Intelligence Engines

Overview

AEO Optima runs five computation engines that analyze your snapshot data and generate actionable insights. Each engine answers a specific strategic question about your brand's AI visibility — from identifying which citation sources drive your presence, to predicting when a competitor might overtake you, to tracing the root cause of factual errors.

These engines run daily as part of the analytics maintenance cycle. Their outputs power the findings on Findings & actions, the brand dossier on the Entity analysis page, and the Related Insights strip that sits beside the charts on each analysis page. Each engine writes structured insights with a type, severity, confidence score, and evidence payload.

Citation Impact Attribution

Question answered: "Which citation sources are driving your visibility?"

Citation Impact Attribution correlates the appearance of citation domains with per-model visibility changes. When a new domain first appears in AI responses about your brand, this engine measures whether your visibility on that model increased or decreased in the same and following week.

What it produces:

Insight type: citation_impact

An insight is generated when a domain caused more than 10% visibility change on any model. Positive impacts highlight sources worth cultivating; negative impacts flag sources that may be diluting your brand presence.

How to act on it:

• Domains with large positive deltas are your highest-value citation sources. Ensure your content appears on those sites.
• Domains with negative deltas may be introducing competing or inaccurate information. Review them and consider correction submissions.
• Track citation rates over time — a source cited in a high percentage of responses has outsized influence on how AI perceives your brand.

Competitor Trajectory Prediction

Question answered: "Is any competitor about to overtake you on a specific prompt?"

This engine fits a linear trend on the recent weeks of per-prompt mention rates for your brand and each configured competitor. It then extrapolates to predict whether and when a competitor's mention rate will exceed yours.

What it produces:

Insight type: competitor_trajectory

An insight is generated when crossover is predicted within 4 weeks and the competitor's upward slope is statistically significant (two-tailed t-test on the slope, p < 0.05, with at least 4 weeks of data). We gate on significance rather than R² because a high R² on only a handful of noisy weekly points happens easily by chance — it measures how tidy the line is, not whether the trend is real. Confidence reported on the insight is 1 − p. Severity is based on urgency — a crossover predicted in 2 weeks is critical; one further out is high.

How to act on it:

• Focus content efforts on prompts where crossover is imminent (fewer weeks remaining).
• Investigate what the competitor is doing differently for those prompts — check their cited content via Citation Tracking.
• Only significant trends raise alerts; a competitor with a noisy, non-significant rate is not flagged even if a naive line through the last few weeks would point upward.

Prompt Visibility Decomposition

Question answered: "Which prompts are pulling your visibility down, and how much would fixing them help?"

This engine computes per-prompt visibility over the last 7 days, identifies prompts that are significantly below your project-wide average (more than 20 points below), and projects how much your overall visibility score would improve if each weak prompt were brought up to a conservative 50% target.

What it produces:

Insight type: prompt_weakness

An insight is generated when 3 or more prompts are more than 20 points below the project average. The insight includes the total projected lift, giving you a clear picture of what is achievable.

How to act on it:

• Start with the weak prompt that has the highest "lift if fixed" value — that is your biggest opportunity.
• Check which specific models are not mentioning your brand. If most models mention you but one does not, investigate that model's training data and citation preferences.
• Supports prompt segmentation — you can analyze branded, non-branded, and competitor prompts separately.

AI Error Root Cause Tracing

Question answered: "Which citation source is causing AI to state incorrect facts about your brand?"

When AI models produce factual errors about your brand (detected via the accuracy methodology described below, scored against your Brand Facts), this engine traces those errors back to their likely citation sources. It groups inaccurate snapshots by the specific fact that was wrong, then ranks citation domains by how frequently they appear in error-containing responses.

How accuracy is measured (factual precision)

For every captured response that mentions your brand, the platform decomposes the response into atomic factual claims — single, indivisible assertions about your brand that a human could fact-check (e.g. "founded in 2010", "headquartered in Mumbai"). Each claim is then verified independently against the Brand Facts you've configured.

Each claim lands in one of three buckets:

• Supported — corroborated by a specific brand_facts row
• Not supported — contradicted by a brand_facts row, or makes a specific verifiable assertion no fact addresses
• Irrelevant — style, framing, recommendations ("a great choice"), or generic category statements that aren't atomic facts about your brand

The headline score is factual precision:

factual_precision = supported_checkable / total_checkable   ∈ [0, 1]

A response with 9 of 10 brand claims correct scores 0.9; a response with 5 of 10 correct scores 0.5. Responses with no checkable atomic claim about your brand (pure opinion / framing) return null — an honest "not scored" verdict rather than a defaulted "accurate".

