Start with what one answer can prove
One check can still help. It shows what one product returned at one time under one set of conditions.
It cannot prove stable visibility. A brand named once will not appear every time. A brand missing once may still appear for another user, place or date.
Save these details:
| Field | Record |
|---|---|
| Product and mode | ChatGPT Search, Google AI Overview, Perplexity or another named surface |
| Prompt | Exact wording and punctuation |
| Context | New conversation or prior messages |
| Location | Country, city or “not controlled” |
| Account state | Signed in, signed out, memory or personalization where known |
| Model | Named model if the interface exposes it |
| Observation | Brand present, absent, cited or mentioned |
| Evidence | Screenshot, source links and timestamp |
| Unknowns | Inputs the interface does not expose |
Now the screenshot shows a dated result, not a claim about every AI answer.
Small changes to the question can change the answer
Small wording changes can make the tool understand a different task.
OpenAI says ChatGPT Search may rewrite a prompt into one or more searches. “Who is the best provider?” and “Which provider handles this special need?” may use different sources.
Earlier messages matter too. A follow-up can carry over a location, budget, industry or need. Memory may also affect a rewritten search when it is turned on.
For a fair comparison, start a new conversation and use the exact same question. Test follow-up questions in a separate group. Do not mix both groups into one trend.
| Prompt condition | Why it matters | Control |
|---|---|---|
| Wording | Changes the task or qualifying detail | Freeze the exact prompt |
| Conversation history | Supplies context to the follow-up | Use a new thread or record prior messages |
| Language | Changes sources and phrasing | Hold the language constant |
| Memory or personalization | May affect interpretation | Record known account settings |
Location and account settings can change local answers
Location matters much more for “who should I hire?” than for a general definition.
ChatGPT Search says local results may use an IP address, device location, VPN or network location. Google also uses location and personal settings in Search.
“Who provides emergency plumbing?” needs a place. Do not compare a test in Charlotte with one run through a VPN in another country unless location is the point of the test.
Note whether you typed the place, the product guessed it or the location was unknown. Test each market separately for a multi-location business.
Signed-in and signed-out results may differ. Neither is fake. They are simply different test conditions.
The tool may use different sources next time
AI search does not use one fixed list of pages.
Google says its AI features can run several related searches and use links from the wider results. OpenAI can also rewrite a question and run several searches. Perplexity says it searches current web sources before writing an answer.
The available source set can change because:
- a page was published, updated, removed or became inaccessible;
- an index or ranking system changed;
- the rewritten or related queries changed;
- the platform chose different supporting sources;
- a temporary incident affected retrieval or display.
| Variation layer | What can change | Evidence available to the tester |
|---|---|---|
| Query interpretation | Rewrites and related searches | Sometimes visible only through different results |
| Search index | Eligible and indexed pages | Platform tools and page checks |
| Retrieval | Sources selected for the run | Citation and source panels where shown |
| Synthesis | Wording, emphasis and recommendation | Captured answer |
| Time | Page, index and platform state | Timestamp and change log |
Open the cited page. A title in the source panel does not prove that the page supports every claim in the answer.
Product updates can change the answer
The product name does not tell you every part of the test.
Perplexity Pro can offer several models and search modes. ChatGPT and other tools change models and features. Google changes Search and its AI features. Release notes may explain a sudden shift, but not every change.
Save the model or mode when the product shows it. Leave it unknown when the product hides it.
Note known releases and incidents beside the test. A change after an update shows timing, not cause. You need stronger proof before saying the update caused it.
Say exactly what you know: “The inclusion rate changed after the dated release.” Do not pretend the cause is settled.
A citation can still support a wrong answer
Citations make an answer easier to check. They do not make it correct.
OpenAI says ChatGPT can give wrong or misleading answers, including fake citations. A real source can also be old, incomplete or unrelated to the claim.
Check:
- whether the cited page opens;
- whether it supports the nearby statement;
- whether the information is current;
- whether the source is primary, independent or commercial;
- whether the answer added a conclusion the source does not support.
A citation gives you something to check. It does not promise accuracy or a future recommendation.
Keep the test conditions the same
Repeated tests become useful when you set the conditions and scoring rule first.
| Test card field | Decision |
|---|---|
| Buyer question | Freeze exact wording and intended user task |
| Product and mode | Test one named surface per result row |
| Location | Fix or record the market and location method |
| Account/context | New thread, known settings and no hidden follow-up context |
| Criterion | Define present, absent, cited, accurately described or recommended |
| Sample | State the number of runs and dates; use no universal run count |
| Change log | Record platform, model, page and campaign changes |
| Interpretation window | Decide when the evidence will be reviewed |
OpenAI’s guide to evaluations uses set data, criteria and graders. A business visibility check can be simpler. It still needs one exact question and a clear scoring rule.
Run enough tests for the choice you need to make. A quick check and a quarterly benchmark need different sample sizes. State the limit instead of hiding it behind a percentage.
Say what is known, likely or still unknown
Most changed answers do not come with a full explanation.
Use four levels:
| Confidence label | Meaning | Example |
|---|---|---|
| Documented mechanism | The platform states that the input can affect results | ChatGPT Search may use approximate location |
| Observed correlation | Two dated changes appear together | Source inclusion changed after a page update |
| Plausible inference | Evidence supports an explanation but does not prove it | A new source may have changed synthesis |
| Unknown | Available evidence cannot separate the causes | Same setup returned different wording |
A general rule does not prove what happened in one test. “Location can affect local answers” is supported. “Location caused this answer to change” needs a controlled test.
These labels show the reader where certainty ends.
An AI answer is not the final business result
Appearing in an answer shows visibility. The booking, CRM or order system shows the business result.
For a home-service company, save the answer about a service and market. Count the job only when the operating system confirms it. For B2B SaaS, save the category or use-case answer, then look for a qualified demo and pipeline movement.
The same separation applies to Healthcare, Legal and eCommerce. Healthcare work must protect patient privacy. Legal work must protect confidentiality and follow the rules of the relevant state.
One AI check cannot prove that it caused a call, appointment, case, sale or order. Connect the numbers where you can, but do not call assigned credit proof of cause.
Save the answer before you explain the pattern
The AI Visibility service covers the wider work. The AI measurement guide explains how to save and compare a set of questions. The llms.txt guide explains why crawler access does not prove visibility.
Mindflow’s method starts with what can be checked and one clear next step. If you already have a saved answer, contact Mindflow with the question, tool and date. Do not send private customer or account data.