Oracle

​

Purpose

The Oracle page verifies how accurately AI models are talking about a customer’s brand. It checks AI responses against the customer’s knowledge base to spot inaccuracies, hallucinations, and contradictions. It also allows customers to verify their internal knowledge base facts against external third-party sources. Customers use this page to measure their brand’s AI accuracy over time, identify high-impact AI hallucinations, and resolve discrepancies so they can take action to correct the AI’s understanding.

​

What’s on the page

Page header Shows the title “Oracle”. If a run has been completed, it includes a “Run analysis” button to kick off new checks. What AI is saying about your brand A search box allowing customers to ask questions about their brand across AI models (e.g., “What is AI saying about your pricing, competitors, founders…?”). Submitting a query returns a synthesized summary of AI claims, complete with citation pills (like [1]) and a sentiment badge (“Positive”, “Negative”, “Mixed”, or “Neutral”). Metric cards

• Discrepancy rate: The share of analyzed AI responses that contain at least one claim contradicting the verified facts. Includes a sparkline chart showing the trend over recent runs. Tooltip: “Share of analyzed AI responses with at least one claim that doesn’t match your verified facts. Lower is better.”
• Open discrepancies: The total number of unresolved discrepancies, with a sub-value indicating how many of those are high-impact.

High-impact discrepancies A list of the top unresolved, high-impact discrepancies found in the latest run. Each row shows a severity indicator (“High”, “Medium”, “Low”), the incorrect AI claim, and the ground truth. A “See all” button opens the full review window. Coverage

• Accuracy by model: A chart breaking down AI accuracy across different models (like ChatGPT, Claude, Gemini, Perplexity, DeepSeek, Grok, Copilot, AI Mode, AI Overview). Tooltip: “Click a model to drill into its accuracy over recent runs. Accuracy = 1 − (high-confidence critical wrong claims ÷ total response claims). Variants of the same model (e.g. Claude) are merged.”
• Topics AI gets wrong: A chart showing which topic categories have the most errors or the highest error rate. Tooltip: “Topic categories where AI gets claims wrong. Toggle Count vs % wrong: Count surfaces topics with the most wrong claims overall; % wrong surfaces topics where the error rate is highest regardless of volume. Click a bar to see the actual wrong claims for that topic.”

Check responses dialog Opened when running a new accuracy check. Allows the customer to select a date range to scan. Once a run finishes, this dialog updates to show how many discrepancies were found, offering a “Review discrepancies” button. Verify facts dialog Opened to validate the internal knowledge base against external sources. Shows a multi-select dropdown to choose specific Pillars to verify. Verification sources drawer Appears when reviewing a fact-verification run. Shows a table ranking external domains by the number of contradictions found.

• Domain: The contradicting source’s root domain. Can also display “Knowledge base” for internal inconsistencies, with the tooltip: “Places where your KB topics contradict each other (internal inconsistencies)”.
• Contradictions: The count of contradictions attributed to this domain.

Contradictions drawer Opened by clicking a domain in the Verification sources drawer. Shows a paginated view of individual contradiction cards comparing the Knowledge base claim to the AI response claim. Discrepancy review window A full-screen modal opened by clicking “See all” or clicking a specific discrepancy. It provides a split-pane workspace:

• Left rail: A list of discrepancies, filterable by Unresolved, Resolved, or Dismissed.
• Main panel: Compares “Claim A” and “Claim B”, explains “Why this is flagged”, and provides a “Research verdict” from deep research. Contains options to mark which claim is correct and add a note.

Response Details drawer A slide-out drawer showing the full text of an AI response, its date, the model used, the prompt type, and whether the target was mentioned.

