Skip to content
Game Market Copilot Docs
Esc
↑↓navigate↵open⌘Jpreview
On this page

Methodology

How Game Market Copilot defines data coverage, sampled evidence, as-of snapshots, tag counting, and safe claims, the rules behind every number in a GMC response.

This page defines the rules that govern how to read a GMC response and what claims it supports. The gmc-analysis Skill and the guided MCP prompts apply these rules automatically; this page is the reference for anyone reading GMC output directly, human or agent.

Coverage levels

Every title carries a coverage level: basic, partial, or full.

  • basic means catalog and snapshot data only. No deeper review or marketing analysis has been collected for that title yet.
  • partial and full mean progressively deeper collection has completed.

Collection status is always labeled explicitly. A title whose deeper data has not been collected yet is reported as not collected, never as zero, never as an absence of the thing being measured. Most of the catalog sits at basic coverage at any given time; deep cross-game analysis only exists for the partial and full subset, so sizing a cohort’s coverage before running deep analysis tells you both the true denominator and the cost of proceeding.

Sampled evidence vs. market truth

Tools that read reviews or other qualitative data operate over a bounded sample, a fixed number of returned items, not every review ever written for a title. Findings from these tools are evidence-backed observations about that sample, not population measurements of “what all players think.”

Any share or percentage reported from sampled evidence must state:

  • the denominator (how many titles or items the figure is out of), and
  • the observation window (the time range or snapshot the sample was drawn from).

Aggregate tools that report basis: server_aggregate are population-true for their filter, not page-bounded samples, and can be read as counts over the whole matching set. Tools that return sampled evidence should never be conflated with these population-true aggregates.

As-of semantics

Analyses carry an as-of date. GMC data is snapshot-based: a response reflects the state of the catalog as collected as of that date, not a live read of Steam at request time. When comparing figures across time, note that historical price, review, and rating metrics can be read as-of a past date, but metadata fields such as genre and tags always reflect the current, latest values; a historical query does not roll metadata back with it.

Review analysis

Themes, claims, and quotes surfaced from review analysis are drawn from observed reviews and link back to their evidence rather than being asserted as unattributed conclusions. Per-title theme labels are not guaranteed to match verbatim across different titles. Clustering similar labels into a shared, cross-title theme is an interpretive step for the reader (or agent) to perform and label as their own synthesis, not something the API does for you.

AI-derived data and verification

Some of what GMC reports is measured directly; some of it is generated by a language model and needs the same scrutiny you would apply to any analyst’s draft. Counts, prices, dates, and other catalog and snapshot fields are measured: they come from collected data, not from a model’s judgment. Review themes, claims, quotes, and summaries are LLM-extracted from a sampled set of reviews. Copilot’s chat answers, and the cohort clustering it performs at query time, are generated when you ask, not stored facts retrieved from a database.

Generated output can be wrong. A model can produce a plausible-sounding claim that is not actually grounded in the reviews or snapshots it was given, a hallucination, and because review analysis runs over a sample rather than every review, coverage of any given theme is partial even when the model gets it right. Treat an AI-derived finding as a starting point for your own check, not as a settled fact.

GMC narrows this risk in a few ways. Claims from review analysis link back to the observed evidence behind them rather than standing as unattributed assertions (see Review analysis above). API and MCP responses carry machine-readable warnings, coverage, and basis metadata so a caller can tell what a given response does and doesn’t cover. not_collected is never reported as zero, so missing data cannot be misread as a measured absence.

None of this replaces your own judgment. Verify AI-derived findings, whether from review analysis or from a Copilot answer, against the linked evidence or the underlying source data before you rely on them or publish them elsewhere. This matches the “Republishing for Press, Analysis, and Research” section of the Terms: republished Derived Outputs must attribute Game Market Copilot as the source and must not present data that was not collected as if it were zero or absent.

Tag counting bases

Titles can be counted against tags in two different ways, and they diverge sharply:

  • Primary tag basis: each title is counted once, under its single primary market tag. Untagged titles are omitted. Counts under this basis sum to the tagged subset of the cohort.
  • Multi-assignment basis: a title is counted under every tag it carries. A title with five tags contributes to five tag counts. Shares under this basis can overlap and sum to more than 100% of the cohort.

Always name which basis a figure uses when citing a tag share, and never mix or directly compare numbers computed under the two different bases.

Safe claims

  • not_collected means the data has not been collected. It is never reported as zero events or as an absence of sentiment.
  • Marketing and review-derived signals are correlational. Associations are described as such (“is associated with,” “co-occurs with”), never with causal language (“drives,” “causes,” “leads to”).
  • Any figure not computed from GMC data, such as unit-per-review multipliers, wishlist folklore, or outside revenue estimates, must be labeled as an external estimate with its source, and never blended into GMC-derived numbers as if it were one.
  • Locked or preview data stays partial. What a locked or preview payload contains is presented as a preview only; nothing is inferred about content beyond what was actually returned.

Was this page helpful?