Your company appears in Google, but an AI answer about suppliers lists competitors. Sometimes the system knows your name but invents a service, address or certification. Accurate business information and appearing in a supplier shortlist are two different tasks.
In brief: Start with six questions and save the answers and sources. A question containing your name tests information about the company. A question without it tests whether your company appears during supplier discovery. Repeat a defined sample in the buyers’ language; count mentions, recommendations and links separately.
Canem Errant checks a specific market, language and set of commercial questions. We record the answer, source URLs and factual discrepancies. That helps choose a limited correction and repeat the test instead of buying a large project from one screenshot.
Six questions for a first check
You can start a trial in five minutes. Recording several platforms and verifying their sources takes longer. Choose ChatGPT with search, Perplexity or Gemini and record the mode. A fresh chat reduces previous conversation context but does not remove every difference in settings or personalization.
Replace the placeholders and use the language buyers use to select a supplier. An English test alone does not describe German-language demand in Vienna.
- Category: “Which providers of [service] operate in [city or region], and how do they differ?”
- Company facts: “What is known about [company]? What services does it offer, and where?”
- Comparison: “Compare [company] with [competitor A] and [competitor B] by conditions and suitable tasks.”
- Benefits and limitations: “Which tasks suit [company], and when should I consider alternatives?”
- Reputation: “Which reviews of [company] can be verified? Give sources and do not invent missing information.”
- Evidence: “Are there published cases or documented achievements for [company]? Give links and periods.”
The second question already supplies the brand. Its appearance is not an independent recommendation. Comparisons also supply names; keep these tests separate from supplier discovery.
Add controls about a problem without the company name, a specialty, and three local providers with their strengths. These preserve the different tasks in a broader check. Do not ask for a market leader without defining an evidence-based criterion.
What should you record in an answer?
For every attempt, retain the question, date, platform, search mode, displayed model/version where available, language and market. Save the full answer and URLs. Note the company name, competing suppliers and factual accuracy, then classify how it appears: a mention, a recommendation or a source for a statement.
Verify address, service area, services, current conditions, authors and credentials. Open the source behind a claimed award or project result. Confident language is not evidence. If the answer says information is insufficient, record that rather than converting it to a global visibility score of zero.
An outdated profile, ambiguous name or incomplete service page is a possible explanation only after checking sources. Absence does not prove low trust; a correct answer does not certify the whole infrastructure. Keep observations and proposed explanations in separate fields.
A search-enabled attempt should not be directly compared with one without search. Do not paste promotional company text into the conversation before repeating: that changes what the test measures.
Three states: absent, mentioned, recommended
This classification applies to a particular answer and sample. It helps choose a check rather than certify the business.
| Answer state | What to check |
|---|---|
| Absent | Suitable service page, access, market fit and sources used. |
| Mentioned | Correct facts, clear specialty, source links and competing providers. |
| Recommended | Evidence for the recommendation, conditions, identity and repeatability. |
Record factual errors at every state. A strong recommendation with an invented certificate is worse than an accurate admission of insufficient evidence. Presence and accuracy belong together in the report.
For a local business, check Google Business Profile and suitable Austrian directories. Names, contacts and services should match facts. Not every business qualifies for every local profile or directory.
Why can a website be clear while the company remains unclear?
“Brand Entity Gap” is a useful description of a mismatch between a name and verifiable business information. A page claims expertise without explaining the service, customers or evidence; an external profile may still describe an old specialty.
Check business and service descriptions, authorship, published cases and profiles. Genuine reviews and relevant independent coverage help buyers verify claims and can become sources. No universal review count or two-stage verification architecture is established for every AI system.
An internal Canem Errant note dated 10 August 2026 compared two anonymous sites: technical scores of 78/100 with 0 mentions out of 6, and 55/100 with 3 out of 6. The original prompt log is not presented here. This small historical snapshot shows why a technical score cannot stand in for measured recommendation frequency; it does not establish why the results differed.
It cannot prove the effect of reviews, PR or llms.txt. Save a starting sample, choose a specific change and repeat the method to assess your own hypothesis. Automation helps collect observations after the questions and assessment rules are defined; it does not create business authority by itself.
How should you calculate mention share and AI Share of Voice?
Define the denominator: comparable answers in the chosen question set. Keep platforms, languages and modes separately reportable. Define a mention, recommendation and link to your domain before counting.
Illustrative example: a company appears in 45 of 100 comparable answers. Its mention share is 45/100 = 45%. If 12 recommend it, recommendation share is 12%; if 8 link to its domain, the share of answers with a citation is 8%. These describe separate outcomes in the same sample.
