Overview
A support twin is your product's twin rather than your own. It answers in a neutral product voice, not a personality — nobody is pretending to be a person, and that is deliberate. It is created from the twin switcher and lives alongside your personal twin rather than replacing it.
The thing that makes it usable in front of real customers is what happens at the edge of its knowledge. A support bot that improvises a refund policy costs you more than no support bot at all. This one says it doesn't know, shows the searches it ran before giving up, and offers to pass the question on.
Why answer from your docs instead of training a bot on tickets
Your docs are the version of the truth you already decided to publish. Tickets are a record of confusion, often resolved by a person improvising, sometimes wrongly. Grounding in docs means the twin's answer and your documentation cannot silently disagree — and when the docs are wrong, you fix one file rather than retraining anything.
It also gives you a receipt. Every grounded answer carries the passage it came from, so a customer who doubts an answer can open the source, and a support engineer reviewing a bad conversation can see exactly which paragraph misled it. That is a much shorter debugging loop than guessing at a model's memory.
The gaps are the other half of the value. Questions your docs don't answer are the highest-signal product feedback you will get, and they arrive already sorted by how often they are asked.
Step by step
Create the support twin
Open the twin switcher in the dashboard rail and choose New support twin. It asks for two things that cannot be defaulted: the product's name and the handle it lives under.
This creates a twin separate from your personal one, with its own knowledge, its own rules and its own audience. Your public twin stays exactly as it was.
The switcher is where every twin after your first one starts. A support twin and a team twin are different products, not settings on the same one. Give it your documentation
Point it at the docs you already have. You can paste text straight in, upload files, or give it the URL of a help centre or site and let it read the pages itself.
Start with the twenty pages that answer eighty per cent of your tickets rather than the whole archive. A smaller, current corpus produces better answers than a large one full of superseded pages — and you can add more the moment you see a gap.
Every source type in one place. For a support twin the useful ones are usually paste, file upload and your help-centre URL. Read it back before anyone else can
Use the Test chat — a private sandbox that doesn't touch your message allowance, your Insights or your gap queue, and works before you publish. Ask it the five questions your team answers most often.
Check the citation, not just the answer. A correct-sounding answer with a citation pointing at the wrong passage is a problem you want to find now, while nobody is watching.
The answer is assembled from the docs, in a neutral product voice, with the source attached. Open the chip and you get the passage it used. Test the edge on purpose
Now ask it three things your docs genuinely do not cover. This is the test that matters, and it is the one demos skip.
What you want to see is a refusal that looks like a refusal: the dashed treatment, an honest admission, the searches it ran before giving up, and an offer to pass the question to a human. What you do not want is a fluent paragraph inventing a discount policy.
The gap is a designed moment, not an error state. It offers the handoff, and the question goes to your queue so one answer from you closes it permanently. What happens at the edge of the material Wire it into your product
A support twin has no public link, which is the point: your own site vouches for its visitors. Your server mints a short-lived token with the signing secret, and the widget carries it. Nobody can reach the twin with the URL alone.
The Publish tab gives you the secret, the token-minting code to run server-side, and the widget snippet. Keep the secret on your server — never in a browser bundle, never in your repo.
Three steps: the secret, a token your server mints per signed-in user, and the widget. Fifteen minutes of token life is plenty. How a visitor is let in Answer the gaps once, and keep answering them
Watch the gap queue for the first fortnight. Each unanswered question is either a missing docs page or a real product gap, and you will be able to tell which within a week.
Answer it once in the dashboard and the twin uses your answer from then on, overriding older material. That is the loop that makes the twin better than the docs it started from — and it turns support volume into documentation instead of away from it.
What this will not do
Every playbook carries this section, because a recipe that only lists what works is marketing.
- It is text only. There is no voice or audio twin, and no phone support.
- It will not make a twin safe for regulated advice. Answering from your docs is not the same as being qualified to advise, and no configuration changes that.
- It does not connect to Intercom, Zendesk or a CRM today. Handoff is to your inbox; helpdesk integrations are something we are exploring, not something you can switch on.
- A citation proves where an answer came from, not that the answer is right. If the docs are wrong, the twin will be wrong and show you exactly which paragraph made it so.
Questions people ask first
- Can people find the support twin without logging into my product?
- No. A support twin has no public link and is not listed anywhere. It is reachable only through a short-lived token your own server mints with your signing secret, so a URL on its own gets nobody in.
- What happens when it can't answer?
- It says so, in a designed refusal marked as not in the source material, and shows the searches it ran first. It offers to pass the question on, and the question lands in your gap queue where one answer from you closes it for good.
- Does it sound like a person pretending to be my team?
- No. A support twin answers in a neutral product voice rather than a personality, and every twin carries an AI disclosure label that no plan can remove.
- How much does this cost to run?
- The Free plan includes a real published twin, and visitors always chat free — you are never charged for someone asking a question. The paid plan is Founding Creator at $49 a month for scale; the honesty features are never behind it. See the pricing page for what can be bought today.
- How often do I need to re-add my docs?
- Whenever they change materially. A list can already refill itself from a JSON or CSV feed on a schedule, but scheduled re-reading of your docs pages is not shipped — today you re-add or correct, and a correction overrides older sources immediately.
Take it further
- Public twin
Turn repeat questions into booked calls
Open it - Team twin
A team twin for the questions colleagues keep asking you
Open it
Build it free, and test the edge first
The Free plan includes a real published twin, and visitors always chat free. The private sandbox lets you break it before anyone else can — which is the step worth not skipping.