LLM Brand Visibility Audit
Measure AI awareness, sentiment, and co-citations to diagnose brand discoverability.
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 7
- Confidence
- 91%
The LLM Brand Visibility Audit treats generative search as a measurable representation of a brand's public footprint. First, run repeatable prompts for the category, competitors, and relevant customer segments. Then compare how often the brand appears, what sentiment surrounds it, and which descriptive terms or competing products appear alongside it. These co-citations expose the categories the model has learned to associate with the brand. A company might have strong general awareness yet disappear from an enterprise shortlist because public content repeatedly connects it with small businesses instead. The resulting diagnosis guides positioning and content work intended to reinforce accurate associations. Because outputs vary by model, prompt, and time, the audit should use multiple tests and be repeated after changes rather than relying on a single generated list.
Origin
Extracted from Marketing Against The Grain while analyzing why HubSpot appeared in one ChatGPT Search CRM list but not another.
Core principles
- 01AI recommendations reflect patterns learned from internet content.
- 02Relative mention volume indicates how strongly an LLM associates a brand with a market.
- 03Sentiment reveals whether awareness is positive, negative, or neutral.
- 04Co-cited terms help explain which categories and customer segments the model associates with a brand.
- 05Visibility should be measured across several prompts, not inferred from one answer.
How to run it
- 1
Define the Visibility Set
Choose the category, competitors, customer segments, and buying situations in which the brand should appear.
Pro tip Write down the expected associations before seeing the outputs.
Watch out Testing only branded prompts cannot reveal genuine category visibility.
- 2
Run Standardized Prompts
Ask multiple AI systems the same neutral discovery and comparison questions. Save the outputs and citations.
Pro tip Repeat prompts in fresh conversations to reduce context effects.
Watch out One result is too unstable to establish a visibility pattern.
- 3
Measure Mention Share
Count how frequently the brand appears relative to competitors across the test set.
Pro tip Separate first-position mentions from inclusion anywhere in the answer.
Watch out Mention volume alone does not indicate positive perception.
- 4
Analyze Sentiment
Classify the language around each brand as positive, negative, or neutral and identify recurring reasons for that sentiment.
Pro tip Retain supporting passages so classifications can be audited.
Watch out Model-generated sentiment is directional, not a perfect measure of customer opinion.
- 5
Map Co-Citations
Record the customer types, use cases, attributes, and competitors repeatedly mentioned beside the brand.
Pro tip Pay particular attention to desired words that rarely or never appear.
Watch out Correlation between terms does not prove how a proprietary model ranked the answer.
- 6
Close Association Gaps
Create or improve credible public material that accurately connects the brand with missing categories and use cases.
Pro tip Prioritize evidence-rich customer stories and third-party coverage over repetitive claims.
Watch out Do not manufacture endorsements or misleading associations.
- 7
Retest Consistently
Repeat the original prompt set after sufficient time and compare mention share, sentiment, and co-citations.
Pro tip Preserve the same models and prompts for a useful baseline.
Watch out Model updates can alter results independently of marketing changes.
In the wild
A CRM brand appears prominently when an AI is asked about small-business software but disappears from an enterprise comparison. The team audits mentions, sentiment, and adjacent language and discovers that public discussion strongly co-cites the brand with SMB terminology while rarely connecting it to large deployments. It then publishes verified enterprise case studies and encourages accurate third-party coverage before rerunning the original prompt set.
→ The team gains a testable explanation for the omission and a focused positioning plan.
Common mistakes
Treating One Prompt as a Score
An isolated result may reflect wording or output variance. Use a standardized set spanning multiple category and customer questions.
Optimizing Mentions Without Sentiment
More awareness is not automatically beneficial when the surrounding perception is negative or inaccurate.
Claiming Co-Citation Proves Causation
Co-citation is a useful diagnostic signal, but it does not reveal the complete ranking mechanism of a proprietary model.
Is it for you?
Best for
It is best for established brands seeking to understand and improve their visibility in AI-generated recommendations.
Not ideal for
It is not ideal as a substitute for customer research, conventional search analytics, or verification of factual brand claims.
From the transcript
“how much awareness do I have in the llm that is one of the ways that you can start to figure out how I can…”
“the llms can actually tell you how do people feel about your product positive negative neutral”
“the more your product is co-sited with other words that would make the llm think that you belong in either of those categories the more…”
From the episode
Google is overhauling Search - what you need to know