The Method Everyone Claims and Few Execute
Thematic analysis is the default method of qualitative research. It shows up in academic papers, product research decks, UX readouts, and market studies -- anywhere someone has a pile of interviews and needs to say what they mean. And precisely because it is so accessible, it is the method most often done badly. "We did a thematic analysis" frequently means "we read the transcripts and wrote down what stood out," which is not thematic analysis at all. It is impressionistic summary wearing a methods label.
The framework that gives thematic analysis its credibility -- Braun and Clarke's six phases -- is not bureaucratic overhead. Each phase exists to catch a specific way that human pattern-matching goes wrong: seeing themes that confirm what you expected, anchoring on the first vivid quote, mistaking a topic for a theme. Skip a phase and you reintroduce exactly the bias that phase was designed to remove. This piece walks through all six, and flags the corner-cutting that quietly hollows out most analyses.
Phase 1: Familiarization -- Reading Until It Is Uncomfortable
Familiarization means immersing yourself in the raw data before you code a single line -- reading and re-reading transcripts, listening back to recordings, noting first impressions. The discipline is to read the entire corpus before forming conclusions, not to start theming after the third interview because a pattern already "seems obvious."
The corner most teams cut: they read each transcript once, in the order collected, and let the earliest sessions disproportionately shape the frame. That is how an availability cascade takes hold -- the loudest early data point becomes the lens for everything after it. Real familiarization means reading across the whole set, deliberately, before you decide what matters -- and reading the quiet transcripts as carefully as the dramatic ones.
Phase 2: Generating Initial Codes -- Comprehensive, Not Selective
Coding is where you systematically tag features of the data that are interesting or relevant. The rigor here is coverage: you code the entire dataset, not just the passages that support the story forming in your head. Every segment that carries meaning gets a code, including the ones that contradict your emerging narrative.
This is the phase where confirmation bias does its quietest damage. When you only code what fits, you manufacture a pattern that was never in the data -- you consulted your own hypothesis and mistook it for a finding. The safeguard is coding for disconfirmation as deliberately as for confirmation: actively tagging the moments that break your expected story. Comprehensive coding is also what makes the difference between reading a topic and reading a genuine theme visible later, because you cannot compare patterns you never captured.
Phase 3: Searching for Themes -- Themes Are Patterns of Meaning, Not Buckets of Topics
Here you collate codes into candidate themes -- but this is where the single most common conceptual error lives. A theme is not a topic. "Pricing" is a topic. "Participants distrust pricing they cannot predict" is a theme: it carries a shared meaning, a point, an argument. Most weak analyses stop at topics -- they sort quotes into labeled folders and present the folders as findings.
A real theme captures something meaningful about the data in relation to your research question. It has a center of gravity. The test: can you state the theme as a sentence that makes a claim? If the best you can do is a noun ("onboarding," "support," "trust"), you have a topic, not a theme, and you have more work to do. This is the same distinction that separates surface-level coding from analysis that actually earns its rigor at speed.
Phase 4: Reviewing Themes -- The Phase Almost Everyone Skips
This is the quality gate, and it is the phase teams most reliably skip because it feels like rework. Reviewing themes happens at two levels. First, against the coded extracts: do the quotes assigned to a theme actually cohere, or have you lumped together things that only superficially relate? Second, against the entire dataset: does the thematic map represent the data as a whole, or does it over-index on a handful of memorable sessions?
Skipping this phase is how analyses ship with themes that fall apart under scrutiny -- a theme that sounded strong turns out to rest on two participants, or two "separate" themes turn out to be the same idea described twice. This is also where you catch the synthetic saturation illusion, where new data points feel like confirmation but are really the same signal echoed back. Reviewing is not optional polish; it is where the analysis either earns its claims or exposes that it cannot.
Phase 5: Defining and Naming Themes -- Precision Forces Honesty
Defining themes means writing a tight, specific account of what each theme is and is not -- its scope, its boundary, the exact aspect of the data it captures. This sounds like a documentation chore, but the act of writing a precise definition is a rigor check in disguise: a theme you cannot define crisply is usually a theme that does not actually hold together.
Naming matters too. A vague name ("user experience issues") invites the reader to project their own meaning; a precise name ("friction from unexplained state changes") tells them exactly what the theme claims. The discipline of definition is what prevents the narrative coherence bias from smoothing genuine tension into a tidier story than the data supports. If two themes resist clean definition because they keep bleeding into each other, that is the data telling you your map is wrong -- listen to it.
Phase 6: Producing the Report -- Analysis, Not a Quote Parade
The final phase is writing up -- and the failure mode is the quote parade: a theme name followed by three supporting quotes, repeated for each theme, with no analytic argument connecting them. Quotes are evidence, not analysis. The report has to make an argument: why these themes, how they relate, what they mean for the research question, where the tensions and contradictions live.
Good write-up integrates disconfirming cases rather than hiding them, states the strength of evidence honestly, and connects themes into a coherent account of the phenomenon. It is the difference between "here is what people said" and "here is what it means." This is also where reflexivity belongs -- naming how your own position shaped what you saw, so the reader can weigh the analysis rather than take it on faith.
Where AI Fits -- And Where It Does Not
AI genuinely accelerates the mechanical phases: it can code an entire corpus comprehensively in minutes (phase 2), surface candidate patterns across every transcript at once (phase 3), and flag where a theme rests on thin evidence (phase 4). That is real leverage, especially for the coverage problem -- a model does not get bored on transcript forty and start skimming.
But the interpretive core -- deciding what a pattern means, where the theme boundary sits, which tension matters -- remains human judgment. The right posture is AI-assisted, human-owned: let the model do the exhaustive tagging and cross-referencing so the researcher spends their attention on meaning, not mechanics. That is exactly how thematic analysis at 10x speed keeps its rigor -- by automating the labor and preserving the judgment, rather than outsourcing the thinking.
The Bottom Line
Thematic analysis is not hard to describe and genuinely hard to do well. The six phases are not a ritual; each one removes a specific failure mode, and the two most-skipped phases -- comprehensive coding and theme review -- are precisely the ones that separate a defensible finding from a confident guess. If your analysis skips straight from reading transcripts to writing themes, you are not doing thematic analysis. You are summarizing, and calling it method.
The fix is not more sophistication. It is discipline: read the whole corpus, code all of it, test every theme against the full dataset, define each one until it either holds or breaks, and write an argument instead of a quote parade.
Running thematic analysis on a real body of qualitative data? Qualz.AI codes your entire corpus comprehensively, surfaces candidate themes across every transcript, and flags where a theme rests on thin evidence -- so your team spends its judgment on meaning instead of mechanics. Book a demo to see rigorous thematic analysis at practitioner speed.
