AI Claims Verification

AI Visibility Claims shows factual statements made about your organisation, products and configured competitors in monitored AI answers. Use it to inspect what the AI said, check public evidence, record your own assessment and create follow-up tasks on the Action Board.

Claims uses answers already collected by AI Visibility monitoring. Building the inventory does not ask the monitored platforms new questions, and finding a claim does not mean it has been verified.

Open Claims and build the inventory

You need access to the workspace and project, an active AI Visibility persona configuration, and a settled monitoring batch containing processed answers. New extraction, page retrieval and verification can use workspace credits.

1. Open AI Visibility Claims from the application navigation.
2. Select the Project. Its latest saved inventory loads automatically if one exists.
3. If there is no inventory, select Build inventory. Otherwise, select Refresh inventory to update it from the latest monitoring.
4. Wait for publication, or leave the page and return later.

The previous inventory stays available while a refresh runs. Refresh running means the background work is underway; you do not need to start it again.

Check the organisation, product and competitor names in your project settings. Claims uses them to identify statements about your brand and competitors.

Which answers are included

Claims uses the newest settled monitoring batch for each active persona configuration. A batch is settled when its monitoring jobs are no longer pending or running. It includes the expected non-deleted conversations and configured platforms in those batches.

It does not analyse every historical answer or substitute an older answer when a current one is missing. Unchanged answers can reuse saved extraction. New or changed answers are processed, their occurrences are retained, and related claims are grouped into topics.

If the selected answers have not changed, a refresh can reuse the current inventory without new extraction charges.

Check the coverage

Beneath Project, the page shows how many monitored answers were analysed and when the inventory was updated. This is an answer count, not a count of verified claims.

If some expected answers could not be analysed, their results are excluded. An inventory can publish with partial coverage, but missing observations do not prove a claim disappeared. If no expected answer can be analysed, the previous inventory is not replaced.

Select Dismiss to hide the missing-answer warning for that saved inventory in your browser. This does not process missing answers. The coverage count and Coverage details remain available. A new inventory or changed coverage can show the warning again.

Find your way around

The Competitors tab is hidden when standalone competitor claims are not included. Changing that scope applies from the next refresh.

Not every associated source was a visible citation in the original answer. Background sources can be retained as context without being treated as citations the answer displayed.

Needs attention and All claims

The inbox opens in Needs attention. This includes incorrect or conflicting findings, partly accurate findings containing a material contradiction, incomplete or expired checks, and user overrides needing review.

Select All claims to include accurate, not publicly confirmed and unchecked claims as well. Being unchecked does not, by itself, put a claim in Needs attention. Partly accurate is not automatically an attention item: some details may be unestablished rather than contradicted.

An active user override supplies the working assessment used by these views. The public check remains visible separately when you expand the claim.

Topics and claim counts

Related claims appear under a topic. Expand it to see all checkable claims matching the current view. Opinions, predictions and other statements appear separately and do not increase the checkable claim count.

One topic can contain several distinct claims, and one claim can appear in several answers. Claim counts concern distinct checkable claims; mentions and answer counts describe where they appeared. Repetition is not evidence that a claim is true.

Open rows remain readable while background results arrive. A claim that stops needing attention while you are reading it can stay in view until you close it or deliberately change the view.

Filter the inventory

Use the Result and Check shortcuts, or open More filters for Persona, Conversation, Platform, Change, Result, Check and Overrides.

Result describes the current assessment, such as Accurate or Incorrect. Check describes public verification, such as Checked, Not checked or Incomplete or expired. They answer different questions and their counts should not be added together.

Choosing a result, check, change or override filter opens All claims. Remove a tagged filter with its × button, select an active Result or Check shortcut again to clear it, or select Clear filters. Select Needs attention to return to that view.

Find overridden claims

Overrides has three choices:

• All claims — no restriction based on overrides.
• Active overrides — claims currently using a user’s assessment.
• Overrides needing review — a saved assessment needs reviewing against changed information.

The option counts cover checkable claims in the current project and claim tab, respecting the other filters. They ignore the override choice itself and the Needs attention view so you can see what each option would include. Removed overrides remain in claim history and have no separate filter.

Filters only change the displayed inventory. They do not refresh monitoring or restrict Check eligible claims now, which applies to the whole saved inventory.

Read a claim and its evidence

Expand a claim to inspect its wording, What the AI said, where it appeared, and any saved checking results.

The public check researches the original claim, including its company, product, numbers and qualifications. What the public evidence shows explains meaningful findings as Supported, Not established, Contradicted or Conflicting evidence, where those details were retained.

For example, a claim might say “The product provides predictive and generative analytics”. If evidence supports predictive analytics but does not establish generative analytics, the finding can be Partly accurate. The narrower supported finding explains the original claim; it is not a replacement claim sent for a separate check.

The verdict and explanation are saved together. There is no additional paid step solely to fill these explanation boxes. Older valid results keep their saved explanation; they are not automatically rechecked to add newer detail.

