How to Check What AI Says About Your Business

What does checking your business in AI actually tell you?

A useful audit starts by choosing an observable outcome. Recognition, factual accuracy, citations and supplier recommendations answer different questions. Keep them in separate columns rather than combining them into a single unexplained reputation score.

Microsoft makes a related distinction in its Bing Webmaster Tools AI Performance documentation: citation metrics describe source appearances, while the aggregated data does not establish a page's ranking or role in an individual answer. Use the meaning of the metric, not its label alone, when reporting results. Source: Microsoft Bing, AI Performance.

If your main concern is missing supplier recommendations, use our guide to why AI recommends competitors instead of your company. Here, the starting point is the description of your business and the evidence supporting it.

Write down the company facts you will check

Prepare a private reference sheet before asking the first question. Keep it outside the conversation used for the name-only check. Otherwise, you may supply the very information you intended to test.

• Identity: public business name, relevant legal entity and genuine website.
• Location: headquarters or office location, distinguished from the areas you serve.
• Activity: current products, services and intended customers.
• Timeline: important changes, with their effective dates.
• Evidence: a relevant public URL and passage for each fact you want to verify.

Match the evidence to the claim. A company page may explain a current service, but an asserted certification should be checked with the relevant issuing body. A registration record may establish legal identity without establishing that a particular service is available today. Record what each source actually supports.

Separate things you know internally from things a reader can verify publicly. An unpublished product plan may be accurate internal information, but its absence from a public-facing answer is not, by itself, an error.

Choose a manageable set of products and questions

Start with the AI products relevant to your customers. You can use ChatGPT, Gemini and Claude as examples, without assuming they use identical models, settings or sources. Record the product and displayed model or mode for each check instead of treating the brand name as a complete technical specification.

Use the same language and company spelling for the initial comparison. If location is needed to distinguish the business, decide that in advance and include it consistently. Treat later tests of abbreviations, former names or another country as separate variants.

Start with a company-name check before supplying your website

Open a separate conversation and ask about the name alone, adding the country if that is part of your planned test. Avoid first telling the assistant what the business does, who founded it or which achievements it should mention.

What can you tell me about [company name] in [country, if needed]?

Describe its main activity and location if you know them.
If the name could refer to more than one business, say so.
If you do not have enough information, state that uncertainty.
Do not guess missing details.
For this check, do not search the web or open websites.

The final instruction expresses the condition you want; it does not prove the product complied. Where an explicit search or tool setting is available, record it. Inspect visible activity. If you cannot verify the retrieval conditions, label the result accordingly rather than calling it a controlled test without web access.

Anthropic's guidance on hallucinations recommends allowing uncertainty and checking claims against quotations and sources. That supports a cautious review process; it does not make this prompt immune to incorrect answers. Source: Anthropic, Reduce hallucinations.

Keep personalization and supplied context visible in your notes

Record whether you used an ordinary chat, a project, connected material or a workspace with additional instructions. Note any relevant personalization settings you can inspect. A new conversation is a useful starting point, but do not describe it as proof that the product had no other context.

If the response asks for your domain, save that original response before providing it. The next answer will be a test with additional identifying information. If no useful description appears, our guide to why ChatGPT does not recognize your company explains what that observation does and does not establish.

Run a separate check with web search and sources

Now investigate what the product can find through retrieval. Use a separate conversation so the correction you may have supplied in the first exchange is not silently included in the second. Keep the business name, country and language consistent.

Search the web for current public information about [company name]
in [country, if needed].

Identify the business and its main activity.
Link the sources supporting the factual claims you make.
Distinguish the business from other organizations with the same name.
Note conflicting information and relevant dates.
If search or a source is unavailable, say so.
Do not treat an allegation as an established fact.

What to inspect in ChatGPT

OpenAI's documentation says ChatGPT on the web shows search results and citations when it uses web search; the desktop app records search activity in the transcript. Inspect those indicators and the linked pages. Available search functionality can depend on workspace settings. Source: ChatGPT, Web search.

What to inspect in Gemini

Google explains that Gemini may show sources inline or beneath an answer, and that not all responses include sources. Related material may include public pages or supplied and connected documents. Record the source type as well as the URL; a missing Sources button only tells you that no links were provided for that response. Source: Google, Gemini sources.

What to inspect in Claude

Anthropic describes search indicators and citations for Claude's web-search responses. Its current help page also distinguishes experiences with a search toggle from a newer experience where Claude searches when useful. Check the interface you actually have instead of assuming every account exposes the same control. Source: Anthropic, Web search in Claude.

If you provide the website, label that as another condition

You can supply the correct domain to investigate a confused identity or ask the assistant to summarize a particular page. That is useful, but it changes the task. Label it “website supplied” and record whether the page was actually retrieved or whether you pasted its text.

Do not infer a model update from a better answer after adding the domain. You have observed a response with more context. Preserve both versions so the improvement remains understandable.

Check identity, dates and evidence before judging the answer

Break the response into statements you can inspect. “The business is based in one city and provides a particular service internationally” contains more than one claim. Verify the location and service coverage separately.

1. Check the entity. Compare the domain, activity and location. A familiar name is not enough to establish the right business.
2. Check the passage. Open the cited page and find the text supporting the claim. A relevant-looking title is not the evidence itself.
3. Check the date. Separate when a page was published from when the described fact applied.
4. Check the strength of the statement. Does the answer turn a possibility into certainty or a complaint into an established finding?
5. Record uncertainty. If the source is inaccessible or inconclusive, keep that limitation in the result.

