Plain-English guide
AI Visibility Is Not Just ChatGPT Visibility
Short answer: use ChatGPT as one observation. Choose the other contexts your buyers have reason to use, and keep the product, model, sources, and date attached to each check.
Start with the buying decision, then choose the tools
It is easy to turn AI visibility into a list of logos. A small team can spend all its time checking more tools while learning very little about why a buyer would choose the business. Start with the decision you want to understand: finding a local provider, comparing software, or checking whether a company can handle a particular job.
Our AI visibility guide separates being mentioned, being cited, and being recommended. Use those definitions across your checks. A brand name appearing in a long list is different from an answer explaining why the business fits the customer's constraints.
Sample — illustrative, not measured buyer behavior: a small accounting firm might prioritize “Who can help a plumbing business catch up on bookkeeping?” over “Who is the best accountant?” The narrower question gives the team a useful standard for evaluating the answer. Does it address that kind of customer and that kind of work?
A product name does not describe the whole check
Write down where the question was asked. Was it a consumer chat app, an API request to a named model, or a Google Search experience? Then record which tools or web evidence were available. Without those fields, two answers carrying the same company name can look more comparable than they are.
OpenAI documents that ChatGPT can search the web automatically or through a user-selected search option, and that search responses may include citations. Anthropic separately documents a web search tool for the Claude API. Those are specific capabilities with specific contexts. Neither description establishes that an arbitrary API response reproduces a person's consumer-app conversation.
For a lean team, the practical move is simple: add a “test context” column before adding another score. If you cannot identify the model in a consumer product, say it was not disclosed. Do not guess a version to make the report look complete.
Separate Google Search from Gemini observations
Google's Search documentation treats AI Overviews and AI Mode as Search features. It says those features may use different models and techniques, producing different responses and links. Calling all Google-related checks “Google rank” would hide a distinction that matters when a team chooses what to fix.
Google also documents Search grounding for the Gemini API: developers can connect a model to current web content. That describes a way to build an application. It does not make every Gemini API response a captured AI Overview.
Use a specific label such as “direct model check with application-supplied web evidence” when that describes the method. If someone manually observes an AI Overview, store it separately with the query and observation conditions. Our Gemini guide walks through this distinction in more detail.
Google: AI features and your website
Pick coverage you can maintain
Ask the sales or customer-service team what buyers say they used while researching. Review any available referral evidence and ask new customers a neutral discovery question. These are inputs to prioritization, not a representative market survey. A few anecdotes can justify a test; they cannot establish market share.
Choose a small set of important questions and the contexts most relevant to those decisions. Give the set an owner and a repeatable review schedule. Expand when an additional context changes a decision, exposes a meaningful accuracy issue, or answers a customer question you cannot currently address.
Avoid equal spending for the sake of a symmetrical dashboard. If your team cannot review the answers it already collects, more checks will mostly create a larger reading queue. Coverage is useful when it leads to better evidence and a practical next action.
Compare the same need across contexts
Keep the core buyer question and constraints consistent. Save the answer text, date, product or API route, model when known, web-evidence method, named brands, and visible citations. Record a failed or incomplete request separately from a completed answer that omitted the brand.
Sample — illustrative: one completed check names your accounting firm without describing its industry experience; another recommends a competitor for construction businesses; a third fails to complete. The first two are answer observations. The third is unavailable evidence. Reporting all three as “not recommended” would misstate the test.
Review the actual wording before scoring it. A business can appear as an option the buyer should avoid, an alternative for a different budget, or a strong fit. Keep a short reviewer note explaining the classification so a colleague can reproduce your judgment.
Use disagreements to investigate evidence
When answers disagree, start with the simplest useful questions. Did the question change? Did the model or search method change? Are the visible sources discussing the same business? Does one answer contain an unsupported claim? The disagreement itself does not tell you which system is right.
Open the cited pages and compare them with the facts you can verify. If the business website does not explain a service the answer discusses, publish an accurate explanation. If an outside profile is stale, ask its owner to correct it. If the answer is unsupported and no source is visible, record the limitation instead of inventing a cause.
Do not rewrite the business to please each answer independently. Choose a shared improvement that helps real buyers: clearer service scope, current pricing context, stronger approved proof, or a fair explanation of fit and limitations. Then record what shipped before reviewing later checks.
Report a useful pattern without hiding the sample
For a simple report, show each context separately alongside the number of completed questions checked. If you summarize inclusion, state the numerator and denominator and define what counted. Do not describe a result from your selected prompts as the percentage of all buyers who see the business.
Keep historical model labels and method changes visible. If the test setup changes, explain the boundary rather than presenting the new result as directly equivalent to the old one. Repeated observations can support a pattern in the test set; they do not prove a content change caused it.
In The Answer offers buyer-question monitoring and saved answers with model and date information. Your team reviews the facts and publishes changes. Read the methodology before comparing plans so the scope of those checks is clear. The objective is a decision you can defend, not a claim that you have measured every AI experience.
What to do next
- Choose buyer questions tied to an actual decision.
- Label consumer products and direct model checks separately.
- Save complete answers and mark unavailable checks explicitly.
- Review disagreements and assign one supported improvement.
Sources
Conclusion
You do not need to monitor everything to learn something useful. You need a clear question, a well-labelled check, and enough evidence to decide what to improve.
Frequently asked questions
How many contexts should a small team check?
Choose the ones relevant to your buyers and maintain a question set you can review. Expand when the additional evidence changes a decision.
Can I combine all results into one score?
Only with an explicit definition and visible denominators. Keep each context available separately so the combined figure does not hide the method.
