When is keyword searching better than predictive coding, and can you use both together in the same review?

Keyword searching and predictive coding (a form of technology-assisted review, or TAR) are not rivals so much as different tools for different jobs. The two are routinely used together, and the right mix depends on the data, the issues, and proportionality — the principle, reflected in the Federal Rules of Civil Procedure, that discovery effort should match the needs of the case. Note that rules and expectations vary by jurisdiction, including state courts and other countries.

When keyword searching is the better fit

Keyword searching shines when relevance is tied to specific, predictable terms.

• Small or simple collections, where the cost of training a model isn’t justified.
• Distinctive identifiers — project names, account numbers, parties, product codes — that reliably signal relevant material.
• Targeted requests, such as locating every document mentioning a single contract.
• Transparency needs, when parties must negotiate and document an agreed, defensible search protocol.

Keywords struggle, however, with synonyms, jargon, misspellings, code words, and the difference between a term used casually and one used meaningfully. High-volume terms can also sweep in large amounts of noise.

When predictive coding is the better fit

Predictive coding learns from human coding decisions to rank or classify documents by likely relevance. It tends to outperform keywords when:

• Volumes are large and manual review of everything is impractical.
• Relevance is conceptual rather than keyword-bound, so meaning matters more than exact wording.
• You need to prioritize review so the most important documents surface first.

It requires careful workflow design, quality-control sampling, and often disclosure to opposing parties.

Using both in one review

Combining the methods is common and often advisable:

1. Cull first with searches (date ranges, custodians, file types, or keywords) to remove clearly irrelevant material.
2. Apply predictive coding to the remaining set to rank and accelerate review.
3. Validate results with statistical sampling and targeted keyword checks for high-risk terms.

The Sedona Conference encourages cooperation, transparency, and defensibility in whatever method you choose. Document your process, test it, and be ready to explain it.

For broader context, see e-discovery.