KnitKnot

# Investigate a claim in an AI answer

Start with a claim gap, open the captured answer, check the source signal, and compare the wording with your company facts.

## At a glance

  1. 1. **Find the topic**: Use Answer claims to spot a gap or a reported misrepresentation.
  2. 2. **Read the answer**: Open the captured response and inspect the exact claim in context.
  3. 3. **Check the basis**: Review source signals and compare the claim with approved facts.

The **Claims** view groups what measured AI answers say about your company by buyer topic. Start there when a report says AI is missing a capability or making a damaging claim. The counts point you toward an answer; the answer and its evidence are what you need to read before acting.

## Find a topic that needs review

  1. 1. Open **Intelligence → Claims** and choose **Answers**.
  2. 2. Use the **Claim gaps** filter to find topics with too little positive substance or a reported misrepresentation. **Low sample** means the topic has too few answers for a gap verdict.
  3. 3. Open a topic. The table shows answer counts, positive and negative claims per answer, reported misrepresentations, and competitor claims per answer. Those rates help you choose where to investigate; they do not identify a false sentence by themselves.

You can switch the Claims view to **Pages** to inspect claims extracted from your own pages. That view helps you compare what you publish with what the AI answers say.

## Read the captured answer

Open the answer behind the finding. Read the full response and the specific claim together. Keep the buyer question and engine in view: a sentence that looks wrong in isolation may be a comparison about a narrower product or plan.

The answer detail links claims and source evidence where the captured response supports that link. A cited page, an uncited related page, and a company-owned page are different signals. If no reliable source can be bound to the claim, leave its source unattributed. Do not guess which page caused the wording.

## Compare it with the Fact Ledger

Open [Facts → Fact Ledger](https://app.knitknot.ai/facts/fact-ledger) and find the relevant current fact. Check the statement, source, date, checkable value, and approval status. A contradiction needs a current, supported fact; a missing fact is a reason to investigate, not proof that the AI answer is false.

For example, if an answer names an old plan price, compare the exact amount and the plan it applies to with the dated pricing fact. A change in price may explain a stale answer, but the captured answer alone cannot show where the stale value came from.

## Decide what to do

If the evidence supports a material problem, open the linked [issue and playbook](/docs/issues-and-playbooks/) to keep the finding, proposed response, and later measurement together. A playbook can point to content your team should create or revise. If the problem is on a page you control, review the owned-page claim before changing it.

After a change ships, compare a later full benchmark with the earlier one. A changed answer is an observation. It does not, on its own, establish that your change caused it.

Docs navigation
Docs › Guides

Investigate a claim in an AI answer

Start with a claim gap, open the captured answer, check the source signal, and compare the wording with your company facts.

Updated

At a glance

  1. Find the topic

    Use Answer claims to spot a gap or a reported misrepresentation.

  2. Read the answer

    Open the captured response and inspect the exact claim in context.

  3. Check the basis

    Review source signals and compare the claim with approved facts.

The Claims view groups what measured AI answers say about your company by buyer topic. Start there when a report says AI is missing a capability or making a damaging claim. The counts point you toward an answer; the answer and its evidence are what you need to read before acting.

Find a topic that needs review

  1. Open Intelligence → Claims and choose Answers.
  2. Use the Claim gaps filter to find topics with too little positive substance or a reported misrepresentation. Low sample means the topic has too few answers for a gap verdict.
  3. Open a topic. The table shows answer counts, positive and negative claims per answer, reported misrepresentations, and competitor claims per answer. Those rates help you choose where to investigate; they do not identify a false sentence by themselves.

You can switch the Claims view to Pages to inspect claims extracted from your own pages. That view helps you compare what you publish with what the AI answers say.

Read the captured answer

Open the answer behind the finding. Read the full response and the specific claim together. Keep the buyer question and engine in view: a sentence that looks wrong in isolation may be a comparison about a narrower product or plan.

The answer detail links claims and source evidence where the captured response supports that link. A cited page, an uncited related page, and a company-owned page are different signals. If no reliable source can be bound to the claim, leave its source unattributed. Do not guess which page caused the wording.

Compare it with the Fact Ledger

Open Facts → Fact Ledger and find the relevant current fact. Check the statement, source, date, checkable value, and approval status. A contradiction needs a current, supported fact; a missing fact is a reason to investigate, not proof that the AI answer is false.

For example, if an answer names an old plan price, compare the exact amount and the plan it applies to with the dated pricing fact. A change in price may explain a stale answer, but the captured answer alone cannot show where the stale value came from.

Decide what to do

If the evidence supports a material problem, open the linked issue and playbook to keep the finding, proposed response, and later measurement together. A playbook can point to content your team should create or revise. If the problem is on a page you control, review the owned-page claim before changing it.

After a change ships, compare a later full benchmark with the earlier one. A changed answer is an observation. It does not, on its own, establish that your change caused it.