Google warns that Gemini can misrepresent its own operation, including explanations of its training or use of sources. Treat an assistant's account of where it learned a fact as another claim to verify, not as an audit log. Source: Google, About Gemini responses.

Review this answer about [company name]:
[paste the answer]

Evidence available for this review:
[public URLs and relevant passages]

List the factual claims individually.
For each, identify the supporting or conflicting passage, if any.
Use: supported, contradicted, unclear, or not checked.
Keep identity mismatches and date differences explicit.
Do not invent sources or treat missing evidence as proof of falsity.
State whether your review uses retrieved pages or supplied excerpts.

This prompt organizes a review; it does not replace your verification. Inspect the quoted passages yourself, especially before publishing a correction or responding to an allegation.

A promotional description can still contain a serious identity error. A cautious description can be accurate. Assess factual support separately from whether the tone sounds flattering.

Use a simple AI business-audit worksheet

Create one record for each answer and link its claims to the same saved response. Use the following fields in a spreadsheet or document. This is a blank framework, not a dataset or a result from an actual business.

Define your labels before using them. “Supported” means supported by the material reviewed, not certified true for all purposes. “Contradicted” needs conflicting evidence. “Unclear” means the available material does not settle the point. “Not checked” means you have not completed that review.

Keep technical failures outside those labels. A timeout, inaccessible product or usage-limit message is not evidence that the model does not recognize the company. Record the failed check rather than counting it as a negative description.

Do not use agreement as a substitute for evidence

If several answers repeat the same fact, inspect their supporting material. Repetition is an observation about the answers, not independent corroboration. If they disagree, identify the precise contradiction rather than choosing the majority response automatically.

Likewise, counts of mentions depend on which questions and products you tested. State the sample and counting rule before reporting a percentage. A handful of checks is not a measurement of everything customers see across AI services.

What should you do when AI gets your business wrong?

Prioritize errors by their practical effect. A wrong contact route, an incorrect closure statement or an unsupported serious allegation may deserve attention before a minor wording preference. This is an editorial prioritization method, not a legal assessment or a severity score produced by the models.

• Wrong business: compare identity details using our guide to AI confusing your company with another company.
• Old information: establish the timeline and follow the guide to correcting outdated company information.
• Scam or fraud allegation: preserve the claim and investigate the sources behind the accusation.
• A claim that needs escalation: review the different routes for reporting or requesting removal of incorrect information from ChatGPT.

When your own page is wrong, correct the specific fact and retain a clear timeline where relevant. When another publisher is wrong, use its correction process with the exact passage and supporting evidence. When an AI summary misreads an accurate source, show that mismatch in the response report.

A missing description needs investigation too, but it is not the same as a false statement. Avoid purchasing a promised fix solely because one response was incomplete. First establish what you can actually reproduce and what information is missing.

How often should you check what AI says about your business?

Choose a cadence that fits the business and the importance of the issue. Recheck after a material change such as a new name, location or service, and after a source correction you want to evaluate. The sources cited here do not establish one universally correct monitoring interval.

Reuse the saved question and record any changed conditions. If the product, model, mode, personalization or source access differs, note that beside the new answer. Compare the specific disputed claim, not just the overall tone.

If you rerun a check, retain every result you plan to include in the comparison. Do not keep generating answers until a favorable one appears and then report only that version. Work within product usage limits and avoid unnecessary repeated calls.

Separate the outcomes: the public page was corrected; the assistant cited that page; the new answer stated the current fact. One does not automatically establish the others. A changed answer after an edit is an observation, not proof that the edit caused the change.

Where ReputationScan fits into an AI business audit

ReputationScan queries its tested Claude and Gemini models about a company by name, with the country when supplied. Those queries have no web tools and do not receive the company's website. The site collects website information and public identity signals separately for comparison.

This can provide a starting record of descriptions and identity findings. It does not test the ChatGPT consumer product, reproduce every consumer app configuration, measure all customer conversations or determine a company's share of supplier recommendations.

Read the actual answers and evidence, the reported model coverage and the original scan time. Not every model request necessarily produces a usable answer. Repeated requests can return a cached report, and automated assessments can be wrong.

See how the scan works and what its checks cover before comparing it with your manual checks.

ReputationScan refers interested businesses to SecondSideMedia, a partner that publishes corrective public records, and may earn a commission from resulting business. A scan does not change AI answers or guarantee recognition.

Your AI business-reputation audit checklist

The useful output is a record you can investigate: what was said, under which conditions, what supports or contradicts it and what you will do next.

Sources and review notes

Sources checked on 1 October 2026. The worksheet, prompts and checklist are editorial tools. No business experiment, ranking improvement or universal monitoring interval is claimed. Scanner descriptions were checked against its implementation.

1. Microsoft Bing — AI Performance in Webmaster ToolsWhat citation metrics measure and their limits.
2. Anthropic — Reduce hallucinationsUncertainty, evidence and verification of generated claims.
3. ChatGPT — Web searchSearch activity and citations in ChatGPT.
4. Google — View related sources from Gemini AppsRelated links, source types and responses without source links.
5. Anthropic — Enable and use web searchClaude search indicators, citations and interface differences.
6. Google — Learn about responses from Gemini AppsFactual errors and unreliable explanations of the model's own operation.