How Do I Know if AI Is Recommending My Business?
AI Search Visibility
Test the real questions buyers ask without naming your company. Record whether the engine names your business, which competitors appear, and which sources support the answer. Then repeat the test under consistent conditions.
Asking ChatGPT, “What is [my company]?” may confirm that information about your brand exists. It does not tell you whether the brand appears when a buyer has not already chosen it.
Recommendation visibility begins with an unbranded question:
Which companies offer [service] for [buyer, need, or location]?
Brand recognition is not recommendation visibility
Suppose a roofing company asks ChatGPT, “What is Smith Roofing?” and receives a polished summary.
That result feels good. It may even be accurate. It still leaves the important question unanswered:
If a homeowner asks for reputable roofers in the area, does Smith Roofing appear without being named first?
The difference is intent.
A branded prompt tests recognition. An unbranded buyer prompt tests whether the company enters the consideration set.
Both can be useful, but they do not measure the same thing.
Start with the questions that can lead to a decision
A useful baseline does not begin with every possible variation of a keyword. It begins with the moments when a buyer needs an answer.
For a manual diagnostic, start with 10 to 20 questions across these five groups:
Discovery questions
- Which companies provide [service] in [location]?
- Who helps [type of buyer] with [problem]?
- What are the best options for [specific need]?
Use-case questions
- Which [service] providers work with [industry]?
- Who can handle [constraint, integration, or timeline]?
- What is a good option for a business that needs [outcome]?
Comparison questions
- What are the best alternatives to [known provider]?
- Which is better for [use case], [category A] or [category B]?
- How do the leading [service] providers differ?
Trust questions
- Which [service] companies are reputable?
- Which providers have experience with [industry or problem]?
- Who is known for [specific capability]?
Location questions
- Who offers [service] near [city or neighborhood]?
- Which [category] companies serve [region]?
- What are the best local options for [need]?
These are templates, not a script to mass-produce pages. Use the language your real prospects use in calls, proposals, site searches, sales conversations, and customer-support questions.
Use a repeatable AI visibility test
1. Freeze the question set
Write the prompts before testing. Save the exact wording.
If you change a prompt every time you run it, you cannot tell whether the engine changed or the question changed.
2. Record the testing conditions
For each result, note:
- Engine and product used
- Date and local time
- Whether web search was active
- Whether the session was signed in
- Location setting, if relevant
- Whether memory or prior conversation context could affect the result
- Exact prompt
Personalization can be useful to the user, but it can muddy a business visibility test. OpenAI explains that ChatGPT Search may use memory and location when those settings are enabled.[1] Record those conditions or use a neutral session when you want a cleaner baseline.
3. Do not name your company in the discovery test
The main prompt should describe the need, audience, service, and geography without feeding the engine the answer you want.
Weak test:
Is NAMYNOT a good AI visibility agency?
Stronger test:
Which agencies help small businesses track and improve how they appear in ChatGPT and Google AI answers?
You can run branded prompts separately to check factual accuracy, but do not mix those results into the recommendation baseline.
4. Capture names and citations separately
Record every business named in the answer. Then record every cited domain.
Do not assume a cited company was recommended. Do not assume a recommended company was supported by its own website.
ChatGPT Search may display inline citations and a Sources panel.[1] Bing’s AI Performance reporting makes a similar distinction, explaining that citation counts show how often pages are referenced, not their importance, ranking, or placement.[2]
That distinction should carry into your spreadsheet.
5. Repeat the test
One run is a snapshot.
A 2026 preprint repeatedly submitted the same queries across Gemini, SearchGPT, and Perplexity in three consumer-product topics. The median overlap between the cited-source sets from repeated answers ranged from roughly 29% to 50%, depending on the platform and topic.[3]
That finding is useful, but it should not be stretched beyond the study. It covered three consumer topics, not every local, B2B, medical, legal, or professional-service market.
For a directional manual baseline, run each priority question at least three times and space some tests across different days. For formal research or claims about market share, use a much larger, precommitted sample and report uncertainty. The same 2026 preprint found that the query volume needed for tighter confidence intervals varied materially by platform and metric.[3]
6. Compare engines without pretending they are identical
Run the same question set across the platforms your customers are likely to use.
Do not combine the results into one unexplained score. ChatGPT, Google AI features, Microsoft Copilot, and Perplexity can use different retrieval systems, interfaces, sources, and personalization.
Report the platform-level result first. An overall view can come second.
7. Save the baseline before changing anything
Capture the answer text, names, cited URLs, date, and conditions before updating your website or launching a campaign.
Otherwise, a later appearance may feel like progress without giving you a defensible before-and-after comparison.
What should the worksheet measure?
CITE by NAMYNOT · Research visual
One AI search is a snapshot, not a trend.
Select a platform to explore the median similarity between cited-domain sets returned by repeated searches in a specified 2026 experiment.
