Short answer: Discourse analysis is a family of qualitative methods that study how language does things: how it builds identities, frames problems and sustains or challenges power. Unlike thematic analysis, which asks what participants said, discourse analysis asks how and why it was said that way. Choose a tradition (critical, discursive psychology, Foucauldian or Gee-style) before you collect data, because each one implies a different research question and unit of analysis.
Discourse analysis in qualitative research is often the method PhD students reach for when thematic analysis feels too flat. Their interviews are full of hedges, justifications, borrowed phrases and institutional jargon, and the interesting finding is not the topic but the way people talk about it. This guide explains what discourse analysis is, compares its four main traditions, gives a step-by-step process and a worked example, and sets out where AI tools help and where 2026 research says they still fall short.
\n\nDiscourse analysis is the qualitative study of language in use. It treats talk and text as social action: people use words to justify, blame, include, exclude, persuade and position themselves and others. The analyst examines word choice, metaphors, grammar, argument structure and the wider social context to explain what a piece of discourse accomplishes and what it makes possible to think or say.
\nThe data can be interview transcripts, focus group talk, policy documents, news articles, social media posts or meeting recordings. What changes is the question. A thematic analyst asks \"What do nurses say about workload?\" A discourse analyst asks \"How do nurses construct workload as a personal failing or as an institutional problem, and what does each construction allow?\"
\n\nAs Gee notes in An Introduction to Discourse Analysis, no single approach is uniquely right. Different approaches fit different questions. The table below summarizes the four traditions you are most likely to meet in social science, health, education and management research.
\n| Approach | Core question | Unit of analysis | Typical data | Good fit when… |
|---|---|---|---|---|
| Critical discourse analysis (Fairclough) | How does language reproduce or resist power and inequality? | Text, discursive practice and social practice (three dimensions) | Policy, media, organizational documents, interviews | Your question is about ideology, power or social change |
| Discursive psychology (Potter and Wetherell) | What are people doing with their talk in this interaction? | Turns, accounts, interpretative repertoires | Naturally occurring talk, interviews, focus groups | You study attitudes, identity or accountability as things done in talk |
| Foucauldian discourse analysis | What does a discourse make it possible to know, say or be? | Discursive formations, subject positions, historical shifts | Archives, professional texts, guidelines over time | Your question is about how categories (e.g., \"patient,\" \"user\") came to be |
| Gee's approach | How does language build significance, identities, relationships and politics? | Utterances analyzed through \"building tasks\" | Interviews, classroom talk, everyday conversation | You want a structured, teachable toolkit, common in education research |
Choosing between them is a design decision, not a stylistic one. Ong, Beaumont, Caine and Slemon (2024) describe two nursing master's students choosing between Fairclough and Foucault and conclude that the decision depends on the substantive focus, the researcher's methodological knowledge, and their positionality and interests (International Journal of Qualitative Methods, 2024).
\n\nThe three methods are often confused because they can use the same transcripts. The difference lies in the question you ask and in what counts as a finding.
\nA practical test: if your findings would survive paraphrasing the quotes, you are probably doing thematic analysis. If paraphrasing would destroy the finding, because the exact wording, hedging or metaphor is the point, you are doing discourse analysis. For life stories and sequence, consider narrative analysis instead.
\n\nThe steps below are a generic scaffold. Adapt them to the tradition you chose, and cite that tradition's own procedural texts in your methods section.
\nTake a hypothetical interview extract from a hospital manager: \"Obviously, we all want to give patients more time, but the reality is the system just doesn't allow it.\"
\nA thematic analyst might code this as \"time pressure.\" A discourse analyst would argue it shows an accountability-deflecting repertoire, and would then look across the corpus for when this repertoire appears, who uses it and what it shuts down.
\n\nWe use this checklist, created by Qualitati, before recommending discourse analysis to a research team. If you cannot tick at least five items, a different method may serve the study better.
\nPartly. LLMs can draft a first reading of framing and argument structure, but they miss implicit meaning, and a human analyst must make the interpretation.
\nThe most direct evidence comes from Francesca Cristiano and Emiliano Giovannetti at the Institute for Computational Linguistics of Italy's National Research Council (CNR-ILC). In a pilot published on June 22, 2026 in Umanistica Digitale, they built a consensus gold standard from three human raters analyzing thirty opinion articles from ideologically diverse newspapers about media coverage of the October 7 attack. An LLM answered the same analytical questions. It performed well on macro- and superstructural features but struggled with microstructural ones such as euphemism detection. A shorter version appeared at the LLMs4SSH workshop at LREC 2026 in May. The authors conclude that LLMs are best treated as aids to interpretation, not replacements for human analysts.
