ChatGPT, Perplexity, Gemini or Claude do not recommend a company with a universal grid.
A hairdresser, accountant, SEO agency, consulting firm, industrial supplier or regulated solution are not evaluated in the same way. Criteria change with risk, buying frequency, location, specialization, available proof and decision type.
This is central to GEO: being visible in AI is not enough. You need to understand the criteria AI uses to justify a recommendation in your specific market.
Short answer
ChatGPT and LLMs recommend a company by reconstructing a decision grid adapted to the market.
| Market type | Often dominant criteria | Often useful sources | GEO risk |
|---|---|---|---|
| Recurring local service | Proximity, rating, reviews, availability, customer sentiment | Google Business Profile, reviews, local directories | Being absent if the local profile is weak |
| Local intellectual profession | Specialization, target clients, seniority, website clarity | Website, specialty pages, reviews, professional sources | Being compared as a generalist |
| SEO or GEO agency | Visible results, positioning, public proof, published analysis | SERPs, studies, cases, expert content | Being cited without being judged credible |
| Complex B2B supplier | Buying criteria, proof, thresholds, use cases, risks | Criteria pages, studies, customer cases, documentation | Being excluded from the shortlist |
| Technical or regulated offer | Standards, compliance, traceability, verifiable proof | Documentation, certifications, matrices, RFP material | Being compared on a standard that is too weak |
The sentence to remember:
LLMs do not recommend a company only because it is visible. They recommend it when it is easy to justify within the criteria grid of its market.
Why this question already exists in search
Users no longer only search for a list of providers. They ask engines and AI systems to reduce uncertainty:
- “Who is the best accountant for an SMB?”
- “Which SEO agency should I choose?”
- “Which criteria should I compare to choose a supplier?”
- “Which providers do you recommend in my city?”
- “Which risks should I check before signing?”
- “Which company is most credible for this need?”
Gartner predicts that traditional search volume will drop by 25% by 2026 because of AI chatbots and virtual agents (Gartner). Pew Research Center also found that users click fewer links when a Google AI summary appears (Pew Research Center).
The issue is therefore not only to win a click. It is to understand the recommendation that forms before the click, sometimes without a click.
Why AI cannot apply the same grid everywhere
A recommendation depends on the decision type. The riskier the decision, the less simple criteria are enough.
| Factor | Effect on AI criteria |
|---|---|
| Decision risk | The higher the risk, the more AI looks for proof, thresholds, compliance, limits and rejection criteria |
| Buying frequency | The more recurring the purchase, the more customer experience signals can matter |
| Information asymmetry | The less the customer understands the field, the more AI looks for educational criteria |
| Location | The more proximity matters, the more local profiles and reviews matter |
| Offer comparability | The more similar offers seem, the more AI reuses available market criteria |
| Accessible sources | The clearer and more corroborated the sources, the more defensible the recommendation |
AI does not start from your sales promise. It starts from a need, a category and criteria it can justify.
Example 1: recurring local service
For a hairdresser, restaurant, landscaper or garage, AI may rely heavily on public local signals:
- proximity;
- Google Business Profile;
- primary category;
- average rating;
- review volume;
- review freshness;
- customer sentiment;
- availability;
- photos and practical information.
Google explains that local results are mainly based on relevance, distance and prominence (Google Business Profile Help).
In this type of market, AI can recommend a company because it is easy to justify locally: well located, well rated, frequently reviewed and clearly categorized.
Risk
A competent company can be under-recommended if its local profile is incomplete, if reviews do not describe the right services or if the website does not confirm its area and specialties clearly.
Example 2: local intellectual profession
For an accountant, lawyer, architect or specialized consultant, the grid can change.
The Google rating is still useful, but it is not always enough. AI may look for:
- specialization by client type;
- situations handled;
- sectors covered;
- local seniority;
- degrees, certifications or affiliations;
- detailed service pages;
- proof of expertise;
- website clarity;
- anonymized cases or mission examples.
An accountant can therefore be recommended for SMBs, freelancers, agricultural businesses, startups, property tax situations or company acquisitions depending on what AI finds in the sources.
Risk
If the website only carries generic promises, AI may compare the firm as a generalist. The specialization may be real, but it does not structure the recommendation.
Example 3: SEO or GEO agency
In SEO or GEO, declared experience matters less than visible proof.
