Fix wrong or outdated AI Twin answers

When your AI Twin says something wrong, the fix is almost always in the source. Diagnose fast and remediate cleanly.

Updated June 23, 2026 7 min read

Reality Twin answers are grounded in your knowledge graph, with clickable citations for every response. When you spot a wrong answer, resist the urge to hard-code a response — instead trace it to the source, fix the source, and every future user (and future twin-to-twin exchange) gets the right answer.

Step 1 — Reproduce the bad answer

  1. Open Conversations and find the transcript in question.
  2. Ask the same question to your twin from the test panel to confirm it still misfires.
  3. Note the exact phrasing — small wording changes can flip retrieval.

Step 2 — Read the citation

Every reply shows which sources it used. Click the citation to jump to the exact chunk. Ask:

  • Is the chunk factually wrong?
  • Is it outdated?
  • Does it contradict another source?
  • Is it correct but retrieved without enough context?

Step 3 — Apply the right fix

Wrong fact

Update the source page or document. If it's a page, re-crawl. If it's a document, re-upload with the same name so the old version is replaced.

Outdated content

Publish the current version, then trigger a manual re-crawl — see Force a re-crawl of your website.

Contradictory sources

Two sources disagree, and retrieval picked the wrong one. Options:

  • Delete the losing source.
  • Update it to match the winning one.
  • Add a structured Q&A that authoritatively answers the question — Q&A ranks highest.

Missing context

The chunk is correct but retrieved in isolation. Add a summary paragraph at the top of the source, or add a Q&A that stitches the context together.

Step 4 — Verify

Ask the original question again — plus 2–3 rephrasings. Retrieval should now return the corrected source.

Never do this

Do not edit the twin's response text directly. It masks the bug in one conversation while leaving every other visitor exposed. Fix the knowledge, not the answer.

Best practices

  • Review a random sample of conversations weekly — five is enough to spot systemic issues.
  • Turn recurring misfires into structured Q&As so they never come back.
  • Keep a changelog of source edits — future you will thank you.

Frequently asked questions

Why does my AI Twin hallucinate?

It shouldn't. Reality Twin is retrieval-grounded — replies without a citation are refused or handed off. If you see a confident answer without citations, report it to hello@realitytwin.io.

Can I mark answers as wrong from the transcript?

Yes. Every message in Conversations has a thumbs-down. Marked messages appear under Knowledge → Review with the citation, so you can fix them in one place.

How quickly does a source fix take effect?

For document re-uploads: seconds. For re-crawls: 30–90 seconds. For structured Q&A: immediate.

Did this article solve your problem?

If not, email us — a human on the founding team replies, usually within a business day.