Query design guide
Write semantic searches that stay useful
Turn a research question into a semantic content search that catches related language without becoming a broad, noisy feed.
Test the example queryA useful semantic content search describes the subject, audience, problem, and evidence you want. It does not need every possible keyword. The matching system can look for related meaning, but it still needs a clear boundary.
For example, “onboarding” is too open for most research work. “B2B SaaS onboarding friction, implementation delays, time to value, and setup abandonment” states the market, the problem, and the adjacent language that belongs in the result set.
The query test
Someone who did not write the query should be able to say which stories belong, which stories do not, and what decision the feed will support.
Build the query from four parts
Use the parts that sharpen the boundary. A narrow research job may need all four; a named topic may need only two.
| Part | Question it answers | Example |
|---|---|---|
| Subject | What market, workflow, or category is this about? | B2B SaaS onboarding |
| Audience | Whose situation should the coverage reflect? | Product-led growth teams |
| Problem | What tension or decision makes a story relevant? | Setup abandonment and slow time to value |
| Evidence | What kind of material is worth surfacing? | Product changes, operating examples, and customer research |
Broad feeds fail quietly
A broad feed rarely looks broken. It returns plausible stories, which makes the noise easy to accept. The failure appears later when the same feed cannot support a brief, a market note, or a product decision. “Marketing” may be interesting to read, but it does not define a stopping rule.
Tightening the query does not mean stuffing it with synonyms. Add the distinctions that change inclusion. If geography, company size, buyer role, or evidence type would change your judgment, include it. If a word only repeats the same idea, leave it out.
Diagnose the first result set
Review the feed once before saving it. The visible stories and source mix tell you which part of the query needs work.
- 01
Too many adjacent topics
Add the audience or decision that separates the intended subject from neighboring coverage. Replace “content strategy” with the actual job, such as source research for B2B editorial planning.
Completion signal: You skip relevant-looking headlines because they serve a different reader or decision.
- 02
The same phrase repeated everywhere
Remove wording that forces exact phrasing and describe the underlying problem instead. Related terms should have room to enter the set.
Completion signal: The feed looks like an exact keyword alert rather than a view of the subject.
- 03
Useful stories mixed with weak commentary
Name the evidence you want: product changes, original research, implementation examples, standards, or attributed analysis.
Completion signal: The subject is right, but the material cannot support the work that follows.
Use exact search for fixed identifiers
Semantic matching is not the right tool for every lookup. Product names, error codes, quotations, legal clauses, and other fixed identifiers are better handled with exact search. Use semantic search when the same idea can appear under different wording and the relationship between concepts matters more than a literal phrase.
A research workflow can use both. Keep exact alerts for named events and semantic feeds for the broader customer problem or market question around them.
Know when the query is finished
The query is ready when the current set contains material you would actually save, the irrelevant stories fail for understandable reasons, and a teammate can explain the feed boundary without your help. Save that query as the durable research object. Change it only when the decision or market scope changes.
curaft shows the current stories, recurring sources, coverage window, and tagged themes for focused public feeds. Those counts describe the loaded sample. They are useful for inspecting the result set, not for claiming market share or statistical trends.
Continue the research
Use a live topic for current material or move to the guide that owns the next part of the workflow.