Intelligence Scores

Overview

Intelligence Scores are six proprietary metrics developed exclusively for AEO Optima that give you a comprehensive, quantitative view of how AI engines perceive, represent, and position your brand. While standard analytics tell you whether your brand appears in AI responses, Intelligence Scores tell you how well your brand is represented — measuring narrative accuracy, cross-platform agreement, competitive positioning, sentiment stability, citation strength, and topical authority.

These scores are computed from your existing snapshot data. Each score runs on a 0–100 scale, is updated every time you visit the Analytics page, and is displayed alongside a confidence indicator based on sample size. Together, the six scores form a radar chart that visualizes your brand's overall AI health at a glance.

Why This Framework Is Novel

Traditional SEO tools measure rankings, traffic, and keyword positions — metrics designed for search engine results pages. AI answer engines don't have pages, rankings, or click-through rates in the traditional sense. They generate free-form text responses that mention, describe, compare, and cite brands in entirely new ways.

AEO Optima's Intelligence Scores are the first framework purpose-built for this new reality:

Key techniques used:

• Jensen-Shannon Divergence (JSD) — information-theoretic measure for cross-model consistency
• Shannon Entropy — measures competitive landscape fragmentation per prompt
• Kullback-Leibler Divergence — tracks vocabulary and narrative shifts between time periods
• PageRank-inspired influence scoring — weights citation sources by authority and depth
• Multi-dimensional topic authority — evaluates breadth, consistency, depth, and citation rate across 10 topic categories

No other AEO platform, SEO tool, or brand monitoring service provides these specific metrics. They are computed entirely from your snapshot data — no external data sources or third-party APIs required.

The Six Scores

BNCI — Brand Narrative Coherence Index

What it measures: How consistently AI platforms describe your brand's core attributes — value propositions, differentiators, and key facts.

How it's calculated: BNCI analyzes your most recent brand-mentioning snapshots against your Brand Facts and produces a weighted composite of four sub-dimensions:

Value proposition and differentiator scoring uses natural-language inference (NLI) — the same standard formulation as peer-reviewed fact-verification benchmarks (FEVER, MultiNLI, ANLI). For each (response, brand_fact) pair in the value-prop / differentiator subsets, an LLM judge classifies the response's stance as:

• Supported — the response makes a statement about your brand that ENTAILS the claim (not just vocabulary overlap — real semantic backing).
• Contradicted — the response makes a statement that DENIES the claim (asserts an incompatible attribute, or directly negates it).
• Unaddressed — the response does not engage with the claim either way.

The dimension score is 100 × max(0, supported − contradicted) / (supported + contradicted). Unaddressed pairs are excluded from the denominator — they carry no signal either way; a fact with zero mentions is a "missing fact" (attribute-coverage sub-dimension), not a "wrong fact". This means a response that says "their customer support has poor response times" correctly CONTRADICTS the value prop "industry-leading 24/7 customer support" even though every keyword overlaps — something the previous regex-overlap engine would have scored as a positive match.

Attribute coverage and competitor contamination remain rule-based: attribute coverage tracks how many of your brand facts have ANY mention (fuzzy) and FULL-PHRASE mention (exact); contamination detects when competitor names appear in attribution-verb contexts about your brand ("Acme is known for X" where X is a competitor's attribute).

Score interpretation:

How to improve it:

• Add comprehensive Brand Facts in Settings — the more facts you provide, the more accurately the system can measure coverage.
• Publish clear, structured content on your website that states your value propositions explicitly.
• Use schema markup (Organization, FAQ, HowTo) to make facts machine-readable.
• Review the "Missing Facts" list in the score detail to identify which attributes AI platforms are not picking up.

Tip: BNCI requires Brand Facts to be configured. If you see a score of 0, go to Settings and add your brand facts first.

CMCS — Cross-Model Consistency Score

What it measures: Whether different AI platforms (ChatGPT, Claude, Gemini, Perplexity, DeepSeek, etc.) agree about your brand.

How it's calculated: CMCS treats cross-model agreement as a formal inter-rater reliability problem — the same framework used to measure whether human coders agree. Each prompt is a unit, each AI model is a rater, and the rating is the brand sentiment in that model's answer (Negative → Neutral → Positive, an ordered scale). Only answers that actually mention your brand count; a model that doesn't mention you simply gave no rating for that prompt (not a disagreement).