Public-evidence verdicts

Not publicly confirmed does not mean false. Missing evidence is not the same as contradicting evidence.

Evidence links and quotations show the material used in the finding. A consulted page was opened during the check; consultation alone does not establish support. Source domains can include your own website and competitor websites. It is not a count of independent third-party publishers.

Public check states

No verdict means no completed verdict is available. It is not an Incorrect result. A failed attempt also differs from a completed Not publicly confirmed finding.

The AI answer’s own sources

When Check the AI’s own cited pages is enabled, verification separately assesses whether the visible pages cited by the original answer support its claim.

It uses visible citations rather than background provider sources. Results can show support, partial support, no support or Answer sources not assessable. A claim can be accurate even when the original answer cited nothing that supports it.

Answer sources not assessable means there was no usable visible citation or its content could not be assessed reliably. It does not make the claim false or replace the public-evidence verdict.

Check and recheck claims

Check a selected group

1. Select the checkboxes beside the eligible claims you want to check, up to 25 at a time.
2. Select the check button. Check selected claims becomes a count such as Check 3 claims after selection.
3. Watch progress or return later.
4. Expand the claims to read their results and evidence.

This action reuses valid saved results where possible and performs new work when needed. While checks are running, all manual claim selection is paused. You can select another group when they finish.

Request a fresh check

When your selection includes a claim with a saved public result, a separate button appears, such as Recheck 3 claims — uses credits.

Use Recheck when you deliberately want fresh public verification instead of reusing a valid verdict. It applies to the selected group. The previous verdict remains available until a reliable replacement is saved; a failed recheck does not turn it into a new finding.

Check the whole saved inventory

Check eligible claims now appears when the project has a saved inventory and relevant checking or evidence-monitoring settings are enabled.

With Check new and changed claims enabled, it checks eligible incomplete, expired and unchecked claims across the whole saved inventory. Valid current results are reused. It does not require another refresh and is not limited to filtered rows. A deliberate new click can retry an earlier incomplete attempt.

If only Watch saved evidence is enabled, the button follows that monitoring scope; it does not enrol every unchecked claim. You can check selected claims while automation is off.

Understand completion

A process can finish with incomplete results. The message reports how many claims were fully checked and identifies remaining incomplete work. Requested answer-source assessment also affects completion.

You can Dismiss the completion message; it also clears after a short period in that browser session. Individual incomplete or expired states remain until resolved. Dismissing the message does not restart work or change a verdict.

Record your own assessment

Use an override when you want the project’s working assessment to reflect your judgement. It is shared with people who can access the project and remains visibly separate from the public-evidence check.

1. Expand an eligible claim and select Override result.
2. Choose Your assessment from the five result options.
3. Enter a Reason explaining your decision.
4. For Partly accurate, answer What makes it partly accurate?: Some detail is incorrect or Some detail is not established.
5. Optionally add a Supporting URL (optional), then select Save assessment.

Saving records your assessment, explanation and attribution. It does not run a public check, fetch the supporting URL or use credits. It does not change the original AI statement or erase the public result.

An active assessment controls current result badges, result filters, counts and Needs attention. For example, your Accurate assessment can remove a claim from Needs attention, while an incomplete public check remains visible separately. Public verification and automatic checking continue independently.

Select Edit assessment to change your decision. Use public result restores the public result as the working assessment and retains the manual decision in Recent assessment history.

What happens when the claim is updated

For a checkable claim, an override needing review appears in Needs attention and under Overrides needing review. Select Review assessment to compare the previous decision with the current claim and public result. Choose Keep this assessment if it still applies, edit and Save assessment, or choose Use public result.

For a materially changed claim, a new assessment applies to that claim; the earlier decision stays with its earlier context. If several previous assessments might apply, review them and make a fresh decision.

If the claim or another person’s assessment changes while your form is open, reload the saved context and review it before saving. Your typed explanation is retained for review. If a save response is uncertain, Retry same request confirms the original save without deliberately creating another assessment.

Create a follow-up on the Action Board

A follow-up is a task you write, with the saved claim attached as context. You can create one even without a completed public verdict.

1. Expand the statement and select Create follow-up.
2. Enter a Task title and What needs to be done.
3. Review the attached claim context.
4. Select Add to Board.
5. Select View task to open it on the Action Board.

The task starts in To do. Wheremore does not generate its title or action for you, run another claim check or use credits to create it.

Follow-ups opens tasks already linked to that claim, so you can inspect their status and open an existing task before creating another.

The task retains the claim, available public findings and evidence, answer examples and any manual assessment shown when its source was prepared. Later rechecks or overrides do not rewrite that source. View claim in Claims returns from the Board to the current claim where available; the original task context remains available if the claim changes or disappears from the latest inventory.

See the Action Board guide for managing the task after creation.

Set up automation and claim scope

Select Settings to open Claims settings for the project, then Save to apply your choices. Build the first inventory before enabling automation. Saving settings does not start checking; use Check eligible claims now for the current inventory.

The first inventory establishes a baseline; it does not automatically verify every claim. A refresh that reuses the existing inventory does not start a new automatic verification cycle.

