In short
To check whether AI assistants recommend your business, write 15 to 20 real buyer questions, ask each one in ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, and record whether you are named, how you are described and which pages are cited. Repeat monthly, ask each question twice, and treat the results as trends, because answers vary between sessions.
Why test AI assistants at all?
More people now ask ChatGPT, Gemini, Perplexity and Claude for recommendations instead of scrolling through search results. If a buyer in Nairobi asks “which accounting firm should a small business in Westlands use?” and your firm is not in the answer, you do not exist for that customer, even if you rank on Google.
You cannot improve what you do not measure, so start with a simple, repeatable test. It needs no special software, only a spreadsheet and an hour a month.
Step 1: How do you choose your questions?
Write 15 to 20 questions a real buyer would ask, in the words they would use. Mix the types:
- Category questions: “Best SEO agency for a hotel in Kigali”.
- Comparison questions: “Which is better for a small business, X or Y?”.
- Problem questions: “How do I get more customers from Google in Uganda?”.
- Brand questions: “What does [your business] do?” and “Is [your business] any good?”.
Include the cities and languages that matter to you. Do not write questions that already contain your brand name, except for the brand questions, or you will test nothing.
Step 2: Where do you ask them?
Ask the same questions in each assistant your customers might use: ChatGPT, Google’s Gemini and AI Overviews, Perplexity and Claude. Use a fresh chat each time, signed out where possible, so earlier conversations do not shape the answer.
Step 3: What do you record?
| Column | What to write |
|---|---|
| Date and tool | For example, 2026-10-05, Perplexity |
| Question | The exact words you used |
| Named? | Yes or no, and your position in the list |
| Cited? | Which of your pages, if any, was linked |
| Accuracy | Were the facts about you correct? |
| Competitors | Who else was named |
Step 4: Why should you ask each question twice?
Assistants can vary their wording and sources between sessions, and results may also differ by location and settings. Ask each question at least twice and look for patterns instead of single answers. A business that appears in most runs is genuinely visible. One that appears once may be luck.
Step 5: How do you read the results?
- Not named anywhere: you are not yet a known answer. Focus on clear service pages, consistent facts and trusted mentions.
- Named but described wrongly: fix your own pages, Google Business Profile and listings so the facts agree.
- Named but never cited: your content may not state the answer in a quotable way. Add answer-first sections and FAQs.
- Competitors named, you not: look at what those competitors have, such as reviews, directory listings and press mentions.
Step 6: What do you do with the findings?
Turn each gap into a task: a page to write, a fact to correct, a listing to claim, a review campaign to start. Then re-test next month. Be patient: changes in AI answers often take weeks to appear, and no one can guarantee a result.
What does progress look like in practice?
In our LAGE Lightings case study, we tracked a set of buyer questions over several months and recorded how the brand moved in ChatGPT, Gemini and Claude. We also noted what remained unfinished, because honest reporting means showing the gaps as well as the wins.
A recording table you can copy
| Date | Assistant | Exact prompt | Named? | Page cited? | Facts correct? | Others named | Notes |
|---|---|---|---|---|---|---|---|
| 2026-10-05 | ChatGPT (search on) | “Best [service] in Kigali” | Yes / No | Yes / No | Yes / No / Partly | Names | What was wrong or missing |
| (example row) | Perplexity | “[Service] near [landmark]” | No | No | n/a | Competitor A, B | Cited directory pages only |
Keep the sheet in a shared spreadsheet. One row per question, per assistant, per test, with the date, so you can compare months.
