The Three Core Metrics That Actually Measure AI Visibility

---

Metric 1 of 3

Latent Brand Association

What the model already "believes" about your brand before it ever touches the web

Every LLM has opinions about your brand.

Nobody programmed them in. They formed on their own from billions of documents the model absorbed during training.

If your brand showed up a lot in positive contexts on authoritative sites, the model learned to like you. If your competitor dominated those conversations instead, the model learned to prefer them.

That's what Latent Brand Association measures. The strength and direction of what the model has already internalized about you, your competitors, and the categories you compete in.

LBA answers one specific question: what does the model believe about you?

Is the association positive? Negative? Outdated? Dominated by a competitor?

This is different from Top of Mind, which measures whether the model recalls you at all. A brand can have strong recall but carry negative associations. Or it can have favorable associations that are too weak to surface on their own. Both matter, but they diagnose completely different problems.

Why You Can't Fix This Overnight

These associations live in the model's neural weights. They were formed during training, and they stick around until the next major training cycle.

That means the model might "remember" your brand as it was two or three years ago.

You could have completely repositioned your product, launched a new category, or doubled your market share since then. The model doesn't know. It's working from a snapshot baked into its weights.

How We Actually Measure It

Because LLMs are probabilistic, one probe tells you almost nothing. We ask each model five different questions about your brand, run each question five times, and add a control prompt (same question but with "a typical brand in your category" swapped in). The control catches models that are faking recall by riffing on your brand name rather than actually knowing it. Then a cheap extraction model classifies every association by polarity, freshness, factuality, and whether the model flagged its own uncertainty.

We do this separately for ChatGPT, Claude, and Gemini — each has different training data and different biases, so you get a per-model breakdown. Full methodology is on the LBA deep-dive page.

The Six LBA Patterns

After running this methodology against brands ranging from Nike and Ahrefs to a completely made-up name, we've identified six recognizable patterns. Your brand almost certainly falls into one of them.

Pattern	What it looks like	What to do
Product-Strong e.g. Ahrefs	Model names your specific products by name across runs ("Site Explorer," "Content Explorer"). Knows your pricing tiers and reputation signals.	Protect and reinforce. Actively seed coverage for new product launches within 12 months, or they'll miss the next training cycle.
Iconic but Oscillating e.g. Nike	Category is perfectly stable. Specific differentiators (slogans, signature products) show up on only some runs.	Category dominance is fine. Concentrate coverage around the oscillating differentiators — headlines on crawled domains.
Integration-Strong e.g. DomCop	Model knows your integrations and partners specifically ("integrates with Moz, GoDaddy"). Identity defined by "X for Y" framing.	Lean into the integration framing. Also build standalone product-level associations so you're not structurally dependent on partners.
Category-Known, Product-Unknown e.g. Medicube	Model knows exactly what you do and who for, but can't name a single specific product you sell. Most common mid-sized-brand pattern.	Push product-specific content into authoritative reviews and comparisons. Product schema, product names in headlines, Wikipedia where possible.
Entity Collision e.g. Domino (magazine) vs Domino's (pizza)	You share a name with a more famous brand. The model flips between interpretations run-to-run; the famous one usually wins.	Always use a distinguishing descriptor in your content. Teach customers to disambiguate. Accept that LBA will be structurally capped when you share a name with a more famous brand.
Unknown or Confabulated e.g. a brand-new startup	Model either flags uncertainty ("generic placeholder brand") or confidently invents a fake backstory. Either can happen on the same brand.	No shortcuts. Build authoritative coverage (Wikipedia, industry press, founder interviews). Expect 12–24 month payoff tied to training cycles.

Read the full LBA deep-dive → — includes the exact probe prompts, how the score is calculated, and a detailed playbook per pattern.

---

Metric 2 of 3

LLM Authority Score

How often and how prominently AI models feature your brand across dozens of responses

In traditional SEO, rank is a single number. You're position 3 for "project management software" and that's your data point.

AI doesn't work like that.

Ask the same question ten times and you'll get ten different answers with different brands, different features, and different recommendations. Any single response is noise. Authority only becomes visible across many responses.

LLM Authority Score combines two dimensions into one metric: frequency (how often your brand appears) and prominence (where in the response it shows up). Both matter. Neither alone tells the full story.

Frequency Alone Will Fool You

One Response Means Nothing

A single AI response is a coin flip. The model might mention you first today and skip you entirely tomorrow.

Traditional "rank tracking" for AI responses fails because it treats each response like a stable SERP. It's not. The model's output is probabilistic.

LLM Authority Score solves this by measuring across many responses. Instead of reporting "you were #2 in this one response," it tells you: "Across 50 responses to this query, you appeared 72% of the time with a weighted average position of 2.3."

That's actionable. That's something you can track over time and actually improve.

How We Actually Measure It

We run your onboarding prompt set (40 to 70 category queries tagged by intent and weighted by search volume) against each model in both modes: recall (no web access) and retrieval (web search on). Users experience both, so we measure both. For each response, an extraction pass pulls out every brand mentioned and its position in the answer. Frequency × prominence (using log decay, since user attention drops off quickly past position 1) gives a per-prompt score, which is aggregated across prompts, weighted, and averaged across modes (50/50) and across models.

One critical rule: we exclude prompts that name the user's brand. "Ahrefs vs SEMrush" forces Ahrefs into the response, which is inclusion, not authority. We confirmed this matters by scoring a made-up brand called AcmeWidgetsXYZ, with self-referential prompts included it scored 16 out of 100; without them, correctly 0. Authority only counts organic mentions. Full methodology on the Authority deep-dive page.

Patterns We've Observed

We've tested this methodology against real brands. Three distinct patterns emerged, each with its own strategic implications. Three more are expected but not yet observed.

