Semantic Coding: staying at the surface on purpose

A semantic code names something present in the data’s explicit content. If a participant says “I stopped going because the bus route changed”, semantic codes are available for transport barriers and discontinued attendance; a code for institutional abandonment is not — nothing in the utterance said it.

The commitment runs through the whole analysis: themes built from semantic codes summarise patterned explicit content, and the write-up’s claims stay at the level of what participants reported. Progression within the level is from description toward organised meaning — grouping, comparing, and interpreting the significance of what was said, not decoding what was meant.

The distinction with latent coding is a decision, not a discovery. The same dataset supports both; the research question dictates the level, and the methods section should state the choice and hold to it.

The semantic/latent distinction is drawn in Braun and Clarke’s foundational Using thematic analysis in psychology, which remains the necessary citation and the clearest statement of what each level claims.

For a worked demonstration of the full analytic arc — including how semantic themes are developed and named — Byrne’s A worked example of Braun and Clarke’s approach to reflexive thematic analysis is the most useful single paper to imitate.

The neighbouring content-analysis tradition calls approximately the same surface “manifest content”; Graneheim and Lundman’s Qualitative content analysis in nursing research is the standard source for that vocabulary and its quality concepts.

• When stakeholders need findings they can verify. Semantic themes trace visibly to quotes; a programme board can check the chain themselves.
• When the question is about stated experience: reported barriers, expressed needs, described events. The said is the object, so the surface is the site.
• When coding must be consistent across a team or at scale. Surface meaning supports shared definitions and agreement checks in a way latent interpretation resists.
• When the corpus is large and heterogeneous — hundreds of open-ended survey responses reward systematic surface coding and punish deep-reading ambitions.
• Not when the interesting content is unsayable or unsaid — stigma, ideology, identity work. Forcing those questions through semantic coding yields themes about euphemisms.

A clinic analysed 240 open-text survey responses and 18 short interviews about missed video appointments. The question — what do patients report as reasons — was explicitly semantic, and the codebook’s inclusion rules required stated content.

Coding produced 31 codes consolidating into four semantic themes: technology failures at join time (stated by 41% of coders’ units), appointment times colliding with caregiving, not seeing the point for check-ins (“nothing to examine”), and privacy at home (“nowhere in the flat to talk about this”). Each theme’s name stayed within the reported; each carried counts, verbatim exemplars, and its negative cases (six patients explicitly preferred video for privacy — reported alongside).

The level discipline mattered at write-up. A draft sentence reading “appointments were missed because telehealth erodes the felt legitimacy of care” was cut: nothing in the data said it, however plausible. The delivered version — patients state they skip appointments that feel purposeless without examination — supported the same service change (repurposing check-ins) while claiming only what 240 people had actually written.

• Paraphrase presented as themes. “Patients mentioned technology problems” restates data; a theme organises it — prevalence, variants, conditions, exceptions.
• Level drift. Codes that start semantic and quietly become interpretive mid-corpus, so early and late transcripts are coded to different standards.
• Topic coding mistaken for meaning coding. “Transport” as a bucket for everything mentioning buses mixes complaints, praise, and asides; semantic codes capture stated meaning, not vocabulary.
• Counting without context. Semantic coding supports counts, but a frequency table with no analysis of variation within codes wastes the qualitative data.
• Apologising for the level. Semantic analysis is not failed latent analysis; hedging it as “merely descriptive” undersells work that was, correctly, built to be checkable.

The level’s honesty is its ceiling: semantic analysis cannot speak to what participants could not or would not say, and on sensitive topics the surface is systematically curated. The method reports the presented self.

Themes can also inherit the question’s framing — ask about barriers and the data will contain barriers — so instrument wording belongs in the interpretation.

And checkability invites false confidence: agreement on surface codes is achievable, but agreement is not validity, and a reliably applied shallow codebook yields reliable shallowness. The remedy is a codebook built from the data’s own range, revised as coding teaches.

Latent coding is the deliberate alternative: interpreting underlying meaning, with different evidence obligations.

Reflexive thematic analysis is the full framework within which the semantic/latent choice is made.

Conventional qualitative content analysis is the neighbouring tradition for surface content, with its manifest/latent vocabulary.

Framework analysis suits applied semantic work needing case-by-theme matrices for teams.

Semantic coding is the level AI assists best, precisely because the standard is explicit content: rule-based codes over stated meaning are what Evidano applies across interview and survey corpora, with every application linked to its quote so the checkability the level promises is real — reviewers audit by clicking, and frequency and cross-segment comparisons come with the evidence attached. Published comparisons of exactly this kind of codebook-driven coding report agreement with expert human coders in the 92–96% range.

The researcher still owns the codebook’s categories, the theme construction, and the significance argument. The tool guarantees the tether to the said; what the said amounts to remains yours to argue.