\nThat matches what we see in AI-assisted coding generally. Models are good at summarizing what a text argues and weaker at what it implies. Euphemism, irony, strategic vagueness and absent agents are exactly the features discourse analysts care most about. Models can also invent or alter quotes, which is fatal for a method built on exact wording. See hallucinated quotes in LLM qualitative analysis.
\n\n| Task | Delegate to AI? | Why |
|---|---|---|
| Transcription and corpus organization | Yes, with spot checks | Mechanical; verify against audio where wording matters |
| Finding all instances of a phrase, pronoun or metaphor family | Yes | Search and retrieval at scale; the human reads the hits |
| Summarizing the overall argument or frame of a text | As a first draft | Macro-level features are where 2026 evidence shows the most agreement |
| Detecting euphemism, irony, implicature | No, human-led | Micro-level, implicit meaning is where models struggled |
| Linking discourse to power, ideology or history | No, human-led | This is the interpretive claim the study makes; it needs theory and reflexivity |
Qualitati is a European, budget-friendly qualitative research platform for universities and research companies. It does not automate discourse analysis, and we do not think anyone should claim to. It can make the work around the close reading faster:
\nWho this is for: PhD students, research labs and qualitative research firms who want transcripts and first-pass coding handled so they can spend their time on interpretation. You can start on the free tier (30 credits per month, no credit card). See pricing.
\n\nMethodology note: the AI evidence above comes from a single pilot on one corpus of news opinion articles. It may not generalize to interview data or other languages. Have a methods supervisor review any AI-assisted discourse design.
\n\nIt is the study of how people use language to do things, such as justify, persuade, include or exclude, and what that language use reveals about identities, relationships and power in a given context.
\nThematic analysis identifies patterns in what participants say. Discourse analysis examines how they say it and what that wording accomplishes. Paraphrasing quotes usually leaves thematic findings intact but destroys discursive ones.
\nThe four most common in the social sciences are critical discourse analysis (Fairclough), discursive psychology (Potter and Wetherell), Foucauldian discourse analysis and Gee's approach. Conversation analysis is a closely related but distinct method.
\nThere is no fixed number. Discourse studies prioritize depth, so small corpora analyzed line by line are common. Justify the size by your question and the richness of the data; see our guide to qualitative sample size and information power.
\nThey can draft a first reading of overall framing. A 2026 pilot found good agreement with human analysts on macro-level features but weaker performance on implicit, micro-level features such as euphemism. Human interpretation remains necessary.
\n\nDiscourse analysis in qualitative research rewards researchers who care about how things are said, not just what is said. Pick a tradition that fits your question, keep the corpus small enough to read closely, show your extracts and use AI for the mechanical work around interpretation rather than the interpretation itself. Ready to collect and organize the data? Start free with 30 credits, or compare Qualitati with NVivo.
\n\nCristiano, F., & Giovannetti, E. (2026, June 22). Can large language models support critical discourse analysis? A pilot experiment. Umanistica Digitale, 10(24), 47–65. doi.org/10.60923/issn.2532-8816/23143.
\nGiovannetti, E., & Cristiano, F. (2026, May). Exploring the use of large language models in critical discourse analysis: A consensus-based pilot study. LLMs4SSH @ LREC 2026, 18–22. ACL Anthology.
\nOng, Beaumont, Caine & Slemon (2024). Selecting approaches for discourse analysis: Thoughts on nursing research. International Journal of Qualitative Methods, 23. doi.org/10.1177/16094069241299658.
\nGee, J. P. An Introduction to Discourse Analysis: Theory and Method. Routledge.
\n\nLast updated: October 15, 2026. This guide is an independent editorial summary; Qualitati is not affiliated with the authors or publishers cited.
\n","related":[{"slug":"second-cycle-coding-pattern-coding-2026","title":"Second Cycle Coding: Pattern, Focused & Axial (2026)","description":"Second cycle coding explained: pattern coding, focused coding and axial coding, with a worked example, a method selector, and what AI can and cannot do.","category":"Qualitative Data Analysis","date":"2026-10-08","readTime":"11 min read","author":"Qualitati Research Team"},{"slug":"open-axial-selective-coding-2026","title":"Open, Axial, and Selective Coding: A 2026 Guide","description":"Open, axial, and selective coding explained: what each grounded theory stage does, worked examples, the paradigm model, and where AI coding tools help in 2026.","category":"Qualitative Data Analysis","date":"2026-09-24","readTime":"11 min read","author":"Qualitati Research Team"},{"slug":"llm-vs-machine-learning-open-ended-survey-coding-2026","title":"LLMs vs Machine Learning for Open-Ended Survey Coding","description":"Should you code open-ended survey responses with an LLM, a fine-tuned model, or classic machine learning? A 2026 evidence review with a decision matrix.","category":"Qualitative Data Analysis","date":"2026-09-10","readTime":"9 min read","author":"Qualitati Research Team"}]}