The headline is the chance-corrected coefficient (revised 2026-05-30). We report three figures, leading with the rigorous one:

1. Consistency score (headline, 0–100) = max(0, Krippendorff's α) × 100 — the chance-corrected agreement. α answers the question that matters: do the models agree more than they would by luck? This is essential because brand sentiment is usually overwhelmingly positive — so two models both saying "positive" agree most of the time for trivial reasons. The chance correction discounts that prevalent-category agreement. α near 0 (→ headline near 0) means "almost all the apparent agreement is what you'd expect by chance; there's little independent consensus." α is the standard, citable inter-rater coefficient (Hayes & Krippendorff 2007); leading with raw percent agreement under skewed sentiment would overstate consistency (the high-agreement / low-kappa paradox, Feinstein & Cicchetti 1990).
2. Raw match % (context) — how often two models that both rated a prompt literally assigned the same sentiment. Shown alongside the headline, never as the headline.
3. Bootstrap 95% confidence interval for α — 1,000 resamples of the rated prompts, so the precision of the figure is honest: on a handful of comparable prompts the interval is wide, which is the truth, not a falsely exact number.

Consistency score (headline) = max(0, α) × 100
α                            = ordinal Krippendorff's alpha over prompt × model sentiment ratings
                               (one rating per model per prompt — repeat captures collapsed to that
                                model's modal sentiment, so no model is double-counted as a rater)
95% CI                       = 2.5th–97.5th percentile of α over 1,000 bootstrap resamples

If too few prompts have been rated by 2 or more models, CMCS reports "insufficient cross-model data" rather than a number computed from too little signal. The system also identifies the most divergent platform (the outlier) and the most consistent platform (closest to consensus).

Score interpretation:

How to improve it:

• Investigate the "most divergent" platform and check what it says differently.
• Ensure your content is accessible to all major AI crawlers (check your robots.txt via Crawler Intelligence).
• Publish consistent messaging across all channels — inconsistent source material leads to inconsistent AI output.
• Capture snapshots from at least 3 different AI providers for meaningful comparison.

Tip: CMCS requires snapshots from at least 2 different AI providers. If you only capture from one model, this score cannot be computed.

MEI — Market Entropy Index

What it measures: How fragmented or concentrated the competitive landscape is within AI responses to your prompts. In other words: does one brand dominate AI answers, or do many brands share the space equally?

How it's calculated: For each prompt, MEI builds a mention-share distribution over the brands actually mentioned (your brand + the competitors AI named) and computes Shannon Entropy. There is no synthetic "other / unmentioned" bucket — counting unmentioned snapshots as if they were a brand mixed two different units and distorted the index, so it was removed. The entropy is then normalized against the fixed size of your competitive field — your brand plus every distinct competitor AI surfaced anywhere in the window — and scaled to 0–100. This fixed-field normalization is what makes the number mean fragmentation, not just evenness: a two-player duopoly now scores well below ten equally-mentioned players, instead of both pinning at 100. The overall MEI is the average entropy across all analyzed prompts.

Each prompt receives a market assessment:

Entropy alone doesn't say who dominates. A low-entropy "Locked" market is excellent if you are the brand that dominates it, and bad if a competitor does. MEI therefore reads two companion signals alongside entropy when wording its assessment: mention share (your average share of mentions across analyzed prompts) and brand-is-top (the fraction of prompts where you hold the largest share). That's why a locked market resolves to either "Brand dominates AI answers — strong position" or "Market locked by a competitor — hard to break in," not a single entropy-only verdict.

Important: MEI uses inverted coloring in the UI. A lower MEI is better for your competitive position (it means you dominate). In the radar chart, MEI is displayed as 100 - raw score so that higher = more favorable.

HHI companion (industry-standard concentration metric). Alongside the entropy number, MEI now reports the Herfindahl-Hirschman Index — the same concentration metric the US Department of Justice and FTC use in horizontal-merger review. HHI is the sum of squared market shares (on a 0–10000 scale): HHI = Σ (share_i × 100)². We show it because entropy and HHI capture different shapes of the same distribution:

• Entropy is sensitive to the long tail — adding many small competitors raises entropy a lot.
• HHI is sensitive to top-share dominance — the leader's share is squared, so a single dominant brand drives HHI up regardless of how many small players are below.