Pattern	What it looks like	Diagnosis
Recall-Led Leader e.g. Ahrefs (Combined 82, gap +8)	High score in both modes, but recall beats retrieval. Training-data dominance outpaces current web footprint.	Protect the moat. Competitors publishing heavily in the next 12-18 months will erode your retrieval lead first, then your recall lead.
Retrieval-Led Challenger e.g. DomCop (Combined 42, gap -10)	Retrieval higher than recall. Current web knows you, training data lags.	Emerging brand. Short-term lever is retrieval, long-term lever is authoritative mentions that will land in the next training cycle.
Floor e.g. AcmeWidgetsXYZ, made-up brand (Combined 0)	Zero in both modes. Neither training data nor current web knows you.	Build from zero. Wikipedia, press, industry listings, founder interviews. 12-24 month recovery tied to training cycles.
Category Ruler expected pattern	High in both modes (85+), gap near zero. Iconic brand with matched training-data and web presence.	Protect. Watch competitor retrieval growth for erosion signals.
Oscillating Challenger expected pattern	Mid-range score with high intra-run variance. Appears strongly in some iterations, not at all in others.	Unstable brand association. Push content weight to stabilize.
Shadow Authority expected pattern	High frequency, low prominence. Mentioned a lot but always buried mid-list.	The model has you categorized wrong. Re-frame content around the exact category phrasing users query for.

Read the full Authority deep-dive → covering exact scoring formulas, the fragility gap, and a per-pattern playbook.

---

Metric 3 of 3

Top of Mind

Whether the model recalls your brand from memory, or only finds you through web search

AI systems have two ways to bring your brand into a response.

The first is recall: the model already knows about you from training data and mentions you from memory. The second is retrieval: the model searches the live web, finds your content, and pulls it in.

To the end user, these look identical. For your brand, they're completely different.

Top of Mind measures the first path. When the model can't search the web, does it still mention you?

Where LBA tells you what the model believes about your brand, TOM tells you whether it recalls you at all without external help.

Recall vs. Retrieval: The Gap Most SEOs Miss

Why This Changes Everything

Retrieval-dependent visibility is fragile.

If the AI tool decides not to search the web for a particular query (which happens), you vanish. If a competitor's content outranks yours in the retrieval step, they replace you. If the AI system changes its retrieval algorithm, your visibility shifts overnight.

Recall-based visibility is the opposite. It's encoded in the model's weights. It persists whether or not the model searches the web.

Think of it as the AI equivalent of a consumer instinctively naming your brand when asked about a category.

How We Actually Measure It

TOM uses a dedicated set of 15 discovery-intent category prompts for your brand ("best X for Y", "top-rated X", "most popular X in 2026"). Each prompt runs five times per model in recall mode, web search explicitly off. That's 75 data points per model per brand. We score frequency × (0.5 + 0.5 × prominence), so the primary question is "did the brand surface?" with position as a secondary modifier. Weighted by search volume, aggregated per model, then averaged across ChatGPT, Claude, and Gemini.

Full methodology on the TOM deep-dive page.

Five Patterns We've Observed

We tested this methodology against five brands during validation. The results sort into five distinct patterns.

Pattern	What it looks like	What to do
Category Ruler e.g. Nike (TOM 97.9)	Near-perfect dominance. Position 1 in ~90% of discovery queries, every iteration.	Protect the position. Monitor competitor growth for erosion signals.
Category Leader e.g. Ahrefs (TOM 87.0)	100% frequency, mostly position 2-3. Always in the conversation but shares top spot with 1-2 rivals.	Identify position-1 weak spots. Push content density to overtake closest competitor on those queries.
Specialty-Recall e.g. DomCop (TOM 55.2)	Strong on niche queries (position 1, every iteration). Invisible on broad category queries.	Seed content that frames you in the broader category phrasing you want to own, not just the sub-category.
Prompted-Recall-Only e.g. Medicube (TOM 0.0, LBA ~50)	Model knows you when asked directly (LBA > 0) but never volunteers you in category queries. Not in the model's competitive set.	Get listed in "Top N" category articles alongside established competitors. Co-mention density is the single biggest lever.
Floor e.g. a made-up brand (TOM 0, LBA 0)	Both TOM and LBA at zero. Model has no training signal for this brand.	Build foundational coverage: Wikipedia, press, industry listings, founder interviews. 12-24 month horizon.

Read the full TOM deep-dive → includes the exact probe prompts, scoring formula, cross-metric diagnostics, and a per-pattern playbook.

---

How the Three Metrics Work Together

Each metric measures a different layer of your AI visibility. Together, they tell you exactly where you stand and why.

Metric	What It Measures	The Question It Answers
Latent Brand Association	Pre-trained model bias toward your brand	What does the model already believe about us?
LLM Authority Score	Frequency + prominence across responses	How seriously does the model take us overall?
Top of Mind	Unprompted recall without web search	Would the model mention us without looking us up?

Two Layers of Model-Side Signal

LBA and TOM are both pure-recall measurements. They measure what the AI model has internalized from training data. You can't change them quickly because they depend on training cycles that take months or years.

LLM Authority Score sits as the bridge, reflecting both what the model recalls and what it pulls from live web search. It's the closest-to-user metric, the signal that most directly tracks what people see when they interact with AI assistants.

Layer	Metric	You Control It?	Timeframe to Change
Model-side	Latent Brand Association	Indirectly, through content that enters future training data	6-18+ months (next training cycle)
Model-side	Top of Mind	Indirectly, through sustained authoritative presence	6-18+ months (next training cycle)
Bridge	LLM Authority Score	Partially, combines model recall and retrieval performance	Weeks to months (retrieval side) + months to years (recall side)