CMC AI Question Keys: The Coin and Market Questions Answered

How to Read a Key

Two identifiers do different jobs, and mixing them up is the common integration error.

Coin-Level Questions

The last column is CoinMarketCap’s published figure for how many of the 100 covered assets had content for that key, given here so the table is usable on its own. It is an undated snapshot, and the two price keys sit far below the rest for reasons specific to how they fire. Checking coverage before you call carries the dated measurement, the per-key analysis and the reason for that gap.

* Six of the eight strings above are taken from the example payloads in the coin insights reference. latest_news and sentiment appear in no reference example; those two are taken from the launch page template and confirmed on the rendered pages. These strings are not what the rendered pages display, and the next section is the measurement of how far apart they are.

All eight are published per coin across four page URLs. The latest-updates page carries four questions on its own: latest_news, sentiment, codebase and roadmap. The other page types map one to one: what-is for overview, price-analysis for the up and down pair, and price-prediction for future_price.

The Title Is Not Stable Across Surfaces

CoinMarketCap publishes these question strings on three of its own surfaces, and the three do not agree. Measured 14 September 2026 across 20 assets and all four /cmc-ai/ page types, 80 page fetches:

First-party measurement, 14 September 2026. Reference column from the coin insights API reference; rendered column from /cmc-ai/{slug}/{type}/ across 20 assets.

Four differences matter to anyone matching on text.

1. The substitution rule is not uniform.

   The launch page writes every question with an X placeholder. Seven of the eight resolve X to the ticker. overview resolves it to the coin name in the reference example and to name plus ticker on the page, so overview is the one key where a ticker substitution produces a string that appears nowhere. Measured on all 20 assets, the rendered overview heading is “What is {name} ({ticker})?” without exception.

2. future_price does not render as a question at all.

   All 20 price-prediction pages return HTTP 200 with an H1 of “{name} ({ticker}) Price Prediction” and no question heading anywhere in the document. The documented string “What could affect BTC’s future price?” appears on none of them. The URL still carries the future_price content, so the mapping in the previous section holds; what does not hold is any assumption that the page shows the question.

3. The apostrophe differs by surface.

   Four of the eight strings contain an apostrophe. The reference uses ASCII ' (U+0027) and the rendered pages use the typographic ’ (U+2019), as the table above reproduces. A literal comparison between the two surfaces fails on every one of those four, and it fails silently, because both strings look identical in a terminal and in most diffs.

4. Three of 20 assets render a duplicated token.

   Where the coin name equals its ticker, the name-plus-ticker format produces “What is XRP (XRP)?”, “What is BNB (BNB)?” and “What is USDC (USDC)?”. That is cosmetic rather than breaking, but it is the kind of string a product will want to rewrite rather than pass through.

Market-Level Questions

Six keys are market-level, and all of them regenerate every 30 minutes.

All six rendered on the CoinMarketCap homepage on 10 September 2026, alongside a seventh question, “Why is the market down today?”, which is not one of the six keys. The market-wide up and down questions render on the homepage but have no entry in the published question_key enum, so they arrive as an unkeyed insight and have to be matched on title rather than on a key. Treat that as a display fact rather than an API contract.

Filtering: Keys Work on Coins, Not on the Market

The two insight endpoints do not take the same parameters, and the difference decides where filtering has to happen.

Both endpoints accept type with the same three values, which is the practical lever on the market feed: type=fixed_question drops the trending questions and the headlines, which is most of the payload. A typical market response carries 26 items, of which 6 are fixed questions, 3 are trending and 17 are headlines.

The Shape of an Answer

Every answer arrives pre-split into answer.tldr and answer.body, both markdown. The split is applied server side using the same rule the CoinMarketCap web front end applies, and nothing else is rewritten, so what you receive is the generated text with a section boundary marked rather than an edited version of it.

The reference documents the two halves and names the body’s sections as Deep Dive and Conclusion. It does not specify what goes inside them, so the structure below is measured from the rendered price-analysis pages of 12 assets on 14 September 2026, not read from the schema. It describes generated copy, which can change without a version bump.

Measured 14 September 2026 across the price-analysis pages of bitcoin, ethereum, xrp, solana, bnb, dogecoin, cardano, tron, chainlink, avalanche, hedera and pepe.

Sources appear inline in the body as parenthetical attributions next to the claim they support, and the sources array carries the URLs. A source object has exactly one field, url, which is therefore also its identity for deduplication. Two further fields describe what you did not receive: one gives the number of sources the answer actually drew on, the other is a boolean saying whether the array was cut to fit sources_limit. Both are defined on the launch announcement, and how every answer is attributed covers what the source mix looks like in practice.

Requesting One Question

# One question, one coin
GET /v5/cmc-ai/coins/latest?crypto_id=1&question_key=price_up

# Two questions across three coins, sources trimmed to a count only
GET /v5/cmc-ai/coins/latest?slug=bitcoin,ethereum,solana&question_key=overview,latest_news&sources_limit=0

# Market feed, fixed questions only, keys filtered client side
GET /v5/cmc-ai/latest?type=fixed_question

Identifiers are mutually exclusive: crypto_id, slug or symbol, one type per request, comma-separated. sources_limit defaults to 10 and caps at 100, and a limit of zero still returns sources_count, so a citation count can be shown without paying for the array. Pagination uses start and limit, which default to 100 and cap at 250. The unit is the cryptocurrency, not the insight, so a page of 100 can carry several hundred insights and limit does not bound the response size in the way it usually would.

Three behaviours are worth knowing before writing the polling loop.

Answers are read from storage, not produced when you ask for them, so a request that arrives inside the regeneration window receives the copy that is already there. A coin whose content has not changed since your last call is not billed for again, so the cost of a conservative schedule is closer to the number of regenerations than to the number of requests.

And there is a second staleness term on top of the cadence. The reference gives these endpoints a cache window of up to one minute, so a response can be a minute behind the stored answer even when a regeneration has just landed. It is immaterial against a 1-hour key, and worth knowing if you are polling the 30-minute market feed on a tight schedule and expecting to see a change the moment it is written.

Which Keys Have Content Right Now

A key being defined does not mean a given coin has an answer for it today, and the map endpoint reports that per key rather than per coin.

/v5/cmc-ai/coins/map returns available_question_keys, an array drawn from the same eight coin-level keys naming what currently has content for that asset. That array is the reason to call the map before the insight endpoint: it is the only place the key enum is reported as a live per-asset fact rather than as a schema. Alongside it, num_insights counts insights across all types, and last_generated_at is null where nothing has been generated.

The behaviour behind an absent key, what the empty responses look like, what they cost, and the measured fill rates by key are all covered in checking coverage before you call, which is the page to build the empty state from.

Availability

Two facts govern whether you can call these endpoints at all today. The family is gated to a single plan tier for now, and a later phase moves it onto its own AI credit allowance and opens it to the rest. The launch announcement is the canonical statement of both, and is the page to check rather than this one, because the phase boundary will move and a reference page that restates it will go stale.

One availability fact does belong here, because it shapes what you build rather than what you buy: answers are generated in English only, so any localised surface is translating them.

FAQ