Two markets can have similar entropy but very different HHIs, and vice versa. Reading both side by side characterizes the competitive landscape more honestly than either alone. Concentration buckets (DOJ 2010 Horizontal Merger Guidelines applied to the 0–10000 scale):

The HHI bands above are the literal 2010 Horizontal Merger Guidelines thresholds, applied to the same 0–10,000 scale the DOJ uses.

Score interpretation:

• Low MEI (0–30): Your brand dominates AI responses for these prompts. Maintain your position through continued content investment.
• Mid MEI (40–70): This is the "sweet spot" for opportunity. There is room to grow your share without the market being impossibly fragmented.
• High MEI (70+): The market is highly fragmented. Focus on differentiating content and building topical authority to stand out.

How to improve it (lower your MEI):

• Focus content strategy on prompts currently in the "Open" (40–70) range — these are your best growth opportunities.
• Avoid wasting resources on "Locked" prompts where a competitor dominates unless you have a specific strategy.
• The MEI Bubble Chart on the Analytics page visualizes each prompt's entropy, making it easy to identify opportunity zones.

SDI — Sentiment Drift Index

What it measures: How much the AI narrative about your brand is shifting over time. It tracks changes in word frequency and sentiment between consecutive time periods.

How it's calculated: SDI divides your snapshot history into time periods and, between each consecutive pair, separates two distinct questions: how much the narrative shifted (magnitude) and whether the shift is real (statistical significance).

1. Word frequency analysis — Extracts the top 50 meaningful words from brand-mentioning sentences in each period, filtering out stop words and low-frequency terms.
2. Magnitude (the SDI score) — Computes the Jensen-Shannon Divergence (JSD) between the two periods' word-frequency distributions. JSD is symmetric and bounded in [0, 1]; its square root is a true distance metric. The score is round(100 × √JSD), so 0 means an identical narrative and a higher score means the vocabulary moved more.
3. Significance — A permutation test (1,000 reshuffles of the period labels, with a fixed seed so the result is reproducible) estimates a p-value: the probability the observed shift could have arisen by chance. A shift is only flagged as "drift" when it is both sizeable and statistically significant (p < 0.05). This prevents small samples from raising false alarms.
4. Sentiment shift — Tracks how the average sentiment score changes between periods, used to label drift as positive or negative.

The period grouping auto-adapts to your selected window:

• Window < 60 days → weekly periods. A 30-day view yields ~4 weekly groups, enough for meaningful drift analysis even on short ranges.
• Window ≥ 60 days → monthly periods. Preserves the multi-month narrative signal that's the natural cadence for slower brand-perception shifts.

SDI (magnitude) = round(100 × √JSD(previous period, current period))
significance    = permutation-test p-value (separate "significant" flag)

The overall SDI is the average magnitude across all period pairs that had enough data to measure. The system also identifies rising words (terms gaining frequency) and declining words (terms losing frequency), giving you concrete visibility into how the narrative is changing.

The Sentiment Pulse reading (in the intelligence summary) combines the drift index with the sentiment direction to give a clearer headline: "Major shift · positive" when vocabulary changed dramatically toward a more favorable framing, "Stable" when neither shifted, and so on. The numbers behind the headline (drift index 0–100 + average sentiment shift in points) appear in the subline.

Important: Like MEI, SDI uses inverted coloring in the UI. A lower SDI is better (stable reputation). In the radar chart, SDI is displayed as 100 - raw score.

Score interpretation:

Bands apply only when the shift is also statistically significant; an insignificant magnitude is reported as stable regardless of size.

How to improve it (lower your SDI):

• Publish consistent messaging over time — avoid frequent brand repositioning.
• Monitor the "Rising Words" and "Declining Words" lists to understand what is changing.
• If you intentionally changed your positioning, a temporary SDI spike is expected and healthy.
• SDI requires at least 2 periods of snapshot data and 10+ brand-mentioning snapshots to produce meaningful results. With weekly grouping that means ~14 days of capture; with monthly grouping it means snapshots that span at least two calendar months.

Tip: A stable (low) SDI is generally good, but some drift is natural. A sudden spike may indicate a significant event worth investigating.

CIPS — Citation Impact & Positioning Score

What it measures: Your brand's share of all citations in AI responses, plus the influence and authority of sources citing you.

