Why a B2B audit is different
Consumer questions to AI assistants are short: a dentist nearby, a cafe with parking. B2B buyers ask longer ones: "ERP vendor in Dubai with Arabic VAT invoicing", "ISO 27001 consultancy in the UAE for a fintech", "stainless steel fabricator in Sharjah for food plants". The engines answer these from a narrow set of sources, and the gap between the companies they name and the companies that exist is wide. The audit below takes about four hours. You need a spreadsheet, a private browser window and one person who knows the buyers.
Step 1: write 20 buyer questions
List the questions a buyer asks before a shortlist exists, not after. Each one should include your category, Dubai or UAE, and one buyer detail such as company size, industry, a certification or a budget signal. Mix four kinds:
- Vendor search: "which companies supply X in Dubai"
- Comparison: "X or Y for a UAE company with 200 staff"
- Procurement: "what does X cost in the UAE" and "what to ask an X supplier in Dubai"
- Brand check: your own company name, and "is [your company] a good X vendor in Dubai"
Keep the wording natural. A buyer types what they would say to a colleague.
Step 2: run them logged out in four engines
Open a private window with no account signed in, so no earlier chats color the result. Run every question in ChatGPT with search enabled, Claude, Gemini and Perplexity. That is 80 runs. Copy each answer into the log before moving on; answers change, and you will not be able to reproduce them later.
Log columns:
Question | Engine | Named (Y/N) | Companies named | Sources cited | Source type | Notes
Under "Sources cited" paste the URLs the engine links to. If an answer shows no sources, write "none" rather than leaving the cell blank; a run with no citations is itself a finding.
Step 3: group the sources
Sort every cited URL into one of six types: your own site, directories, review sites, LinkedIn, news and trade media, and lists (articles titled "top 10 X companies in Dubai"). Count how many times each type appears across the 80 runs. The type with the most citations is where the engines go to decide who exists in your category. If lists and directories lead and your own site barely appears, the engines are not reading you; they are reading people who write about your category.
Step 4: score share of voice
Share of voice is the number of runs that named you divided by the total number of runs. Calculate it overall and per engine. Then do the same for the three companies named most often, so the gap is a number rather than an impression.
| Engine | Runs | Named | Share of voice |
|---|---|---|---|
| ChatGPT (search) | 20 | 3 | 15% |
| Claude | 20 | 1 | 5% |
| Gemini | 20 | 4 | 20% |
| Perplexity | 20 | 2 | 10% |
| Total | 80 | 10 | 12.5% |
The figures above are a worked example, not a benchmark. A large difference between engines usually means one engine reads a source the others do not; check its citations first.
Step 5: check entity consistency
Open your website, Google Business Profile, LinkedIn company page, Crunchbase and the two or three directories that appeared most in Step 3. Compare the name, legal suffix, address, phone, founding year and one-line description. Engines build their picture of a company from these records. When the records disagree, the engine either picks one at random or leaves you out. A company listed as "ABC Technologies FZ-LLC" on one record and "ABC Tech" on another is two weak entities instead of one strong one.
Step 6: check the key pages
Take your five most important pages: home, the main service or product page, pricing or how you work, about, and the page that answers your most common buyer question. For each, check three things:
- A direct first-paragraph answer. Does the page say what you do, for whom and where, in the first 50 words, in plain text an engine can quote?
- FAQ schema. Does the page carry FAQPage structured data with questions written the way buyers ask them? Run it through a schema validator.
- An llms.txt file. Is there a file at yourdomain.com/llms.txt that summarizes the company and lists the key pages? Not every engine reads it yet, but it takes an hour to write.
Step 7: turn gaps into a fix list
Order the fixes by effect, not by ease:
- Entity conflicts in name, address and description across all records.
- Missing presence on the two source types the engines cited most.
- First-paragraph answers on the five key pages.
- FAQ schema and llms.txt.
- New pages for buyer questions where nobody was named well, because those are open.
Rerun the same 20 questions in six weeks and compare share of voice. The log is the asset; keep it.
If you would rather have it done for you, Samura's AI visibility audit runs the same method with 50 prompts across five engines for AED 1,999.
Keywords to build pages around
AI visibility audit UAE, LLM presence audit Dubai, how to check if ChatGPT recommends my company UAE, B2B AI search audit Dubai, share of voice in AI answers UAE, ChatGPT brand mentions check UAE, is my company cited by Perplexity UAE, generative engine optimization audit Dubai B2B, AI answer visibility check UAE, entity consistency audit Dubai company, FAQ schema audit UAE B2B, llms.txt for UAE companies, AI citation sources UAE B2B
Want to see where your business stands?
The AI Visibility Audit tests 50 real buyer questions across ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews and shows who is being recommended instead of you. AED 1,999, credited against any program within 30 days.
Frequently asked
How can a B2B company audit its presence in LLM-generated responses?
Write 20 buyer questions that include your category and Dubai or UAE. Run each one logged out in ChatGPT with search, Claude, Gemini and Perplexity. Record whether you were named, who was named and which sources were cited. Divide named runs by total runs for share of voice, group the sources by type, and compare the result with your own records and key pages.
How many prompts do I need for a useful AI visibility audit?
Twenty questions across four engines gives 80 runs, which is enough to see a pattern in an afternoon. Fewer than ten makes the result swing with wording. A paid audit uses 50 prompts across five engines, mainly to map the sources more completely and to measure competitors with less noise.
Why should the prompts be run logged out?
A signed-in account carries memory of earlier chats and settings, and may mention companies you have typed before. A buyer who has never heard of you has no such history. A private window with no account signed in gives you the answer a new buyer sees, which is the one that matters.
What is share of voice in AI answers?
The number of runs in which an engine named your company, divided by the total number of runs. If you were named in 10 of 80 runs, share of voice is 12.5 percent. Calculate it per engine as well, because a large gap between engines points to a source one engine reads and the others do not.