An agency can claim that it understands Google, ChatGPT, Perplexity or LLMs. But AI may give more weight to signals that are easier to verify:
- visible SERP positions;
- published analysis;
- original studies;
- documented cases;
- explained methodology;
- citable content;
- consistency between claims and observable results;
- ability to position itself on the queries it works on.
The logic differs from professions where seniority or professional status plays a stronger role. For an SEO or GEO agency, proof through results can become a central credibility criterion.
In this context, a company that publishes useful, structured and verifiable analysis can enter a proof loop: the more its work helps engines answer correctly, the easier it becomes to treat that company as a credible source or provider.
This loop does not depend only on brand citation. If an analysis structures an engine’s answer, it can influence the decision even when the click or explicit citation does not appear.
Example 4: complex B2B supplier
For a B2B supplier, AI does not only search for “who exists”. It may search for who is defensible in a shortlist.
Criteria often become more demanding:
- use cases;
- integration;
- fit with context;
- verifiable proof;
- risks avoided;
- performance thresholds;
- comparable references;
- usage limits;
- rejection criteria.
In this type of market, an AI recommendation looks less like a directory and more like a first buying grid.
Risk
A supplier can be cited but absent from the shortlist if AI cannot explain why it should be selected over a competitor that is easier to justify.
Example 5: technical or regulated offer
For a regulated, industrial, medical, cybersecurity or compliance offer, criteria can be stricter.
| Evaluated element | What AI must understand |
|---|---|
| Standard | Which standard applies, and in which context |
| Proof | Which document, test, audit or certificate verifies compliance |
| Threshold | When the offer becomes insufficient |
| Exposure | Which risk level changes the requirement |
| Limit | When the offer should not be selected |
| Rejection | Which missing element should exclude a supplier |
When these elements are missing, AI can recommend a minimum standard. The risk is not a spectacular hallucination. The risk is a plausible but under-specified recommendation.
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.
These figures do not describe every market. They show how a controlled corpus of AI answers mobilizes sources, proof, categories, criteria and shortlist logic.
| Observed signal | Volume in the corpus | Reading for market-specific criteria |
|---|---|---|
| Source dependency | 4,137/4,320 (95.8%) | Available sources strongly structure the answer |
| Proof reuse | 3,714/4,320 (86.0%) | Visible proof can become justification |
| Documentation and proof | 2,687/4,320 (62.2%) | Documentation acts as decision infrastructure |
| Shortlist and vendor evaluation | 2,168/4,320 (50.2%) | One answer in two activates selection logic |
| Criteria reuse | 1,464/4,320 (33.9%) | Explicit criteria can structure the comparison |
| Category compression | 1,061/4,320 (24.6%) | A wrong category can change applied criteria |
| Specification gap | 1,130/4,320 (26.2%) | The requirement level can be formulated too low |
The important reading is simple: LLMs do not recommend only from popularity. They reuse what the market makes readable: sources, proof, criteria and categories.
The 6 factors that make AI criteria vary
1. Decision risk
The more costly the wrong decision, the more AI must justify the recommendation.
| Low risk | High risk |
|---|---|
| Reviews, availability, proximity | Proof, compliance, thresholds, scenarios, rejection |
2. Buying frequency
A frequent local service often generates more usable reviews. A rare or complex purchase generates fewer simple signals, so AI must look for more structured proof.
3. Specialization
The more specialization matters, the less a generic page is enough.
| Market | Useful specialization criterion |
|---|---|
| Accountant | Client type, sector, tax situation, management context |
| Lawyer | Legal domain, case type, jurisdiction, client profile |
| Consultant | Problem handled, method, context, decision level |
| Industrial supplier | Use case, constraint, standard, exposure environment |
4. Visible proof
Google recommends creating helpful, reliable, people-first content (Google Search Central). For LLMs, this connects to an operational principle: proof must be understandable by a human and extractable by a machine.
5. Source format
Google also explains that structured data helps understand page content, but should describe content visible to users (Google structured data).
Structured data can help. It does not replace a page that clearly explains criteria, proof and limits.
6. Category used
The category determines competitors and criteria. If AI places you in a category that is too broad, it can apply the criteria of a simpler market.
This connects with Category Compression Risk: an offer can be visible, but evaluated in the wrong family.
How to know which criteria AI applies to your market
The test must use natural questions, not only your brand name.
