Part of the 20–26 October editorial series. Prepared and source-checked on 26 October 2026.
Use the score to organise review
A search API's authority score can help organise a research queue. It should not decide whether a particular sentence is true. Keep the provider's assessment, then inspect whether the source actually supports the claim, how current it is and whether it is the original evidence.
DeepSeek opened its in-house Search API on 17 October 2026. Its reference documents an authority field with tiers from zero to four and minus one for unknown. That is provider metadata, not a Google authority score. The review policy below is my recommendation; no paid API experiment was performed for this article.
Keep the returned fields separate
The reference also describes a rerank score combining relevance, authority and freshness, a publication date that can be unknown, and distinct snippet, summary and content fields. Preserve those distinctions when normalising results for another agent.
Do not collapse everything into a single internal label such as trusted. A relevant result can support a narrow point while failing to support the proposed conclusion. An unknown publication date should remain unknown, not silently become the retrieval date. A generated search summary is a discovery aid; where the claim matters, inspect the source material rather than cite the summary as if it were the original.
Preserve unknown instead of converting it to poor quality
Missing metadata and negative evidence are different. If a provider cannot assign an authority tier, retain that state so a reviewer can distinguish an unassessed source from a low-ranked one. Record failed fetches separately from pages that were opened but did not support the claim.
For a fictional specialist query, a small standards group's original technical note might have sparse metadata while a large publication summarises it. A rule that discards unknown values could remove the primary source. A useful review queue would preserve both, identify their relationship and let the reviewer decide which evidence answers the question.
Check support at the claim level
For each material statement, retain the source URL, relevant passage, stated date, evidence role and limitation. Ask whether the page directly establishes the statement or only repeats someone else's assertion. An authoritative institution can still publish a forecast, opinion or announcement rather than an independently measured outcome.
Give the reviewer room to narrow the wording. A source may establish that a vendor announced a capability while leaving its reliability untested. The resulting sentence should say what was announced and under which conditions. Raising the authority threshold cannot turn that announcement into evidence of a completed business result.
Review what your threshold removes
Before adopting a filter, use a small authorised set of representative questions and inspect the excluded results as well as the retained ones. Include niche topics, unknown dates and primary documents. Write down which useful sources disappear and which unsupported claims still pass.
Set the policy around the cost of the decision. A low-stakes reading queue and a consequential procurement recommendation need different review requirements. Keep the provider's original score so you can revise the policy without rerunning every search. Do not imply that choosing a particular threshold is a validated defence against misinformation.
Make source selection explainable
Retain the reviewer decision and its reason alongside the collected metadata. If a result is admitted despite an unknown tier, record the direct evidence it supplies. If a high-tier result is rejected, record the missing support. Those reasons make future changes easier to assess than a silent numeric cutoff.
This is a useful design choice for an AI brief generator or a scheduled research system. It connects retrieval to an accountable evidence decision. If evidence remains insufficient after review, use a separate stop rule for the conclusion rather than forcing the highest-scored result to become the answer.
Link Map
Sources and editorial review
Sources checked on 26 October 2026. The publication date is assigned to the 20–26 October editorial series; this edition was prepared and reviewed on 26 October 2026. Proposed workflows and illustrative examples are JQ AI SYSTEMS editorial analysis, not reports of completed client tests.
- DeepSeek: 17 October change log (17 October 2026; First-party launch announcement).
- DeepSeek: Search API reference (publication date not shown; checked 26 October 2026; Official API reference; publication date not shown).