Interview Agents

Writing good interview questions

The difference between an interview that learns something and one that just generates noise — phrasing, sequencing, and what to skip.

An AI moderator is only as insightful as the questions you give it to work from. This page covers the craft of asking questions that actually learn something.

One script, two modes

Custom interview agents support two approaches:

  • Specific questions — you write the questions; the agent asks them in order and follows up naturally within each.
  • Research goals (Pro+ tiers, coming soon ) — you write goals like "understand why users abandon checkout at step 2"; Seena generates questions at conversation time based on the goal and the specific session context.

Both work. Specific mode gives you predictability; goals mode gives you adaptivity. When you're learning a new domain, start with specific questions so you can see what's returning useful answers and iterate.

The structure that works

A custom agent should have one to three core questions, not seven. More than three and response quality drops — visitors fatigue and give short answers.

A good shape:

  1. A broad opening. Gives the visitor permission to talk. "What brought you here today?"
  2. A specific probe tied to the trigger. "You spent a while on the pricing page — what were you trying to figure out?"
  3. A forward-looking wrap. "What would make this decision easier for you?"

The agent will naturally follow up inside each question based on the visitor's answer — you don't need to script those follow-ups.

Open questions beat closed ones

Closed questions ("Did you find what you needed? Yes / No") get one-word answers. Open questions ("What were you trying to find?") get paragraphs. At Seena's analysis layer, paragraphs are gold and single words are almost useless.

| Avoid | Use instead | | ------------------------- | ---------------------------------------------------------------------------- | | "Did this page help you?" | "What did you come to this page hoping to do?" | | "Is the pricing clear?" | "If you were explaining our pricing to a friend, where would you get stuck?" | | "Are you happy with X?" | "Tell me about the last time you used X." |

Don't ask questions you can answer with analytics

You already have analytics. Don't spend a visitor's 60 seconds asking them things you can look up.

| Don't ask | Because | | ------------------------------------- | -------------- | | "How did you find us?" | Referrer data. | | "Which pricing tier did you look at?" | Pageview data. | | "How long have you been using us?" | Account data. |

Use interviews for the questions analytics can't answer: intent, reasoning, emotion, and the gap between what a user did and what they wanted to do.

Avoid compound questions

"What did you like, and what was confusing?" is two questions. The visitor will answer one of them — usually the easier one — and skip the other. Split them:

  • "What's working well so far?"
  • "What's been confusing?"

Ask them one at a time.

Reference session context

Seena is aware of the session when it interviews the user — their journey, pages visited, time on each, the page they're on now, any custom session metadata you've attached. Write questions that lean into that context:

  • For users who bounced from docs to the pricing page three times "what are you cross-checking?"
  • For users on a paid plan "Tell me about the last decision we helped you make."

These questions feel personal and earn longer, more honest answers than generic ones.

Length discipline

The agent wraps when it runs out of useful scripted material or when the visitor's answers stop gaining information. As an author, your job is to set the ceiling — write one to three questions and let the agent decide when to end. Don't pad.

Bad phrasings to retire

  • Leading: "You'd like more features, right?" — visitors answer what you clearly want to hear.
  • Vague: "What do you think?" — about what?
  • Double-negatives: "Is there anything you don't not like?" — confusing, skip entirely.
  • Hypothetical with no anchor: "If we built feature X, would you use it?" — unreliable; people will almost always say yes.

Anchor hypotheticals in real behavior: "Tell me about the last time you needed something like feature X — what did you do?"

Iterate on the questions

The first version of a custom agent's questions is rarely the best. After the first 5-10 interviews land, read the transcripts (not just the insights) and rewrite anything that consistently produced shallow answers. It's a quiet hour a week and doubles your signal quality.

What to read next