How to manually check your company’s visibility in ChatGPT and other AI tools

Before I try to improve OptFor.AI’s visibility, I need a baseline. Do AI tools know the company? Do they recommend it to potential clients? If they do, why?

The first measurement does not require a specialist platform. A spreadsheet, a list of customer questions and access to several popular AI tools are enough.

A manual test cannot replace regular monitoring across a larger sample. It can reveal the main problems quickly and help define a better method for later measurements.

I will begin with five tools:

• ChatGPT, because it can search the web automatically and add citations to an answer.
• Gemini, because it can show sources and content related to an answer.
• Microsoft Copilot, because its web search uses Bing.
• Claude, because it can analyze current web pages and include citations when web search is enabled.
• Perplexity, because it operates as an AI search engine and builds answers from online sources.

The purpose is not to decide which tool is best. Each may use different models, search mechanisms and sources. A company may be visible in one system and completely absent from another.

The Search, Web Search or equivalent web-browsing function lets a tool retrieve information from available pages before preparing its answer.

Without search, an answer may rely mainly on the model’s existing knowledge, the earlier conversation and information contained in the prompt. That result does not reliably show what the tool can find about a company online on the day of the test.

When search is active, the system interprets the question, creates search queries, finds material, selects some sources and constructs an answer. ChatGPT may start searching automatically. Microsoft Copilot creates a short query and sends it to Bing. Claude provides citations after using web search, while Perplexity answers from material found online. The official documentation from OpenAI, Microsoft, Anthropic and Perplexity describes these mechanisms.

I will therefore test two variants:

• an answer without explicitly requesting search,
• an answer with search enabled or explicitly requested.

The first variant helps assess whether the model already associates the company with the topic. The second shows whether it can find the business and justify a recommendation using available web pages.

I should not begin with: “Do you know OptFor.AI?” That prompt supplies the brand name and tests only whether the tool can recognize it.

The test needs questions a potential client might ask, such as:

• Which companies in Poland help organizations prepare to work with AI?
• Recommend companies that audit whether a website is ready for AI systems.
• I need a partner to transform a software delivery process with AI. Who should I consider?
• Which companies help improve brand visibility in ChatGPT?
• Compare companies that offer AI consulting to medium-sized and large organizations.
• Who would you recommend to design and deliver AI-assisted software?

The questions should reflect the company’s real services, markets and clients. Visibility for an unrelated query has no business value.

Ten to fifteen questions are enough for a first test. They should cover several intentions, including finding suppliers, creating a shortlist, comparing companies and choosing a specific partner.

Each question should start in a new conversation so that previous answers do not affect the next result.

I record:

• the tool and model,
• the date of the test,
• the exact wording of the question,
• the language and specified location,
• whether search was active,
• the full answer,
• the sources shown.

This matters because results may depend on time, location, account settings and conversation context. One answer is not a reliable measurement by itself.

I will run every question three times under the same conditions. Three repetitions do not constitute a statistical study, but they help distinguish a one-off response from a result that appears more consistently.

The first check is simple: does the company name appear in the answer?

A yes or no result is not enough. I also record:

• where the company appears on the list,
• whether it is recommended or merely mentioned,
• which service the tool associates with it,
• whether the description matches the actual offer,
• whether the system confuses it with another organization,
• whether the recommendation fits the client described in the question.

The most useful result is not a brand mention alone. The company should be presented as a suitable choice in a situation where it can genuinely help the client.

When an answer includes the company, I ask a follow-up question:

Why do you recommend this company? Identify the specific information that led to the recommendation.

This helps reveal how the tool understands the brand. Its reasoning may refer to a specialization, service description, experience, completed work, client reviews or inclusion in external rankings.

I classify every reason as:

• accurate and confirmed,
• vague or unclear,
• false or impossible to verify.

If AI recommends a company for the wrong reason, the result may look positive in a report but does not represent useful visibility. It may also direct unsuitable prospects to the company.

My next question is:

Which sources support your recommendation of this company? Provide the page URLs and state which piece of information each source confirms.

I then open every link and check:

• whether the page contains the claimed information,
• whether it concerns the company being tested,
• whether it describes the relevant service,
• whether it is current,
• whether it belongs to the company, a client, a partner, a publisher or a directory,
• whether the citation supports the recommendation or is only loosely related to it.

Source lists require careful interpretation. Gemini may show content related to an answer, while its double-check function searches for pages that are similar to or contradict generated statements. Google notes that these pages are not necessarily the sources used to create the original response. The distinction is explained in the Gemini documentation.

A visible source is not automatic proof that it caused the recommendation. It is still a useful indication of which pages the system associates with the company and its offer.

The following columns are enough for the first audit:

This spreadsheet will not yet explain how to improve visibility. It will show where the main problem lies:

• the tool cannot find the company,
• it finds the company but misunderstands the offer,
• it understands the offer but lacks grounds for a recommendation,
• it recommends the company using weak or incorrect sources.

That is a more useful baseline than counting brand mentions alone.

I will run this test for OptFor.AI and retain the complete answers and their sources. I will group the results by tool, question type and search mode.

Only then will I decide what to improve first: the way services are described, the content on the website, the technical accessibility of the material or the company’s presence in external sources.

The next step will be to publish the baseline results, including weak results if that is what the test reveals.