​

What you can do here

• Search AI claims: Type a query into the “What AI is saying about your brand” search box and press Enter. Click the [1] citation pills to view the underlying AI response.
• Run a response check: Click “Run analysis” and select “Check responses”. Choose a date range and click “Start”.
• Verify facts: Click “Run analysis” and select “Verify facts”. Choose the Pillars you want to check and click “Start verification”.
• Review discrepancies: Click “See all” or click on any individual discrepancy row to open the review window.
• Resolve a discrepancy: In the review window, select the correct claim (“Claim A” or “Claim B”). You can also click ”+ Add a note” to save an explanation for your team.
• View response: When viewing a discrepancy or a search citation, click the AI response chip (e.g., “Response 1” or “View response”) to open the full Response Details drawer.
• Open external site: In a discrepancy card or contradiction view, click the domain link or “Open site” to open the cited third-party source in a new tab.
• Cancel a running analysis: When a run is active, a progress overlay appears in the bottom right. Click “Cancel” on this overlay to stop the run.
• Hide progress overlay: Click the ‘x’ icon on the floating progress overlay to dismiss it from your screen without cancelling the background run.
• Review pending claims: In the Verify facts dialog, if there are pillars with unreviewed claims, click the “Review them” or “Review” link to navigate to the Knowledge Base review page.

​

Data shown

The page displays data generated from analyzing the AI responses collected for your brand against the verified facts configured in your Knowledge Base. It highlights mismatches (discrepancies), measures accuracy rates across models, and identifies external domains that contradict your brand’s established facts.

​

Common workflows

Run a brand-accuracy check

1. Click “Run analysis” in the top right and select “Check responses”.
2. Select a date window (like “Last 7 days” or “Last 30 days”) or pick custom dates.
3. Click “Start”. A floating overlay will track the progress.
4. Once completed, a dialog will summarize the findings. Click “Review discrepancies” to start resolving them.

Resolve a discrepancy

1. From the Overview tab, click on a high-impact discrepancy row or click “See all”.
2. Read the comparison between Claim A and Claim B, along with the “Why this is flagged” explanation.
3. Review the “Research verdict” for additional context.
4. Click on the correct claim (or leave the recommended one selected) to resolve it.
5. Click ”+ Add a note” if you want to leave context for your team.

Verify knowledge-base facts against external sources

1. Click “Run analysis” and select “Verify facts”.
2. Use the dropdown to select which Pillars you want to verify.
3. Click “Start verification”.
4. When finished, open the results to see the Verification sources table. Click any domain to review the specific contradictions.

Search what AI says about the brand

1. Locate the “What AI is saying about your brand” search box.
2. Type a specific question (e.g., about pricing or a competitor) and press Enter.
3. Read the AI-generated summary. Click on the numbered citation pills to drill down into the specific AI responses that support the summary.

​

Empty, loading, and error states

• No analysis run yet: If no analysis has ever been run, the page displays a dashed-border empty state with a shield icon, prompting you to “Run first analysis”.
• Running your first analysis: While the very first run is in progress, a spinner appears with text explaining that the process typically takes 5–15 minutes.
• Loading states: While fetching data, the page shows skeleton placeholders mimicking the shapes of the metric cards, charts, and lists.
• Check complete: When an analysis finishes, a dialog appears summarizing how many discrepancies were found.
• Couldn’t verify access: If there is an issue validating your subscription plan, an error screen appears with a “Retry” button.
• Locked feature: If your subscription plan does not include Oracle, a lock screen will appear prompting an upgrade.

​

Linked from / links to

• Linked from: Typically accessed via the main sidebar navigation under “Oracle”.
• Links to:
  • The “Review them” or “Review” links in the Verify facts dialog take you to the Knowledge Base review page to approve pending claims.
  • Links to external domains open third-party websites in a new tab.

​

Common support questions

Why can’t I click “Run analysis”? Starting, cancelling, or resolving an analysis run requires Write access for Analytics. If you only have View access, the button will be disabled, and hovering over it will show a tooltip explaining the required permissions. Why is there a cap on the number of responses scanned? To ensure the analysis completes efficiently, a single run analyzes a maximum of 3,000 responses. If your selected date range contains more than this, the system will sample them evenly across all models. Why can’t I change a discrepancy that says “Both valid” or “Dismissed”? Once a discrepancy is resolved as “Both valid” or is “Dismissed”, it becomes locked to prevent accidental changes. If you need to update it, click the “Edit” button on the card to unlock the choices. What does “Knowledge base” mean in the verification sources list? When running a “Verify facts” check, you might see “Knowledge base” listed as a contradicting source. This indicates an internal inconsistency where two approved facts within your own Knowledge Base contradict each other.