A prompt library by industry
| Industry | Example prompts |
|---|---|
| Hotels | “Hotels in Kigali with meeting rooms for 40 people”, “Boutique hotel in Kigali with airport transfers”, “Family hotel in Kigali with a pool”, “Where to stay near Kigali Convention Centre” |
| Restaurants | “Best restaurant for a business dinner of eight in Kigali”, “Vegetarian restaurants in Kigali”, “Restaurants open late on Sunday in Kigali”, “Where to eat Rwandan food in Kigali” |
| Real estate | “Best neighbourhoods in Kigali for a diaspora buyer”, “Reliable property agents in Kigali”, “How does buying property in Rwanda work for foreigners?” |
| B2B and professional services | “Accounting firm in Kigali for foreign companies”, “Who can help register a company in Rwanda?”, “Logistics company for imports to Kigali” |
| Retail and shops | “Where to buy [product] in Kigali”, “Best [product] shop with delivery in Rwanda” |
| Clinics | “Dentist in Kigali that speaks English and French”, “Clinic in Kigali open on Sunday” |
Swap in your own neighbourhood, product and language. Include brand questions too: “What do you know about [your business]?”
A simple scoring system
| Score | Meaning |
|---|---|
| 0 | Not named |
| 1 | Named, but with wrong facts |
| 2 | Named, with correct facts |
| 3 | Named with correct facts, and your own page cited |
Add the scores and divide by three times the number of answers. The percentage is your visibility score. It is a way to compare months, not a guarantee of anything.
What a month-three report might look like
This is an invented illustration, with made-up numbers, to show the shape of a report:
| Measure | Month 1 (baseline) | Month 3 |
|---|---|---|
| Questions tested | 20 | 20 |
| Answers naming the business | 3 | 8 |
| Answers citing your own page | 0 | 3 |
| Answers with a wrong fact | 2 | 1 |
| Visibility score | 12% | 33% |
Real results will differ. What matters is that the same questions, in the same way, show a trend.
Should you log in or stay logged out?
Both are valid, but be consistent and note which you used. Logged-in assistants may personalise answers using your history. Logged-out tests are closer to what a new customer sees. Free access can be limited; for example, when we ran a small panel in October 2026 one assistant stopped answering after a few questions. See what we found in our Kigali audit of AI answers.
Common questions
How many questions do I need?
Fifteen to twenty is a good start for a single business. More can help later, but a small, fixed set you repeat is more useful than a large set you run once.
Why do answers change between runs?
These systems can vary wording and sources between sessions, and the web changes. That is why we ask each question twice and look at trends.
Which assistants should I test?
At least ChatGPT with search, Gemini, Perplexity and Google AI Overviews, and add Claude if your buyers use it. See what each crawler documents in OpenAI: crawler documentation, Perplexity: crawler documentation and Google Search Central: AI features and your website.
What mistakes should you avoid?
- Testing with questions that already contain your brand name.
- Judging by one answer on one day.
- Ignoring accuracy. A wrong description is worse than no mention.
- Buying any service that promises a guaranteed AI mention.
How can Vatam help?
Our AI visibility tracking and reporting runs a fixed prompt panel across the main assistants every month and explains what changed. To learn what drives the results, read our guide to getting cited by ChatGPT, Gemini and Perplexity.
Sources and further reading
- OpenAI: crawler documentation
- Perplexity: crawler documentation
- Google Search Central: AI features and your website
- Pew Research Center: Google users and AI summaries (July 2025)
Want a baseline? Request a free AI visibility audit at /contact/ and we will run the first test for you.
Frequently asked questions
How often should I run the test?
Monthly is enough for most businesses. AI answers change over time, so a regular check shows the trend. Running it daily mostly records noise.
Why do I get different answers each time?
AI assistants can vary their wording and sources between sessions, and their results may differ by location and settings. Ask each question at least twice and look for what appears consistently.
What counts as being recommended?
The strongest result is being named in the answer and linked or cited as a source. A mention without a link is weaker, and being described wrongly is a problem to fix. Record all three.
Can I pay to be recommended by ChatGPT or Gemini?
Not through a guaranteed arrangement. Be cautious of anyone who sells a guaranteed mention. Visibility comes from clear facts on your site, trusted mentions elsewhere and useful content.
What do I do if the AI gets facts about my business wrong?
Fix the source first: correct your website, Google Business Profile and directory listings so the facts agree. Then check again in a few weeks, because assistants update at their own pace.