Tracking Brand Visibility in ChatGPT, Claude, and Perplexity: The Math & Metrics of AEO

Google Analytics, PostHog, Mixpanel, and Semrush have powered digital marketing and growth engineering for fifteen years. They track pageviews, session duration, referrers, bounce rates, and organic keyword positions.

Yet, when an engineering leader spends 45 minutes inside ChatGPT architecture-planning an enterprise migration, your entire traditional analytics stack registers absolutely zero events.

No pageview is generated. No referral header is sent. No UTM parameter is captured. The buyer evaluates your software, compares it against two rivals, forms a purchasing conclusion, and moves forward—entirely within the walled garden of a conversational neural network.

This is the Dark Funnel of Generative AI.

To understand and influence this funnel, we have to invent an entirely new analytical discipline. Here is how we formulated the core metrics of Answer Engine Optimization at SeerSight.

---

The Statistical Dilemma: Probabilistic Outputs vs. Deterministic Analytics

In traditional SEO, if you rank #3 on Google for "vector search postgres", you rank #3 for almost every user in that geography. The index is deterministic and updates slowly.

Large Language Models do not work this way. LLM generation is inherently non-deterministic. When sampling tokens at temperature $\tau > 0$, the probability distribution of token generation creates structural variance across runs:

$P(w_t \mid w_{<t}) = \frac{\exp(z_t / \tau)}{\sum_j \exp(z_j / \tau)}$

Because token sequences diverge, a model might surface your brand in one generation, relegate it to an afterthought in the next, or replace it with a competitor in a third.

To extract reliable business intelligence from this stochastic environment, an AEO platform must:

1. Conduct repeated automated batch evaluations across defined prompt sets.
2. Filter out structurally invalid prompts where metrics carry zero informative signal.
3. Quantify presence, ranking, attention share, and sentiment using rigorous mathematical formulations.

---

The Four Core AEO Metrics

1. Visibility Score (Mention Rate)

Visibility represents the empirical probability that an AI model will surface your brand when answering a non-branded, category-defining prompt.

$\text{Visibility Score} = \left( \frac{N_{\text{present}}}{N_{\text{valid}}} \right) \times 100$

Where:

• $N_{\text{present}}$ is the count of successful evaluation runs where your brand entity was explicitly matched.
• $N_{\text{valid}}$ is the total number of valid evaluations executed for that prompt.

The “Pre-determined Question” Trap (Validity Filtering)

In SeerSight’s backend evaluation engine, we enforce strict validity rules (is_visibility_valid):

• If the user prompt specifically names your brand (e.g., “How does Acme Cloud work?”), the model’s mention of your brand is a foregone conclusion. Marking this as a “win” creates false-positive vanity metrics. In SeerSight, such runs are flagged is_visibility_valid = false.
• If the prompt is generic and unbranded (e.g., “What are the best cloud storage providers for HIPAA compliance?”), the prompt tests true organic retrieval. The evaluation is marked is_visibility_valid = true.

---

2. Rank Score within the Competitive Set

Does the model recommend you first, second, or as a passing footnote?

When an AI engine generates a comparative response, it typically structures the answer as an ordered list or assigns prominence via hierarchy. SeerSight evaluates rank as an integer $R \ge 1$:

• $R = 1$: Primary recommendation or highest-ranked tool in the output.
• $R = 2$: Second recommended alternative.
• $R = 0$: The brand was either absent, or mentioned solely as a contrast or caveat outside the actual recommended set (e.g., “Unlike BrandX, the top recommendations are: 1. BrandA, 2. BrandB”).

Tie Handling

If an answer states “Both ClickHouse and Snowflake are outstanding choices for petabyte-scale analytics”, our evaluation engine programmatically scores them as a tie ($R = 1$ for both), rather than introducing arbitrary noise based on which word happened to appear first in the sentence.

---

3. Share of Voice (SoV)

Being mentioned is not enough if your competitor receives four detailed paragraphs of praise while your product gets a half-sentence mention.

Share of Voice measures the proportion of total generative real estate and attention allocated to your brand in the model’s answer, scaled from $0$ to $100$:

$\text{SoV}_i \approx \frac{\text{Token Attention}(\text{Brand}_i)}{\sum_{j \in \text{All Brands}} \text{Token Attention}(\text{Brand}_j)} \times 100$

In practice, SeerSight calculates SoV across four structural factors:

1. Mention Frequency: How many times the brand or its approved aliases appear.
2. Dedicated Token Volume: The proportion of sentences and paragraphs focused on analyzing the brand’s capabilities.
3. Depth of Analysis: Whether the brand is analyzed with dedicated subheadings, feature bullet points, and code snippets, or merely listed in a comma-separated string.
4. Centrality to Prompt Resolution: Whether the brand directly solves the user’s explicit constraints.

If your Visibility Rate is 90% but your Share of Voice is only 15%, the model recognizes your brand exists but considers your competitors far more relevant to the buyer’s needs.

---

4. Sentiment & Qualification Index

Traditional sentiment analysis algorithms rely on generic polarity dictionaries (positive vs. negative words) that fail catastrophically on nuanced technical documentation.

When an LLM says:

“ToolX offers enterprise-grade fine-grained RBAC and strict tenancy isolation, though setting up SSO requires contacting sales.”

A naive sentiment tool might flag “requires contacting sales” as negative and “enterprise-grade” as positive, cancelling each other out.

SeerSight employs specialized evaluation prompts to classify sentiment into three discrete ground states:

• $+1$ (Positive): Enthusiastic recommendation, favorable comparisons, or validation of core technical strengths.
• $0$ (Neutral): Balanced pros/cons, factual encyclopedic descriptions, or standard feature enumerations.
• $-1$ (Negative): Warnings, documented reliability complaints, steep pricing criticisms, or explicit recommendations to avoid.

Crucially, if your brand is absent from the output, sentiment is flagged as invalid (is_sentiment_valid = false). Absence is not “neutral”—there is simply no narrative tone to measure.

---

Automated Competitor Discovery: Why You Shouldn’t Have to Guess Rivals

In legacy SEO tools, you must manually guess and input your competitors: “Track my site against competitor1.com and competitor2.com.”

In AI search, the models frequently recommend emergent startups, open-source projects, or unexpected alternatives you didn’t even know existed.

When SeerSight runs daily evaluations across OpenAI, Anthropic, Google, and Perplexity, our engine:

1. Automatically identifies every third-party brand and tool mentioned in the model’s answer.
2. Disambiguates parent companies from product lines.
3. Computes relative Visibility, Rank, and Share of Voice for every discovered rival.
4. Alerts you to Competitor Displacement—the moment a rival begins appearing in prompts where your brand previously held 100% visibility.

By automating this across continuous daily evaluations, SeerSight gives engineering and product teams empirical, reproducible data to dominate the new frontier of search.