scored retrieval visible and puts GEO manipulation on the same slide as SEO spam — without exposing filter logic. Marketers can probe authority and rerank today; they should still measure mentions and citations in real answers, and avoid abusive tactics that future filters are designed to catch.