Discourse Analysis: what language is doing, not just saying

Language is constructive. Accounts do not neutrally report a world; they assemble a version of it, from a range of possible versions, and the assembling is selective. The analytic question is which version, built how, and at whose cost.

Language is action-oriented. Talk performs social acts — justifying, blaming, excusing, claiming entitlement, disclaiming interest. A speaker who says "I am not being funny, but…" is doing pre-emptive defensive work, and that work is the data.

Variation is a resource, not error. The same person will give incompatible accounts within one interview. Most methods treat that as inconsistency to be resolved or averaged; discourse analysis treats it as the clearest available evidence that different accounts were doing different jobs at different moments.

From these follows the concept of the interpretative repertoire: a recognisable, culturally available cluster of terms, metaphors and figures of speech that people draw on to construct a version of events. Repertoires are the units most discursive analyses report.

"Discourse analysis" names several distinct enterprises, and studies get into trouble by not saying which. The two that matter most for qualitative research are these.

Discursive psychology, developed by Jonathan Potter and Margaret Wetherell, focuses on how psychological matters — memory, attitude, motive, blame — are constructed and managed in ordinary talk. Their Discourse Analysis in Rethinking Methods in Psychology is the standard methodological statement, and Potter's later Discourse: Noun, Verb or Social Practice? clarifies what the term is being asked to do.

Critical discourse analysis adds an explicit interest in power and ideology, asking how discourse reproduces domination and how it can be contested. Teun van Dijk's Critical Discourse Analysis in Discourse and Power is a compact statement, and his Discourse Analysis as Ideology Analysis sets out the ideological reading directly.

The practical difference is where the analysis lands. Discursive psychology stops at what the talk is doing in its immediate context; CDA connects that to structures of power beyond the text. Both are legitimate; conflating them produces analyses that gesture at ideology without evidencing it.

A study of how selection panels account for their decisions recorded twelve post-interview debriefs. A content analysis would report the criteria mentioned. The discursive reading found something else.

Two repertoires ran through every panel. A merit repertoire — "the strongest candidate on paper", "objectively the best fit" — constructed decisions as the outcome of dispassionate assessment. A fit repertoire — "he would work well with the team", "you can just tell" — constructed them as intuitive recognition. They are incompatible accounts of how decisions are made, and panels moved between them fluently.

The movement was systematic. The merit repertoire appeared when the chosen candidate's advantage was documentable, and the fit repertoire when it was not. More tellingly, the fit repertoire was almost always accompanied by stake inoculation — "and I know how that sounds", "this is not about him being one of us" — which is a speaker defending against a charge nobody had made. The pre-emptive defence marked precisely the moments the panel treated as vulnerable.

One deviant case tested the reading: a panel that used the fit repertoire with no defensive work at all. On inspection they were describing a candidate who had already worked in the team, which made the appeal to fit documentable and moved it back into merit territory. The exception supported rather than undermined the pattern.

The finding is not that panels are biased — the data cannot show that. It is that panels themselves treat fit-based reasoning as accountable, and manage it accordingly. That is a discursive finding, and it is directly useful to anyone designing panel training.

• Content analysis in discourse vocabulary. Counting how often a word appears is not discourse analysis. The question is what the word is doing.
• Treating talk as evidence of belief. "Participants believed X because they said X" is the assumption discourse analysis exists to reject.
• Sanitised transcripts. The features the analysis depends on are the ones tidy transcription removes.
• Naming the tradition too late, or never. Discursive psychology and CDA make different claims and need different evidence. Say which you are doing.
• Ideological claims without discursive evidence. CDA is vulnerable to finding the power relations it went looking for. The link from a textual feature to a structural claim has to be shown, not asserted.
• Ignoring variation. Smoothing out a speaker's inconsistencies discards the method's best evidence.
• No deviant cases. A reading that has never been tested against material that does not fit is an illustration.

Discourse analysis has an unusually concrete quality standard: the reader should be able to see the claim in the extract. Analyses are expected to carry substantial verbatim material with the analytic reading laid alongside it, so a sceptic can look at the same lines and disagree.

The other markers are deviant-case analysis, attention to the sequential context of an extract rather than lifting it out, participants' own orientations as evidence (does the next speaker treat the utterance the way the analyst does?), and explicit acknowledgement of the researcher's part where the data is interview-based.

Discourse analysis says nothing about what is true or what people privately think, and it is not designed to. Researchers who need to know whether an account is accurate are using the wrong method.

It is also slow and small-scale. Detailed transcription and line-by-line reading make large corpora impractical, so findings are about how something is done rather than how widespread it is.

The interpretive latitude is real. Naming a repertoire is an analytic act, and two analysts can name different ones from the same corpus. The discipline is extract-level evidence and deviant-case testing rather than agreement statistics, and reviewers unfamiliar with the tradition often want the latter.

CDA carries an additional risk worth naming: because it begins from a critical stance, it is structurally prone to confirming it, and the strongest CDA is unusually careful about showing the discursive work behind every structural claim.

Conversation analysis works at a finer grain on the sequential organisation of interaction, and is the more rigorous choice where turn-by-turn structure is the question.

Critical discourse analysis is the variant that connects discursive patterns to power and ideology.

Narrative analysis attends to the shape of an account rather than to what individual utterances accomplish.

Summative content analysis looks at word frequency in context and is the descriptive counterpart — useful, and answering a different question.

The line-by-line reading cannot be delegated, and a tool that summarises what a transcript is about is answering the question discourse analysis declines to ask.

Where tooling helps is corpus management: locating every instance of a candidate repertoire across twelve transcripts with enough surrounding turns to preserve the sequential context, keeping analytic notes attached to specific lines, and assembling the extract set for a claim so that deviant cases surface rather than getting quietly skipped. Evidano supports discourse analysis as a named methodology and retrieves passages with their surrounding turns rather than as isolated fragments. Naming what the talk is doing is the analysis, and it stays with the analyst.