This is the FActScore methodology (Min et al., Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation, EMNLP 2023), refined with the SAFE supported / not_supported / irrelevant trichotomy (Wei et al., Long-form Factuality in Large Language Models, 2024). It is the same methodology a peer-reviewed factuality benchmark would use — every number on the accuracy surface is explainable down to the individual claim, the evidence excerpt, and the brand_facts row it was checked against.

The atomic-claim decomposition + verification runs as an LLM-as-judge inline with snapshot capture, alongside the M2a sentiment judge and M2b brand-extraction judge. The accuracy judge runs only when (a) your brand is mentioned in the response AND (b) you have at least one Brand Fact configured. Competitor-only snapshots and projects without brand facts return no precision row — both are honest "no signal" verdicts, not errors. Manual single-snapshot captures (the Capture Now button on a prompt) still use the legacy batched binary path for latency reasons; bulk-captures + scheduled runs use the new atomic-claim path.

What it produces:

Insight type: error_root_cause

How to act on it:

• The "likely source domain" is your primary target. If an outdated article on a third-party site is being cited by AI models, that article is probably the source of the misinformation.
• Use Hallucination Corrections to submit corrections for the identified errors.
• Prioritize errors that affect multiple models — these indicate a widely-cited incorrect source.
• Update your own published content to make the correct fact explicit and easy for AI crawlers to find.

Leading Indicator Detection

Question answered: "Which prompts predict your overall visibility score days in advance?"

This engine performs cross-correlation analysis between each prompt's daily visibility and your project's overall daily visibility, testing time lags from 1 to 14 days. A prompt is a leading indicator if its visibility changes reliably predict your overall score change several days later.

Lags are measured in calendar days, and a prompt's day is paired with the overall score only when both actually captured data that day — days a prompt wasn't run are left out, never filled with a fake 0% (which would manufacture a correlation out of the capture schedule). So "predicts 5 days early" means five real days, not five capture cycles.

What it produces:

Insight type: leading_indicator

Because the engine scans 14 lags across every prompt, raw correlations would throw up coincidences. Two guards prevent that: each lag's significance is computed at the autocorrelation-corrected effective sample size (smooth daily series otherwise look far more significant than they are), and the whole family of prompt×lag tests is held to a Benjamini-Hochberg false-discovery-rate correction. An insight is generated only for a prompt whose best lag survives that correction with a lag of 3 or more days. The minimum data requirement is 21 daily data points (3 weeks of captures); a prompt also needs at least 14 complete paired days at the tested lag.

How to act on it:

• Monitor leading indicator prompts closely — a visibility drop on these prompts signals that your overall score may decline in the coming days.
• Leading indicators are often your most sensitive prompts. Changes in AI behavior appear here first before spreading to other prompts.
• Use these prompts as an early warning system by setting up Alert Rules for visibility drops on them.

How Insights Are Generated

All five engines run as part of the analytics maintenance cron job, which executes daily at 4:30 AM UTC. The process works as follows:

1. The cron job identifies all projects with 20 or more snapshots — projects below this threshold do not have enough data for statistically meaningful analysis.
2. Each engine runs against qualifying projects and produces results.
3. Results that meet each engine's significance thresholds are written to the insights table.
4. Each insight record includes:
   • Type — The engine that generated it (citation_impact, competitor_trajectory, prompt_weakness, error_root_cause, leading_indicator)
   • Severity — critical, high, medium, or low based on the engine's assessment of urgency
   • Confidence — A numeric score reflecting statistical reliability
   • Evidence — A JSONB payload containing the full computation output for that insight
5. Insights auto-expire after 30 days if not acknowledged. This prevents stale intelligence from cluttering your feed.

Tip: The more snapshots you capture and the longer your capture history, the more reliable these engines become. Consistent daily captures across multiple AI models produce the best results.

Where Insights Appear

Intelligence engine outputs are surfaced across multiple parts of the platform:

MCP Tools

If you use the MCP API, four tools provide programmatic access to intelligence engine outputs:

Webhook Events

When a new insight is generated, AEO Optima can notify your systems via webhooks:

The webhook payload includes the insight type, severity, and evidence, allowing you to route high-severity insights to Slack, email, or other notification channels.

Related Features