An expired result or inaccessible source does not itself mean the AI changed its claim. If Watch saved evidence finds the saved quotation still present, that does not trigger fresh AI verification. A timeout, access block or incomplete capture leaves the prior verdict unchanged and can be retried on a later monitoring cycle. An inconclusive check does not renew an expired result.

Watch saved evidence is off by default and independent of Refresh weekly and Check new and changed claims. Enabling either of the other settings does not enable it.

Older project settings may show Partially on for Check new and changed claims, because only one of those checks was enabled previously. Leaving the control untouched preserves that choice. Changing it enables or disables both together.

To stop scheduled inventory refreshes and evidence monitoring, turn off Refresh weekly and Watch saved evidence. To stop automatic checks after later manual refreshes too, turn off Check new and changed claims. Work already running may finish.

Review changes and trends

The Change filter can show New claim, The AI changed its claim, Returned, Platforms differ, Verdict changed, Evidence unavailable and Evidence changed.

Platforms differ means platforms currently make materially different assertions; it is not necessarily a change since the previous inventory. Verdict changed concerns public checking. Incomplete answer coverage prevents missing observations from being treated as proof of disappearance.

In Trends, choose Last 12 reports, Last 26 reports or Last 52 reports, then Count or Percentage. These are saved reports, not calendar-day ranges. The charts distinguish claims found, public-evidence results and support from the answers’ own sources. Open available chart details to inspect a saved report.

Manual overrides affect the current inbox assessment. They do not rewrite historical public-check results or count as completed public verification in Trends.

Receive completion emails

The Email notifications page has two Claims preferences:

• Claims Inventory Updates — a new or refreshed inventory is published, including any partial coverage.
• Claims Verification Completed — an eligible verification process finishes with completed results available, including any incomplete checks within that process.

Verification emails cover a manually selected group of at least two claims or the whole automatic or whole-inventory checking process. They are not separate emails for each claim or intermediate batch. Individual one-claim manual checks, empty or stopped processes, and processes with no fully completed results do not send a verification-completed email.

An inventory email does not mean verification is complete, and a verification email does not mean every claim is accurate. Open Claims shows the notification’s run summary and the current inventory, which may contain later updates.

Email preferences control delivery, not automation or whether work runs. See the Email notifications guide for managing them.

Saved work and credits

Return to the project to load its latest published inventory. Refreshing does not erase public verification or manual assessment history. Saved results can be reused for equivalent current claims while valid. Reloading during background checking reconnects to saved work.

Credits can be used for new extraction, AI grouping, uncached page retrieval, public verification, evidence recovery and answer-source assessment. A bounded additional evidence check can use credits when an initial finding needs corroboration or a usable source passage.

Filtering, reading saved results, recording an override and creating a user-written follow-up do not themselves run paid AI work. Recheck explicitly requests a fresh public check and uses credits.

Claims does not have its own Analyse with AI recommendations report. Supported AI Visibility conversation, segment and cross-persona analyses can use Claims context. A manual assessment is supplied as attributed project context, separately from public evidence.

If something does not look right

No claims appear

Select All claims and Clear filters, then check the project and inventory coverage. An empty Needs attention view does not mean the inventory contains no claims. If a project is unavailable, check your workspace, project access and active AI Visibility persona configurations.

Check eligible claims now does nothing

There may be no eligible work after valid results are reused, or only evidence monitoring may be enabled. Check Settings and the result states. Use Recheck on selected claims if you deliberately want fresh verification. Do not refresh the inventory solely to repeat that check.

Check didn’t finish

The check could not save a reliable final result. A provider request may have failed, required evidence could not be retained reliably, or the final assessment could not be validated. It is not proof the claim is false.

Select the affected claim. If a current saved verdict remains, use Recheck to request a fresh check. Otherwise, use Check selected claims. New work can use credits. If the problem continues, contact support with the project, claim and approximate time. Any previous saved verdict remains available.

The process finished but some claims are incomplete

The process has finished, but not every selected claim has a fully completed check. Inspect individual results and retry those that need it. Dismissing the completion message only clears the notice.

My override no longer supplies the result

Look for Manual assessment · needs review and use Review assessment. Changed claim meaning or public evidence can require a new decision. If manual assessments could not be loaded, reload before editing; the displayed public result does not prove the override was deleted.

A source status is unknown

An access block or timeout is not a confirmed broken page. A page can also be available without supporting the claim. Read availability information and the evidence assessment separately.

An older result has less detail

It retains its original explanation and may still be reusable. It is not automatically rechecked to populate newer evidence details. Request a fresh Recheck only when you want new verification, with the associated credit use.

Limits

Extraction and grouping use AI and can occasionally split similar claims or group them too broadly. The inventory represents your monitored scope, not every AI answer or the whole market. Public searching cannot prove that no evidence exists anywhere, and pages can change after a check.

A user override records a project judgement rather than independent public verification. Neither a verdict, an override nor a follow-up task guarantees a later AI answer will change.