- List the buying situations where you want to be recommended.
- Ask the same questions in ChatGPT, Perplexity, Gemini and Claude.
- Note the criteria used in each answer.
- Capture the sources cited or visibly reused.
- Identify associated competitors.
- Classify signals: reviews, proximity, specialization, proof, compliance, price, result, risk.
- Spot what favors you, what commoditizes you and what excludes you.
- Publish missing criteria as clear blocks.
| Question to test | What it reveals |
|---|---|
| ”How do I choose a [profession]?” | Generic decision grid |
| ”Who is the best [profession] for [situation]?” | Contextual criteria |
| ”Which providers do you recommend for [need]?” | Likely shortlist |
| ”Which risks should I check before choosing?” | Rejection criteria |
| ”Which proof should I ask before signing?” | Expected proof |
| ”Which suppliers should I avoid if [constraint]?” | Limits and exclusions |
Repeat questions several times. One answer can vary. A repeated pattern reveals a more stable grid.
What to publish depending on your market
| Market | Priority content |
|---|---|
| Local service | Complete Google profile, detailed reviews, location pages, photos, local FAQ |
| Intellectual profession | Specialty pages, situation guides, anonymized cases, proof of expertise |
| SEO or GEO agency | Studies, methodology, SERP analysis, practical cases, proof of results |
| Industrial B2B | Criteria/proof matrices, use cases, standards, thresholds, limits |
| Regulated offer | Compliance documentation, verifiable proof, rejection criteria, buyer FAQ |
| Premium offer | Avoided risk, cost of a poor decision, contextual comparisons |
The ideal format remains simple: question, context, criterion, proof, threshold, limit, rejection.
Why brand citations are not enough
A citation answers one question:
“Does AI know your name?”
Criteria variation answers a more important question:
“Does AI know why you should be selected in this specific market?”
A brand can be cited and poorly recommended. It can also be cited less often, but structure the decision grid if its content explains market criteria better.
This is why GEO should not stop at brand citations. The real value is Decision Presence: being associated with the right criteria, proof and buying situations.
Quick diagnosis
| Symptom in AI answers | Probable problem | Priority correction |
|---|---|---|
| AI cites your brand but recommends a competitor | Presence without preference | Publish criteria that justify selection |
| AI compares your offer with generic players | Category compression | Clarify category and wrong substitutes |
| AI mostly discusses price | Avoided risk is absent | Document the cost of a poor decision |
| AI does not reuse your proof | Low Proof reuse | Turn proof into extractable blocks |
| AI relies on weak or incomplete reviews | Unfavorable source dependency | Strengthen visible sources |
| AI ignores your specialization | Pages are too generic | Create pages by situation, client or constraint |
Key takeaway
LLMs do not recommend every company with the same grid.
They adapt criteria to the market:
- local reputation for some recurring services;
- specialization for intellectual professions;
- visible proof for SEO and GEO;
- criteria, risks and thresholds for complex B2B buying;
- compliance and rejection criteria for regulated offers.
The useful work is therefore to identify the grid AI already applies to your market, then make the right criteria visible, verifiable and easy to reuse.
Read next
- How does AI build a supplier shortlist?: understand selection mechanics.
- Why does AI recommend some providers over others?: understand recommendation signals.
- Why am I cited by ChatGPT or Perplexity but not converting?: distinguish citation and recommendation.
- Category Compression Risk: detect wrong categories that change criteria.
FAQ
Does ChatGPT use the same criteria to recommend every company?
No. ChatGPT, Perplexity and LLMs adapt criteria to the market type, risk level, buying frequency, location and available sources.
Which criteria matter for a local service?
For a recurring local service, AI may give strong weight to proximity, Google Business Profile, reviews, rating, review freshness and customer sentiment.
Which criteria matter for an intellectual profession?
For an accountant, lawyer or consultant, AI may look more at specialization, client types, seniority, proof of expertise and website clarity.
Why can an SEO agency be recommended without being the oldest one?
In SEO or GEO, visible proof through results, published analysis, rankings achieved and the ability to explain the market can matter more than declared experience.
How do I know which criteria AI applies to my market?
Test several natural questions in ChatGPT, Perplexity, Gemini and Claude, then code the criteria used, sources cited, competitors associated, proof reused and exclusion cases.
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?