Agent Policy for News Sites and Blogs
Published November 14, 2026 · Reviewed November 2026 · By the Ingenious AI team
The short answer
The publisher's policy balances two truths: your articles are written to be read, and your archive is an asset that bulk copying and model training can strip. The published blog sample encodes the balance — an allow default, attribution expected, one named agent refused entirely. This guide extends that pattern for news sites and blogs of every size: the reading welcome, the crawl-full-site line, the training clause, and the syndication conversation.
What is the publisher's pattern?
Open-with-attribution. The default welcomes agents; the reading verbs — browse, read, search, summarize, quote — are granted broadly, because an article an agent can summarise and cite is an article with a new distribution channel. The fences sit where publishing's value leaks: crawl-full-site, the verb that mirrors the archive; download at scale; and the after-life, governed by data_use_terms — attribution required, republication refused, training addressed head-on. The sample blog states the refusal side through one named agent switched off for bulk behaviour; the fuller pattern below distributes the fences across all rules.
Reading is distribution; mirroring is not
The vocabulary hands publishers the distinction their industry has wanted for a decade. read and quote are per-article, question-shaped verbs — an agent answering a reader, citing you. crawl-full-site is the archive-shaped verb — systematic copying of everything, whose product is a corpus, not a citation. Permit the first pair and disallow the third, and your policy states the industry's settled intuition in machine-readable form. The middle verb, download, deserves its own thought: single-article downloads for a reader's convenience are reading-shaped; at pace, they are mirroring with extra steps — which is what your rate_limit is for. Set reader pace generously for questions and let bulk ambitions meet the contact term instead.
The training clause and the attribution price
Two sentences in data_use_terms do most of a publisher policy's economic work. The training sentence states whether article text (and images) may be used to train models — refuse it, permit it, or make it a contact conversation, but state it, because silence is the one position nobody can honour or breach. The attribution sentence prices the welcome in the currency publishers actually collect: credit and the link back, as the sample's terms require. Together they convert the policy from a technicality into the publisher's public licensing posture for the agent era — the document licensing conversations start from, as comparison terms do for stores. Write both sentences in the plain register of the data-use guide: summarisable without distortion.
Syndication, archives, and the named exceptions
Publishers accumulate exceptions the way they accumulate syndication deals — one argued relationship at a time. Wire services, research partners, accessibility tools, and archive licensees each deserve a named rule reflecting their actual arrangement: wider verbs for the licensee (download, a higher pace, retention permitted), tighter ones where the deal is narrower. The blog sample's single no rule is the same mechanism pointed the other way — a named refusal for a named behaviour pattern. Both directions share a discipline: exceptions live in the file, dated, where a successor editor can find them, rather than in email threads the next editor will never see. The review cadence is where the exception list gets re-argued, quarterly, in the open.
Publishers should expect their policy to be read by one more audience than they planned for: their own journalists and editors, the first time an agent-mediated story breaks — a summary that misattributes, a quote that travels without its link. In that newsroom moment the file becomes the reference document everyone reaches for, and its clarity under that pressure is a gift the publisher gives its future self. Write the attribution and training sentences as if a colleague will read them aloud in a meeting about your own content, because that is a realistic fate for them. The policy that survives being quoted in its owner's own newsroom is the policy that was written as a position, not as a settings file.
Worked example: a magazine's attribution-first policy, year one
Fictional example. The Saltmarsh Review, a fictional monthly magazine with 3,900 articles online, publishes: allow default; catch-all granting the five reading verbs free at 30 requests per minute; crawl-full-site and download disallowed for all; data-use terms requiring attribution with a link, refusing republication beyond quotation, and refusing model training on article text and photography. Named rules: 4 — a research partner permitted download at 60 per minute under its licence; a wire service at conditional pending its renewal; 2 agents set to no for observed mirroring. Outcomes in year one: agent-cited referrals became a visible analytics segment; the training clause was quoted back in 3 licensing enquiries that began as policy reads; mirroring continued from the 2 refused agents — handled, as the format intends, by the magazine's host acting on a clean published record.
The publisher policy checklist
- Welcome the reading verbs.
browse,read,search,summarize,quote— granted broadly, free, at a reader's pace. - Refuse the archive verbs.
crawl-full-sitedisallowed by default;downloadpaced or named-rule-only. - Write the two sentences. Training position and attribution requirement, in data-use terms, on every rule.
- Name your exceptions. Licensees wider, abusers refused — each as a dated rule a successor can find.
- Review with the licensing desk. The policy is your public licensing posture; its review belongs beside syndication, not under IT alone.
Keep reading
Agent Policy for SaaS and API Products · Agent Policy for Travel and Booking Sites · What Is an Agent Access Policy (v0.1)? · Spec samples · Conformance · Pricing · All guides
Frequently asked questions
Does allowing summarize mean agents can republish my articles?
No. Summarize and quote are reading verbs with their own scale; republication is an after-life question governed by data_use_terms, where the publisher pattern refuses republication beyond quotation. An agent whose summary becomes a substitute for the article has crossed the stated terms — which is why stating them, precisely, is the policy's job.
How do I stop training on my archive without blocking readers?
Separate the verbs from the after-life. Reading stays permitted; the training refusal lives in data_use_terms, which bind what may be done with what was read. crawl-full-site — the verb training pipelines actually need at scale — is disallowed outright. The combination welcomes the channel and refuses the corpus.
What pace suits a news site?
News is freshness-shaped: a working newsroom's output turns over daily, so reader agents re-read sections rather than archives. Thirty requests per minute serves question-shaped traffic comfortably on most publishing stacks; breaking-news peaks argue for reviewing the cap alongside your CDN posture rather than in isolation.
Should paywalled articles be treated differently?
The policy governs agent operation, not your paywall mechanics — subscription walls remain your platform's business. What the policy can do is keep its promises consistent with the wall: do not grant verbs against content agents cannot legitimately reach, and let the notice point subscription questions to the place humans answer them.
Can a policy help with attribution in practice?
It makes attribution a stated term rather than a hope. Agents built to honour declared terms attribute because the file asked, in the field designed for it. When attribution is missing, the published requirement is the record you act on — with the operator, the platform, or the host — instead of an expectation you never wrote down.
Generate a policy — free · Validate a policy · What an agent access policy is