News and Publisher Visibility in DeepSeek

DeepSeek handles news differently from every retrieval-first engine: by default it answers from a frozen training snapshot, so your last six months of reporting may simply not exist to it. Publisher visibility therefore splits into two separate games — being well represented in the open crawl corpus that feeds training, and being fetchable and extractable when users toggle on live web search.

Why does DeepSeek confidently describe old news as current?

Parametric models have no clock. DeepSeek-V3 (December 2024) and R1 (January 2025) answer from whatever the corpus contained at cutoff, and they will summarize a year-old development in the present tense unless the user activates search. For publishers this creates a reputation exposure: your outlet may be quoted as the source of framing that you have since corrected or superseded. The mitigation is corpus hygiene — keep corrections, updated headlines, and follow-up stories on crawlable URLs that supersede the originals, because the next training snapshot inherits whatever the open web says.

What determines which publishers the model "remembers"?

Corpus share and reuse. Wire-service syndication multiplies one story across hundreds of crawlable domains; widely aggregated outlets get their framing absorbed as the canonical account. Common Crawl (CCBot) is the dominant open pipeline into training sets, so blocking it — as many publishers did after 2023 — removes future training presence entirely. That is a legitimate commercial decision, but it should be made knowingly: no announced licensing deals exist between DeepSeek and publishers, so unlike ChatGPT (where OpenAI licenses content from partners such as Axel Springer and News Corp) there is no paid path back in.

How does web-search mode treat news content?

Which technical signals still matter without a news index?

DeepSeek runs no Google News-style vertical, but structure still pays. NewsArticle JSON-LD with headline, datePublished, dateModified, and author gives both crawlers and live fetchers unambiguous provenance — which matters when a synthesis engine decides whose account to attribute. Ledes should function as self-contained 40–80 word summaries, because that is the unit a fetch-and-synthesize pipeline actually quotes. And full article text must be present in the raw HTML response: DeepSeek's fetchers do not execute JavaScript, so client-side rendering hides your reporting even from live search.

How should publishers think about referral traffic?

Conservatively. DeepSeek's answers cite sparsely compared to Perplexity's numbered-source format, and default-mode answers cite nothing at all — the reporting is absorbed, not linked. That makes measurement, not link-chasing, the realistic first step: know how often your outlet is named or paraphrased in prompts about your beats. Running a beat-specific prompt set through AI mention tracking monthly, and watching citation patterns in web-search mode, tells you whether your coverage is functioning as source material and whether attribution accompanies it.

The strategic posture that fits DeepSeek: protect the crown jewels (hard-paywall genuine exclusives, accepting invisibility), open the authority layer (explainers, timelines, topic pages that build the model's association between your masthead and your beats), and audit quarterly which of your framings the model repeats — because in a parametric engine, yesterday's coverage is the product.