How it's calculated: CIPS analyzes all citations (URLs) found in your snapshots, builds a citation graph, and produces a 0–100 composite of three components:

CIPS = citationComponent (0–40) + hubComponent (0–30) + diversityComponent (0–30)

1. Citation component (0–40) — Based on your brand citation rate (the share of all citations that point to your own website — summed across all your brand-owned domains, not just one), shaped by an exponential-saturation curve: 40 · (1 − e^(−rate / 0.10)). The curve is smooth, monotone, and asymptotic to 40 — no hard cliff. Reference values: 5% rate → 16 pts, 10% → 25 pts, 20% → 35 pts, 30% → 38 pts, 50%+ → 40 pts. Brand-owned citations across the open web are naturally a minority, so a 10% rate is genuinely strong (~63% of max), but a 30%-rate project still scores higher than a 10% one — which a flat cap at 10% would hide.
2. Hub component (0–30) — Your brand-owned sources' share of the total citation-hub influence in your market (revised 2026-05-30 from "average brand-hub influence ÷ the single most-influential hub", which collapsed the component whenever one large third-party hub dominated, regardless of your real authority). A source's influence is citation_count × source_type_authority × hub_bonus, where the authority multiplier rewards trusted source types (academic 1.5×, earned media 1.3×, directory 1.1×, brand-owned 1.0×, social 0.8×, unclassified 1.0× neutral).
3. Diversity component (0–30) — The Shannon entropy of your classified citation source-type mix, normalized to the fixed taxonomy size (log2(5) for the five real source types), computed from raw citation counts, then scaled by a sample-size confidence factor min(1, classifiedCitations / 10). A balanced portfolio across earned media, brand-owned, academic, directory and social sources is more defensible than concentration in one — but only once there are enough citations to make the spread meaningful. Entropy is a ratio, so two citations in two source types is mathematically as "diverse" as two hundred; without the sample-size scaling, a brand-new project with a handful of scattered citations could out-score a mature one (live example: a 16-citation project was scoring higher on diversity than a 69-citation project). The confidence factor reaches full weight at 10+ classified citations and scales down linearly below that. Unmatched domains are classified unknown (not silently lumped into "corporate", as before 2026-05-30 — that made "corporate" an ~85% catch-all that distorted both diversity and authority) and are excluded from this component, so spreading citations into unclassified sources never inflates your diversity. The share of unclassified citations is tracked separately as a data-quality signal.

A note on the influence formula: CIPS uses citation frequency and source-type authority — it does not claim true cross-domain PageRank, because snapshot citations don't tell us which domains link to which. The hub bonus is an honest, bounded proxy for "frequently-cited = more influential," not a link-graph computation.

Methodology change — 2026-05-28 (M1 audit). The citation component formula was previously a flat-cap linear function min(40, rate × 400), which gave every project at ≥ 10% citation rate the same 40-point ceiling and erased information at the high end (10% and 30% looked identical). It now uses an exponential-saturation curve so that marginal gains keep registering all the way up to the 40-point asymptote. Existing projects in the 10–30% citation-rate band may see a CIPS drop of 2–15 points with no underlying behavior change. That drop is the math being more honest, not the score getting worse — a project at 30% citation rate is still genuinely stronger than one at 10%, and now the score reflects that.

The system also identifies:

• Citation hubs — Sources cited 5+ times (high-impact, authoritative sources)
• Citation gaps — Sources that cite competitors but not you, ranked by priority (competitorMentionCount × sourceAuthority × recencyBoost)

Score interpretation:

Note: The three components reward different strengths — you can score well through a high brand-citation rate, influential hubs, or a diverse source mix. The same composite number is shown identically on the dashboard, Findings & actions, generated reports, and via the API, so every surface agrees on what your CIPS is.

How to improve it:

• Create original, authoritative content that AI models cite as a source.
• Review the citation gaps list — these are sources citing competitors but not you. Pursue coverage on those platforms.
• Invest in earned media (press, industry publications) — these carry higher authority multipliers.
• Ensure your content is crawlable and well-structured so AI models can cite specific pages.

ETAS — Entity & Topical Authority Score

What it measures: How strongly AI models recognize your brand as an authority across specific topics relevant to YOUR brand.

How it's calculated (M3 Stage 4, 2026-05-29): ETAS is now a two-stage LLM judge process:

Stage 1 — Topic discovery (per project, refreshed every 30 days). An LLM reads a stratified sample of your brand-mentioning responses and identifies the 10 most-discussed topics SPECIFIC to your brand. So a hospitality brand gets topics like "guest amenities", "location & beaches", "family-friendliness"; a SaaS brand gets "API design", "pricing", "integrations"; a marketing agency gets "campaign results", "industry expertise", "client portfolio". The topic list reflects what AI actually says about your brand, not a generic 2024 category list.

