Compare measurements without mistaking noise for change

Use comparison tools when the question is whether a result changed. Two percentages, or two separate confidence intervals, are not a test of the difference.

Choose the comparison

Question	Tool	Important limit
How does a metric move over several runs?	`get_metric_timeseries`	A trend requires at least two completed runs.
What are the largest increases or decreases?	`list_timeseries_movers`	A ranked movement list is descriptive.
Has the citation landscape shifted?	`screen_drift`	Cross-run change, not within-run convergence.
Which sources changed?	`attribute_drift`	Inspect each returned contrast and its scope.
Did a domain’s citation share change between two runs?	`get_paired_contrast`	Domain-level citation `share` or `count` only in this edition.
What happened to ranks?	`get_rank_displacement`	Rank movement and share movement are different measures.
Which changes deserve investigation?	`get_significant_findings`	An exploratory screen; inspect correction metadata.

Make an explicit paired comparison

{
  "name": "get_paired_contrast",
  "arguments": {
    "project_id": "PROJECT_ID",
    "run_id": "NEWER_RUN_ID",
    "run_id_a": "OLDER_RUN_ID",
    "run_id_b": "NEWER_RUN_ID",
    "entity_or_domain": "colorescience.com",
    "score_family": "citation",
    "metric": "share"
  }
}

Set run_id to run_id_b. Both runs must belong to the project. Use a bare domain, not a page URL. For a within-run comparison, use condition_a and condition_b, such as two platform conditions, instead of the two run-ID arguments. Read data.contrast.value_a, value_b, diff, ci_diff, and significant. The difference is B minus A. For share, a difference of 0.02 is 2 percentage points, not 2% relative growth. significant: false does not prove that nothing changed. It means the comparison did not establish the claimed direction at the returned confidence level. Null contrast fields mean that the comparison could not be estimated; do not describe them as “no difference.”

Select a reference run for drift tools

Omit reference for the previous-run default, or pass {"type":"first_run"} or {"type":"specific_run","run_id":"REFERENCE_RUN_ID"}. A bare string such as previous_run is not the JSON object format. Use a native object or its JSON-encoded string.

Be careful with broad screening

get_significant_findings reports correction metadata. In this edition it does not correct for the full set of comparisons. A long list can contain chance findings. Treat it as a shortlist to investigate, retain family_size, and do not call every candidate a confirmed business change. A confidence interval is required for a defensible estimate of change when the tool supplies one. Statistical evidence still needs business context: a detectable difference can be too small to matter, and an important question can remain uncertain. Next: Worked example.

Get started

Connect securely

Read measurements

Build dependable integrations

Build reports

Report recipes

Reference

Session and discovery

Visibility measurements

Sources and evidence

Changes and trends

Google data

Pages and report planning

HTTP operations

Compare measurements without mistaking noise for change

Choose the comparison

Make an explicit paired comparison

Select a reference run for drift tools

Be careful with broad screening

​Choose the comparison

​Make an explicit paired comparison

​Select a reference run for drift tools

​Be careful with broad screening

Choose the comparison

Make an explicit paired comparison

Select a reference run for drift tools

Be careful with broad screening