Verify
Pick the approved fact that settles the claim, or add one with its source, then choose AI is wrong, Outdated, Ambiguous or AI is right. “AI is right” closes the case with the answer as proof.
PRODUCT DOCUMENTATION · METHOD 2026.09.1
How HowAITrust works, step by step.
Everything a brand, legal or web team needs to catch a wrong AI answer, correct it at the source and prove it changed.
Getting started
- Open the workspace. Sign in at trust.howaicite.com/workspace. One HowAI account opens HowAICite, HowAITrust, HowAISafe and HowAIAds.
- Add your approved facts. The prices, plans, policies and claims your company can back with a public page. Everything else compares against these.
- Add watched questions (Pilot and Team), such as “How much does Acme cost?”. HowAITrust asks them for you. On the free workspace, record an answer yourself with Record AI claim.
- Work the queue. Each case shows its step and the one thing to do next.
Approved facts
A fact is one statement, its claim type (identity, product, pricing, availability, policy, leadership, safety or other) and the public HTTPS page that states it. Add an effective date when a fact starts on a known day, and retire a fact when it stops being true: retired facts stay on the cases that used them but are no longer used for detection. Reviewers, admins and owners manage facts.
Watched questions and suspects
Watched questions are checked on a schedule and use your workspace's AI answer allowance. Each new answer is compared with your current facts:
- With the AI model configured, detection reads the answer against every current fact. A finding is kept only if it quotes a sentence that is really in the answer and points to one of your facts.
- Without it, a rule check catches prices in the answer that differ from your pricing facts.
A finding becomes a suspect in the workspace with the sentence, the fact and the reason. Confirm it to open a case (you choose the potential impact), or dismiss it. Answers HowAICite collects also arrive in the evidence inbox, ready to become a case.
Cases and priority
A case holds one claim from one answer: brand, website, question, AI platform, market, the answer text, up to eight cited URLs and when it was seen. Priority is a review order, not a verdict:
| Potential impact | Limited 0 · customer confusion 1 · commercial 2 · legal or reputation 3 |
|---|---|
| Claim type | Safety +2 · policy or leadership +1 |
| Repetition | Seen in more than one answer +1 |
| No source | The answer showed no supporting URL +1 |
| Result | 6 or more critical · 4–5 high · 2–3 moderate · 0–1 low |
The four steps
Every open case sits in one step. The case page shows it with a progress bar and a Next card.
Correct
HowAITrust builds the correction plan: fix your own pages that repeat the claim, ask other sites to update (email drafted from your fact, with the contact found on their page), publish the fact with structured data, and report the answer to the AI. Mark each step To do, Sent, Done or Declined, or add your own.
Recheck
When no step is left To do, the case moves to recheck. For cases from watched questions, HowAITrust asks the AI the same question now and again after 7 and 21 days, and says whether the answer still contradicts your fact. For cases recorded by hand, paste the new answer.
Resolved
When the answer no longer contradicts your fact, Resolve with this proof closes the case with the steps that were done and the answer before and after. Declined steps are never listed as done. Reviewers can also close a case without a correction, with a reason, and reopen any case.
Source tracing
Check sources reads each page the AI cited and records what it says about the claim. New findings add steps to the correction plan.
| Finding | What it means |
|---|---|
| Repeats the wrong claim | The page states the claim the AI repeated. Your page: fix it. Another site: ask them to update it. |
| States both versions | The page has the old and the new version. Remove the outdated passage. |
| States the correct fact | The page already agrees with your fact. No action on this page. |
| Mentions you, not the claim | The page names your brand but not this claim. |
| Does not mention you | The page does not name your brand. |
| Could not be read | The page did not load or blocked reading. The error is shown. |
Roles and teamwork
| Role | Can do |
|---|---|
| Owner / Admin | Everything, including approved facts, verdicts, resolving and closing cases |
| Reviewer | Approved facts, verdicts, resolving, closing and reopening cases, plus the correction plan |
| Contributor | Recording cases, confirming suspects, owners and due dates, the correction plan, rechecks and comments |
| Viewer | Reads cases, facts and reports |
Give a case an owner and a due date. Comments and @mentions sit in the case activity next to every recorded change. Owners and mentioned teammates get a notice in the workspace, and an email if they turned on mission and case email in Settings. Cases without a verdict after 14 days, and overdue cases, trigger a weekly reminder.
Reports and export
The workspace report summarises your cases, verdicts and resolved corrections with their recheck evidence. Export every case as CSV or JSON from the workspace header. When an AI system your company runs repeats a claim, Add to AI governance turns the case into a HowAISafe control for that system.
Limits
- Case capacity and watched questions depend on your plan; the workspace shows how many cases you have used.
- Automatic checks and rechecks use your workspace's AI answer allowance.
- AI answers vary. One corrected answer shows the change on that question, platform and market, not everywhere.
- Do not paste secrets, personal data or privileged legal material into answers, facts or comments.