How to audit AI brand perception

In short: An AI brand perception audit is a controlled review of what named AI products say about a company across a fixed prompt panel. It records recognition, factual accuracy, framing, sentiment, recommendations, and citations separately. The results describe the sampled answers. They cannot reveal a model’s private reasoning or predict future rankings.

The phrase “AI brand perception” is often used for two different subjects: how consumers feel about a brand that uses AI, and how an AI-generated answer represents a brand. This audit addresses the second. Its unit of evidence is an observed answer under recorded conditions.

Separate the dimensions before scoring

A useful audit keeps eight questions apart:

1. Access: Can a relevant page be fetched and rendered under the controls you set?
2. Retrieval: Did the answer product appear to use a source for this prompt?
3. Entity understanding: Did it distinguish the company, product, parent, category, and namesakes?
4. Mention: Did the answer name the brand?
5. Citation: Did it link to a source, and does that source support the nearby claim?
6. Sentiment: Was a particular attribute framed positively, negatively, neutrally, or with mixed evidence?
7. Recommendation: Was the brand proposed as suitable for the stated buyer and constraints?
8. Business value: Did any observed referral or conversion follow, without assuming the answer caused it?

Treat these as distinct observations rather than a guaranteed funnel. A crawler request records access. A critical mention still counts as a mention, and a citation may support a definition while the recommendation rests on other material. Leaf’s guide to an AI visibility audit explains why each field needs its own denominator.

Official platform guidance helps define the observable boundaries. OpenAI publishes controls for its web crawlers in its crawler documentation, while Google explains eligibility for its AI search experiences in Search Central guidance. These documents explain access and eligibility, without offering a formula for predicting inclusion or citations.

Build a controlled 30-prompt panel

Create six groups of five prompts. Use actual buyer language and preserve exact wording:

• identity and canonical company facts
• category and problem recognition
• audience and use-case fit
• comparison and alternatives
• objections, risks, and limitations
• proof, implementation, and purchasing criteria

Keep neutral discovery prompts separate from branded checks. “What does Acme do?” tests recognition and accuracy. Whether Acme appears when a buyer asks for category options is a separate result. Include persona and constraint details only where they reflect a real buying situation. Keep recommendations for a small local firm separate from those for a regulated global enterprise.

When the question is specifically whether buyer context changes the recommendation, use the persona-dependent AI recommendation test rather than duplicating its one-variable-at-a-time matrix here.

Before collection, record the product, visible mode, signed-in state, market, date, panel version, and repetition policy. Logged-in experiences may use account context or product features that anonymous sessions do not. Therefore, compare like with like rather than treating one session as universal.

For high-value prompts, repeat runs on a declared schedule. Variability is itself a finding. Do not quietly rerun until the preferred answer appears. Preserve absences, refusals, and incomplete answers alongside favorable results.

Use an LLM Brand Perception Card

Create one row per prompt run and include these columns:

The card is the audit’s central artifact. If executives need a summary, show raw counts beside rates: “recommended in 4 of 30 recorded runs.” A label such as “13% of AI” claims a scope the sample cannot support. Publish the denominator, prompt mix, and missing runs.

For sentiment work, code claims at attribute level. An answer can praise ease of use while warning about enterprise controls. Calling that response simply “positive” destroys useful information. The ChatGPT brand sentiment method provides a dedicated annotation approach, while the defensible sentiment scoring guide covers reviewer agreement.

Compare models and repetitions fairly

Do not compare a browsing mode in one product with a non-browsing mode in another and attribute the difference to model quality. Report each tested experience as configured. Save complete answers and citations because interfaces and source displays change.

Review disagreement at three levels: between repeated runs in one product, between products, and between reviewers. A stable factual error deserves different action from one isolated omission. A citation that appears repeatedly may indicate an important source path, but it does not expose internal weighting.

Competitors should use the same prompts and coding rules. This controls the most obvious source of bias: testing your brand with generous branded prompts while testing competitors only in neutral discovery. The competitor recommendation audit follows the full shortlist from inclusion through rationale and caveats.

For a complete product-to-product protocol, use the ChatGPT, Perplexity, and Gemini brand test and keep product modes and conditions explicit.

Turn findings into a bounded action plan

Prioritize issues that could change a buyer’s decision. Wrong compliance claims, discontinued products, and confused namesakes belong near the top. A missing slogan usually belongs in routine monitoring.

Classify each action:

• repair robots, status, canonical, rendering, or discovery problems you can verify
• make canonical company facts direct, dated, and internally consistent
• correct stale profiles or independent sources through their normal editorial process
• add primary evidence where buyers need proof
• submit platform feedback when an interface offers it
• monitor uncertain or one-off observations before changing content

Retest the original panel after changes. A later improvement is an observed sequence, not proof that one edit caused it. Retrieval, model behavior, source availability, and presentation can all change.

For a structured starting point, run Leaf’s AEO assessment. Use the results to scope access, content, entity, and measurement checks, then verify the proposed work against recorded runs.