Persona Development

What should a persona actually contain?

Seven elements earn their place in almost every case, because each one changes a decision: the goal the person is pursuing, the job to be done, the context and constraints they operate under, their current behaviour, the criteria they use when choosing, the barrier that stops or slows them, and the unmet need. Anything beyond that must justify itself against a named decision. The test is specific: would a different value on this attribute change any decision the persona is used for? Favourite coffee, car, family composition and weekend hobbies almost never pass it, unless the category makes them load-bearing.

Is it acceptable to invent details to make a persona feel realistic?

No. This is the central prohibition. Invented age, occupation, household, daily routine and possessions are indistinguishable from measured ones once they are on the page, so downstream users design for a fiction while believing they are following research. If an element has no evidence, write "not established" and leave it. A persona with four evidenced elements and three visible gaps is honest and useful; the same persona with the gaps filled in plausibly is a fabrication that will outlive the study.

Can I write a quote for a persona?

Only if a real participant said it, and it should carry that participant's identifier. A composite sentence in quotation marks is a fabricated quote no matter how faithfully it summarises the theme, and a real quote reassigned from participant P07 to a persona called "Sarah" is misattribution. Where no suitable real quote exists for a point, the persona carries no quote there. In practice, real quotes with identifiers give a persona far more credibility and more emotional weight than invented ones, which tend to read as written by a marketer.

How many personas should we have?

As many as are genuinely distinguishable, which is usually three to five and often fewer than requested. Run two tests. The substitution test: take a claim from persona A and put it in persona B; if it still reads as true, they are not distinguishable on that element, and personas surviving substitution on fewer than two or three elements should be merged. The decision test: would the team make a different choice for A than for B? Beyond four or five, teams stop distinguishing and default to the one they remember, so the extra personas cost attention and change nothing.

Should personas be built from a segmentation or from qualitative research?

Either, and they carry different confidence. A persona built on a validated segmentation can carry population sizes, tested differences and declarative statements. A persona built on depth interviews carries a described pattern with participant counts, uses hedged language, and does not carry percentages, because qualitative evidence does not support population estimates. A set built from both should say, element by element, which stream each claim came from, or the mixed artefact silently promotes the weaker evidence to the confidence of the stronger.

How do I check whether existing personas are evidence-based?

Take each statement and try to source it. Produce a marked-up version showing sourced claims, unsourced claims and claims contradicted by the data. In most inherited sets, somewhere between a third and two-thirds of the statements have no traceable source. The marked-up version is more persuasive than a rebuild, because it names the specific claims currently circulating as fact, and it usually makes the case for re-evidencing on its own.

Why do personas stop being used?

Three common reasons. They contain nothing decision-relevant, so the team has no occasion to consult them. They are not distinguishable, so the team defaults to one. Or they read as fiction to people who know the customers, which happens when they are full of invented character. A fourth reason is silent decay: a persona looks identical on the wall in year four, and nothing on the artefact says when the evidence was collected. Put the fieldwork date and a re-validation trigger on the face of every persona.

What is compression damage?

A persona is an average with a face, and the people it hides are the ones the design will fail. A team designs for the persona and misses the third of the group who behave differently. The fix is two lines on every persona: what varies within this group on the elements that matter, and what evidence cut against this characterisation. It is a small addition that prevents the largest downstream error, and it is never present in a fabricated persona, which makes it a useful audit signal.

Can AI generate personas?

It generates them extremely fluently, and that is the problem. A persona template has slots, and a language model fills them plausibly whether or not evidence exists, producing a uniformly complete set with no gaps anywhere. Uniform completeness is a warning sign rather than a mark of quality: real evidence bases are uneven. Use AI for the mechanical work of extracting elements from a corpus and attaching source references, and check the output by trying to source every statement in it.

What if the client asks for names and photographs?

Offer the strongest honest alternative rather than refusing flat. Label personas by their defining difference rather than a first name, so nothing invites the reader to invent the rest. Use no photograph that could be read as a picture of a participant; an abstract illustration or none at all is safer, and any image used is credited as an illustration. Carry the human weight with real, identified quotes instead. Teams generally find the real quotes more compelling than the name and face they asked for.