Build a keyword set

Use the Keyword Tool to move from an initial topic idea to a search query that can be defended and reproduced.

Run the Keyword Tool

1. Enter the initial topic terms.
2. Review the suggested related terms.
3. Accept the suggestions that capture the intended concept.
4. Reject the suggestions that would create false positives or ambiguous matches.
5. For a term whose relevance is unclear, open its example sentences and check how the term is used in real earnings-call language.
6. Add terms the tool did not suggest, such as plural forms or exact phrases.
7. Save the search protocol so accepted terms, rejected terms, and review notes can be audited later.

The accepted terms become the query for the Risk Tool search, joined with OR unless the construct requires stricter query syntax.

Rejecting a suggestion removes it from the accepted list and stops it from being shown again. It does not train the suggestion model.

Start narrow

Begin with the clearest terms for the construct. Add synonyms, phrases, and plural forms after checking whether they match the intended concept in earnings-call language.

For example, a first pass might separate general terms from phrase variants:

inflation OR pricing OR margin pressure

A bare multi-word term is already treated as a phrase; quotes are only needed for reserved characters and make matching case-sensitive (see query syntax).

Validate suggestions

Suggested related terms still need review. A keyword set is not validated just because it returns many matches.

For a suggested term whose relevance is unclear, use the example-sentence review: the tool shows randomly selected corpus sentences containing the term. Read them and track whether each sentence uses the term in the intended sense. If most examples use the term correctly for the construct, the term is a better candidate for the final query.

Use settings deliberately

Keyword Tool settings change which suggestions you see:

• Length filters suggestions by phrase length: one-word, two-word, longer, or mixed-length phrases.
• Number of keywords controls how many suggestions are returned, up to 100 per request.
• Model selects which earnings-call period the term suggestions are learned from (for example 2012–2022, 2020–2022, or 2023). A recent-period model surfaces newer vocabulary; a long-period model favors stable vocabulary.
• Capitalization is respected by default and can be turned off in the tool settings.
• Nested-phrase suppression hides suggestions that contain a phrase already in your set, and is on by default; turn it off to see longer variants of accepted phrases.

Record non-default settings when they affect the final keyword set.

Balance precision and recall

Precision means the matched sentences usually describe the intended concept. Recall means the query captures more of the relevant language.

To improve precision, use exact phrases or exclusions:

silicon AND NOT valley

To improve recall, add variants manually:

company OR companies

Record the query

Save the exact query string outside the product. Also record the date range, selected options, and any later refinements — the reproducibility checklist lists everything worth keeping.

Save the Keyword Tool search protocol with the final query definition so accepted terms, rejected terms, and example-sentence review can be audited later.