Stage 2 — Per-response authority verdicts. For each (response, discovered topic) pair, the judge classifies the brand's stance as:

Topic authority score = breadth × 0.30 + consistency × 0.30 + depth × 0.20 + citation rate × 0.20, where:

• Breadth = your brand's share of the topic across the full answer population — of all AI answers that discuss this topic (whether or not they mention you), what fraction mention your brand? A brand that genuinely owns a topic shows up in a large share of that topic's answers; a brand that only appears when explicitly named scores low. This is the most important change in the 2026-05-30 methodology revision: breadth used to be measured only within answers that already mentioned your brand, which let a low-visibility brand score deceptively high (a selection bias known as Berkson's paradox) — so ETAS could move against your actual visibility. Breadth now uses the full-population denominator (the SemEval target-aspect "share" formulation), so ETAS tracks real visibility.
• Consistency = how reliably the brand owns the topic across providers and responses.
• Depth = how substantively the brand is discussed when present.
• Citation rate = how often the brand's own sources are cited on the topic (square-root compressed to 0–100, so a 40% and a 58% self-citation rate map to genuinely different scores rather than both saturating at 100).

The judge path and the keyword fallback use the same four-dimension weighting and the same 0–100 scale — so the headline doesn't jump when the judge toggles on or off. The two paths still measure those dimensions differently — the judge reads NLI authority verdicts, the fallback matches keywords — which is why every score carries a method marker (see below); treat a long-run trend as comparable only within the same method.

This is the SemEval target-aspect formulation (Pontiki et al., 2014, 2016) restricted to brand-authority verdicts. Vocabulary overlap is no longer enough on the judge path — a response saying "non-premium pricing" no longer scores authority on "Pricing & Value" just because both words appear. Authority requires the brand to be NAMED with a specific authoritative stance on the topic. Every ETAS (and BNCI) score carries a method marker — LLM-judge vs keyword fallback — so a fallback number produced when the judge is temporarily unavailable is never presented as judge-grade.

The overall ETAS is the volume-weighted average score across topics that have data — each topic is weighted by how often your brand is actually discussed on it, so the headline reflects where your authority genuinely concentrates rather than being swung equally by a rarely-discussed topic.

Topic assessments:

The Analytics page shows the count of strong topics, weak topics, and gap topics (topics your prompts cover but where your brand is never mentioned).

How to improve it:

• Focus content strategy on weak topics and gap topics where you want to build authority.
• Publish in-depth, authoritative content per topic — breadth across sub-topics matters.
• Ensure content is cited by creating original research, data, or frameworks for each topic area.
• Capture snapshots from multiple AI providers to improve the consistency dimension.

Score Comparison at a Glance

Note on the radar chart: On the Score Radar displayed in the Analytics page, MEI and SDI are inverted (100 - raw score) so that all six axes follow the same convention: higher = more favorable for your brand.

Data Requirements

Each score has minimum data thresholds to produce meaningful results:

If a score cannot be computed due to insufficient data, it displays "Insufficient data" instead of a number.

Confidence Indicators

Each score card shows a confidence indicator based on sample size. Larger sample sizes produce more reliable scores. As a general guideline:

• Low confidence — Fewer than 20 snapshots analyzed. Scores may shift significantly as more data is captured.
• Medium confidence — 20–50 snapshots. Scores are directionally reliable.
• High confidence — 50+ snapshots. Scores are stable and trustworthy.

Where to Find Them

Intelligence Scores are displayed on the Analytics page under the "Intelligence Scores" section. Navigate to Dashboard > Analytics and scroll to the score cards.

The section includes:

• Six score cards — One per metric, showing the numeric score, description, abbreviation, and confidence indicator.
• Score Radar — A radar chart plotting all six scores for a visual overview of brand health.
• Cross-Model Consistency Matrix — A heatmap of pairwise JSD values between AI platforms (visible when CMCS data is available).
• Market Entropy Landscape — A bubble chart showing per-prompt entropy, snapshot count, and brand count (visible when MEI data is available).

Plan Requirements

Intelligence Scores are computed from your existing snapshot data and are available on all plans. However, to build up the data needed for meaningful scores, you will benefit from higher-tier plan features:

Related Features