AI Share of Voice can mean different things: share of answers containing a brand, recommendation share, or share among all cited companies. State the formula. Without it, two SOV figures may be incomparable. Separate answers with at least one link from the total link count: three URLs in one answer are not three independent recommendations.
Keep branded questions apart from unbranded discovery. Repeat questions in fresh chats on a stated schedule and retain missing or failed attempts separately. Count comparable responses consistently rather than quietly removing inconvenient results. The sample does not represent the entire market.
Mention frequency is not direct ROI. Check visits, enquiries, qualification and deals separately. Later branded visits might follow an AI answer, but attributing the sale requires evidence. The GEO guide explains the broader measurement process.
Two illustrative scenarios
A local professional service
Suppose a Vienna firm offers a specific service, but its website and directories describe it broadly. Start with questions about the service, city and selection criteria. Verify existing sources, then correct scope, service area and terms.
Update Google Business Profile with facts. Ask genuine customers for voluntary accounts of their experience. Do not write invented results for them. Explain the consultation and required documents on the website, then repeat the same questions and compare accuracy and sources. This is a practice scenario, not a promise of a top-three recommendation in six weeks.
B2B SaaS with a large blog
Suppose product compatibility questions lead to other sources. Check the answer, documentation and links from the product page, then investigate crawler access, server restrictions and bot protection. Page count and a third-party authority score alone do not explain the omission.
PerplexityBot supports search indexing; Perplexity-User handles user-triggered fetches. A visit in the log does not prove a recommendation. Check both the user-agent and documented IP ranges when identifying requests. Perplexity documentation.
For ChatGPT, distinguish OAI-SearchBot for search, GPTBot for possible training and ChatGPT-User for user actions. Training restrictions and search access are independent. OpenAI documentation.
For Google, Googlebot and the effective Search generative AI control are relevant to Search. Google-Extended is a separate preference and does not determine ordinary Search inclusion. Google crawler documentation, Search Console control.
Identify the intended system and agree data-use policy before changing restrictions. Repeat the particular question after an approved change. An automatic crawl, user fetch and citation are distinct observations; no indexing deadline or promised number of links follows from permission alone.
Five actions for this week
- Save a starting sample. Run the six questions and separate unbranded controls. Retain settings, answers and sources in one record.
- Check company facts. Verify name, address, telephone, service area and services on the site and relevant profiles. Correct factual discrepancies rather than cosmetic formatting.
- Check independent evidence. Identify real reviews and relevant industry coverage. Choose useful information to supplement rather than manufacturing mentions.
- Improve one important page. Clarify the answer, conditions and evidence. Match Organization, appropriate LocalBusiness and FAQ markup to visible facts. Consider llms.txt separately for a system that actually uses it.
- Define the diagnosis. Choose a market, service and repeated question. Specify the starting check, a testable obstacle and the concrete change. A general promise to increase authority is too vague.
A monthly repeat is a useful starting schedule; add a separate check after important factual corrections. Keep the same question set, and record new questions as a separate group. AI workflow automation can help with collection once the method is stable.
Frequently Asked Questions
Why am I in Google but absent from an AI answer?
These are different search and response outcomes. Check the question, mode, market and sources. A Google position does not guarantee another platform’s recommendation, and absence does not establish one cause.
Can I pay to make AI recommend my company?
Advertising placement and an organic recommendation are different. Check the platform’s current product, rules and geography. Paying an agency for GEO does not buy the model’s organic decision to include your brand.
How often should I repeat the check?
Monthly with the same question set is a useful starting point. Save previous responses and settings. Add a separate check after corrections; one change in output should not immediately be attributed to your work.
Which factor matters most?
There is no universal published factor for every AI platform. Start with an accessible answer to the buyer’s question and verifiable information about the company.
Is llms.txt compulsory?
No. It is a proposed format for systems that use it, not a universal condition for recommendations. Google Search does not use it for visibility or ranking.
Do links and external publications matter?
They help buyers check a company and can be sources for an answer. Accuracy and relevance matter more than a domain count alone. Each external link is not a proven increase in AI trust.
For the format’s scope, see the llms.txt proposal and Google’s current guidance.
Discussing your results
Use the preliminary GEO check for an initial snapshot. For a repeated commercial problem, GEO diagnosis defines one market, language, question set and limited change. For Google-specific reporting, see AI Overviews, traffic and leads.
