AI Crypto Data Sources: How Every CMC AI Answer Is Attributed

What a Source Is

The source object is deliberately thin. Its one required field is the canonical URL of the page a claim came from, and because that URL is the only field, it is also the identity used to deduplicate. Two entries pointing at the same article, arriving with different query strings, are the same source and collapse into one.

That puts weight on URL normalisation, which the documentation describes as forcing https, stripping tracking and rendering parameters, stripping www., normalising twitter.com to x.com, and removing fragments before the array is deduplicated. Across the 206 published source links measured, every one used https and none carried a query string, so scheme forcing and parameter stripping both hold on the rendered surface.

The Measured Source Mix

Measured 9 September 2026 across 206 linked sources on 62 coin pages, all of one question type.

Method: same-origin fetch of the latest-updates page for 64 coin slugs, parsed with DOMParser, counting every anchor the page labels “View original post” and classifying by hostname. 62 pages carried linked sources, 2 returned 404.

Two things follow for anyone building on this. There is no parameter for filtering sources by kind, so a product that does not want to surface social posts has to filter client-side on the hostname, and it will be discarding most of the citations rather than a fraction of them. And a product that surfaces citations to a professional audience should expect handles rather than mastheads, and design the credit line accordingly.

The concentration matters as much as the mix. Thirteen hostnames across 206 links means the same small set of publishers and accounts recurs, so a citation display will show repeat names to any user reading more than a few answers.

What the Published Answers Show

A discussion answer renders considerably more than a bare link. Four sources behind one Bitcoin answer, read on 8 September 2026:

Each carries a short verbatim quotation from the source, an interpretation of what it means for the asset, and a link labelled “View original post” that resolves to the original page.

Linked Sources Sit on the Discussion Answer

Attribution is not uniform across question types, and this is the practical thing to know before building a citation UI.

So a claim in a discussion answer can be checked in two clicks. A claim in a price answer names its source but leaves the reader to find it. Build the citation surface around the answer type rather than assuming one pattern.

Filtering Sources Is a Client-Side Job

The API returns every source in one array with no kind attached, so classifying them is work the caller does. The four groupings used on this page are a description of what turns up, not values the endpoint accepts.

A product that needs to exclude one of these matches on the hostname after the response arrives. Given the measured mix, excluding social removes roughly five in six of the citations behind a discussion answer, so plan the empty state as carefully as the populated one.

How Many Sources Sit Behind an Answer

Answers draw on far more sources than any interface shows, which is why the array is capped and the true total is reported separately.

The gap is wide: published pages average 3.32 linked sources while a single answer’s sources_count can run into the dozens. Together those fields make an honest citation display possible: show what came back, and state the count that did not, rather than implying the visible handful is the whole basis.

# Count only, no array transferred
GET /v5/cmc-ai/coins/latest?slug=bitcoin&sources_limit=0

# Raise the cap to see more of what an answer drew on
GET /v5/cmc-ai/coins/latest?slug=bitcoin&sources_limit=100

Parameter names and defaults verified against the shipped Pro API reference. The family is Enterprise-gated in Phase 1, so these are not callable on a lower plan.

A citation line built from those fields, rather than from the array length:

def citation_line(insight):
    """Build the line from the fields, never from len(sources)."""
    shown = len(insight["sources"])
    total = insight["sources_count"]

    if not insight.get("sources_truncated"):
        return f"{total} source{'s' if total != 1 else ''}"

    return f"Showing {shown} of {total} sources"

Where an Answer Says It Does Not Know

Generated answers disclose their own gaps rather than filling them. On the live codebase answer for Bitcoin, the text states that specific patch notes for one release “are not detailed in the provided sources” and reasons from what such a release generally consolidates. On the news answer, a corporate purchase is described as reported, with the note that transaction details “remain unconfirmed by an official filing”.

For anyone assessing whether to republish an answer, that hedging marks the boundary of what the sources support.

Checking an Answer Yourself

The published pages are the reference implementation. Open the answer, read the claim, follow the source link, compare. Four questions render per covered coin on the latest updates page, so news, discussion, codebase and roadmap answers can be checked side by side, and price analysis covers the price-move questions.

FAQ