To know whether ChatGPT, Perplexity or Gemini really understands your offer, do not only check whether AI cites your brand.
Check how it explains you, which competitors it compares you with, which criteria it uses, which proof it cites and when it recommends you.
AI can know your name and still misunderstand your value.
Short answer
An LLM understands a B2B offer when five elements remain coherent.
| Element | Control question | Good signal |
|---|---|---|
| Category | Which family are we placed in? | Precise and useful category |
| Competitors | Who are we compared with? | Truly comparable alternatives |
| Criteria | How is the offer evaluated? | Differentiating criteria reused |
| Proof | What justifies the reading? | Verifiable proof cited |
| Recommendation | When are we selected or rejected? | Correct use cases and limits |
If these five elements are coherent, AI is not only citing you. It understands you in a decision logic.
Why you should test several LLMs
ChatGPT, Perplexity, Gemini or Claude do not always use the same sources, answer modes or caution thresholds. An offer can be well classified in one engine and poorly understood in another.
| Engine to test | What the test often reveals |
|---|---|
| ChatGPT | Perceived category, associated competitors, generic choice criteria |
| Perplexity | Cited sources, proof reuse, dependency on specific pages |
| Gemini | Alignment with Google results, adjacent categories, brand signals |
| Claude | Reasoning quality, limits, objections and risks |
The right diagnosis is therefore not to ask once, “what do you think of my offer?” It is to compare several answers, across several prompt formulations, then code the gaps.
What Beyond Mentions data shows
In the first consolidated wave of the Beyond Mentions Observatory, we analyzed 4,320 Perplexity sonar answers over 3 UTC days, with 6 passes per question per day.
| Observed signal | Volume in the corpus | Reading for understanding |
|---|---|---|
| Source dependency | 4,137/4,320 (95.8%) | Available sources influence explanation |
| Proof reuse | 3,714/4,320 (86.0%) | Proof can stabilize the reading |
| Documentation and proof | 2,687/4,320 (62.2%) | Documentation helps explain value |
| Criteria reuse | 1,464/4,320 (33.9%) | Explicit criteria can be reused |
| Category compression | 1,061/4,320 (24.6%) | An offer can be placed in the wrong category |
| Specification gap | 1,130/4,320 (26.2%) | Requirements can be poorly formulated |
The key point: AI understanding is not binary. It is measured through coherence between category, criteria, proof and recommendation.
The 6 tests to run
1. Category test
Ask:
“How would you define this offer?”
Check whether ChatGPT uses a precise category or a fallback category.
| Answer | Reading |
|---|---|
| Precise category | Good signal |
| Category too broad | Compression risk |
| Competitor category | Wrong-comparison risk |
| Vague description | Insufficient documentation |
2. Competitor test
Ask:
“Which alternatives should be compared?”
If alternatives are too distant, ChatGPT does not yet understand the scope.
3. Criteria test
Ask:
“Which criteria should be used to choose?”
Reused criteria should reflect your real value, not only price, awareness or features.
4. Proof test
Ask:
“Which proof should be requested from this provider?”
A good signal appears when ChatGPT cites verifiable proof, not only declarations.
5. Shortlist test
Ask:
“In which case would you recommend this offer?”
The answer should connect your offer to precise buying situations.
6. Exclusion test
Ask:
“In which case should this offer not be selected?”
AI that can exclude correctly often understands better than AI that recommends too broadly.
Diagnosis
| Symptom | Probable problem | Correction |
|---|---|---|
| ChatGPT cites you but explains you poorly | Weak definition | Publish a category definition |
| Competitors are incoherent | Weak category fit | Clarify wrong substitutes |
| Criteria are generic | Differentiating criteria absent | Publish matrices and thresholds |
| Proof does not appear | Low Proof reuse | Make proof extractable |
| Offer is recommended too broadly | Missing limits | Publish limits and rejection criteria |
| Price dominates | Avoided risk invisible | Document the cost of a poor decision |
Simple scoring grid
| Dimension | 0 | 1 | 2 |
|---|---|---|---|
| Category | Wrong | Broad | Precise |
| Competitors | Incoherent | Partial | Relevant |
| Criteria | Generic | Mixed | Differentiating |
| Proof | Absent | Declarative | Verifiable |
| Recommendation | Vague | Partial | Contextualized |
| Exclusion | Absent | Fuzzy | Actionable |
A low score does not mean your offer is weak. It means your documentation does not yet make the right reading easy enough.
Key takeaway
ChatGPT, Perplexity or Gemini really understand your offer when they can:
- place it in the right category;
- compare it with the right competitors;
- reuse the right criteria;
- cite the right proof;
- recommend the right use cases;
- exclude the wrong cases.
So the question is not:
“Does AI talk about us?”
The real question is:
“Does AI prepare a decision that reflects our real value?”
Read next
- Why am I cited by ChatGPT or Perplexity but not converting?: distinguish citation and recommendation.
- Category Compression Risk: understand wrong categories.
- Why does AI recommend some providers over others?: connect understanding and recommendation.
- Documentation Blind Spot: correct missing proof.
FAQ
Is one prompt enough to know if an LLM understands my offer?
No. A single answer can vary. Test several questions and angles, then code recurring patterns.
What is the best signal of understanding?
The best signal is coherence between category, competitors, criteria, proof and the situations where ChatGPT, Perplexity or Gemini recommends or excludes the offer.
What should I do if AI simplifies my offer?
Publish definitions, criteria, proof, use cases, limits and rejection criteria that make the real category easier to reuse.
What is the difference between being understood and being cited?
Being cited means appearing. Being understood means being explained, compared and recommended with the right criteria.
What question does the buyer ask AI?
Which documentation simplification can lower the standard?
Which technical requirement must be clearly formulated?
Which evidence should be requested or published?
Which criterion excludes an insufficient answer?