Across the three tested consumer-product topics, Gemini’s reported median fell between 0.29 and 0.31.
Jaccard similarity compares two sets. A value of 1 means the sets are identical. A value of 0 means they share no domains.
What you can safely conclude: repeated runs can return materially different source sets, so visibility testing should preserve the prompt, platform, date, conditions and raw answer. These values are study results, not universal benchmarks.
Source: Ronald Sielinski, “Quantifying Uncertainty in AI Visibility: A Statistical Framework for Generative Search Measurement,” arXiv:2603.08924v2, June 9, 2026. Read the preprint.
Use one row per prompt, engine, and run.
| Field | What to record |
|---|---|
| Query | Exact unbranded buyer question |
| Engine | Product and mode tested |
| Date and conditions | Time, location, sign-in state, search mode, and relevant personalization |
| Brands named | Every business explicitly included as an option |
| Your brand named? | Yes or no |
| Your domain cited? | Yes or no |
| Named and cited? | Yes only when both occurred in the same result |
| Cited domains | Every supporting domain shown |
| Display order | Where the brand appeared, without automatically calling it a rank |
| Answer context | Positive, neutral, negative, inaccurate, or unclear |
| Notes | Missing facts, unusual caveats, or visible source patterns |
Doing this by hand works once. It stops working when you try to hold conditions steady across four engines, repeat each question enough times to tell a real change from normal variation, and keep the same comparison running month after month. That is where a manual worksheet becomes a spreadsheet nobody updates.
Four metrics a business owner can understand
These are transparent calculations, not proprietary scores.
Query coverage
The percentage of valid tested prompts in which your business was named at least once.
Formula: prompts naming your business ÷ valid prompts tested × 100
Citation coverage
The percentage of valid prompts in which your website appeared as a cited source.
Formula: prompts citing your domain ÷ valid prompts tested × 100
Named-and-cited rate
The percentage of valid prompts in which your business was named and your domain was cited in the same answer.
Formula: prompts with both outcomes ÷ valid prompts tested × 100
Competitor mention share
The percentage of all recorded business mentions belonging to each competitor.
Formula: mentions of one competitor ÷ all business mentions recorded × 100
Label this as mention share within your test set. It is not market share, revenue share, or a universal ranking.
How to interpret the result
Your company is named and cited
That is the strongest result in the framework, but verify the wording and source accuracy. A citation is not automatically positive, correct, or persuasive.
Your company is named but not cited
The engine knows or retrieves the brand, but another source may be supporting the recommendation. Strengthen first-party evidence around the question and verify what third-party sources say.
Your site is cited but your company is not named
Your content may be helping answer the question without placing the business in the consideration set. Clarify the relationship among the page, author, organization, and service.
Competitors appear and you do not
Study what supports their inclusion. The useful question is not, “How do we copy this?” It is, “What evidence makes these companies defensible answers, and what do we genuinely have that is missing or unclear online?”
No company is named
The prompt may be informational rather than commercial. Try a related recommendation or comparison question before treating the absence as a brand problem.
Do not confuse a dashboard with the diagnosis
A dashboard can show that your business is missing. It cannot tell you the complete reason without reviewing the actual answers, source pages, technical access, brand facts, and outside evidence.
That is why Cite separates the measurement from the work.
The scan shows the gap. The roadmap explains it. The strategy and execution close the parts that are within your control.
Frequently asked questions
Can I test by asking ChatGPT about my company?
Yes, but treat that as a brand-accuracy test. Use unbranded buyer questions to test recommendation visibility.
Is the first company mentioned ranked No. 1?
Not necessarily. Some answers present a clear ordered list, while others group options without a formal ranking. Record display order, but do not label it a ranking unless the response does.
How many prompts should I test?
Ten to 20 well-chosen questions can produce a directional manual baseline. That is not enough for a universal market claim. Formal research needs a larger sample, repeated measurements, a published methodology, and uncertainty reporting.
Should I combine every engine into one score?
Only if the underlying engine-level results remain visible and the weighting is disclosed. A single score can hide that a brand performs well in one system and disappears in another.
Does a citation mean the AI agrees with my page?
No. A page may be cited for one statement, contrasted with another source, or used in an answer that misrepresents it. Review the context, not only the count.
Find out what your buyers see
Run a free Cite scan. We will test one engine, show you the businesses it names, and explain your first visibility gap in plain English.
Sources
- OpenAI, “ChatGPT Search.” help.openai.com
- Microsoft Bing, “Introducing AI Performance in Bing Webmaster Tools Public Preview,” Feb. 10, 2026. blogs.bing.com
- Ronald Sielinski, “Quantifying Uncertainty in AI Visibility: A Statistical Framework for Generative Search Measurement,” arXiv:2603.08924v2, June 9, 2026. arxiv.org