Key Takeaways
- The principle: ChatGPT serves both as a testing tool and as an assistant for analyzing one's own responses, provided that the two roles are kept separate in distinct conversations.
- Settings: Cache disabled, no custom instructions, web search enabled, new conversation for each query.
- The five tests: brand awareness, unbranded recommendations, cited sources, factual accuracy, page review.
- The limit: ChatGPT is just one of six engines; its responses do not predict those of Perplexity or Google's AI insights.
ChatGPT is both the most widely used AI engine in France and an analytical tool capable of providing feedback on its own responses. When used properly, it enables an initial visibility audit. GEO in one hour, with no subscription required. If used incorrectly, it produces flattering but false conclusions. Here are the prompts that work, the required settings, and what this exercise cannot measure.
The settings without which the test is incorrect
Since 2025, ChatGPT has been storing conversations and tailoring its responses based on what it knows about the user. An executive who has previously described their company in a conversation will receive a recommendation biased in their favor. Before running any tests:
| Adjustment | Why |
|---|---|
| Memory Disabled (Settings, Personalization) | Eliminates the influence of past conversations on responses |
| Empty Custom Instructions | Prevents a profile that says “I run an SEO agency” from influencing the answer |
| Web search enabled | Without conducting a search, the answer comes from training data dating back several months |
| Temporary chat or new conversation per request | A previous question about your brand affects the subsequent ones |
| The same model for all tests | The results differ between the default model and the reasoning models |
Ideally, have the audit conducted by someone who has never mentioned the company to ChatGPT, or from an account created specifically for this purpose. The general audit protocol, which applies to all engines, is described in audit its visibility on AI search engines.
Test 1: What ChatGPT Knows About Your Brand
Prompt: “What do you know about [brand]? Describe its activities, offerings, pricing, and location, citing your sources.” Note three things: accurate facts, false or outdated facts, and missing facts. A vague answer (“a French company in sector X”) means that the site is not a sufficiently clear source for the brand itself; this is the most common scenario for small and medium-sized businesses and can be corrected first by updating the About Us page, see Optimize Your "About Us" Page for AI.
Test 2: The Recommendation Without Mentioning the Brand
It’s the test that counts. Ask ten questions the way a prospect would phrase them: “Which service provider for [need] in [city]?”, “What tools are available for [problem] for an SME?”, “Compare the top three solutions in [category].” For each answer, note the brands mentioned, their order, and the sources cited. Repeat each question three times in three separate conversations: a brand mentioned once out of three times is a contested brand, not an established one. The query formulations to be tested are constructed using the method of Keyword research for GEO.
Test 3: What sources does ChatGPT use for your topics?
After each answer in Test 2, ask: “List all the pages you consulted to answer the questions, including their URLs, and indicate which one you used for each statement.” The response isn’t always exhaustive, but it reveals the dominant domains within your topic: media outlets, comparison sites, forums, competitors, and Wikipedia. In 2026, for B2B search queries in French, comparison sites and blog posts from publishers often accounted for more than half of the sources cited; an SME not featured on these third-party pages is unlikely to be recommended. The list of identified domains becomes your roadmap for media relations and partnerships.
Test 4: Accuracy and Tone
Prompt: “What are [brand]’s weaknesses? Are there any negative reviews?” followed by “Is [brand] suitable for [use case]?” Here, you’re assessing reputational risk: an AI that picks up on an isolated review from 2022 or confuses your company with a namesake can turn prospects against you. Record each false claim along with its likely source; corrections must come from the source, not the AI engine. Ongoing monitoring of this risk is covered in monitor its presence in generative AI.
Test 5: Have the search engine evaluate a page
In a separate conversation, paste the text of a strategic page and ask: “If a user asked the question [query], which passages from this page could you quote verbatim? What information is missing for you to consider it a reliable source? ” The answers are remarkably consistent across tests: lack of numbers, the answer buried after 400 words of introduction, paragraphs that don’t make sense on their own, and an unidentified author. This test doesn’t measure visibility—it measures extractability—and it takes five minutes per page.
What This Protocol Doesn't Say
| Limit | Consequence | Supplement |
|---|---|---|
| A single engine | Perplexity, Gemini, and Google’s AI insights cite other sources and use different criteria | Repeat Tests 2 and 3 on at least three other engines |
| No volume | ChatGPT doesn't say how many people ask the question | Cross-referencing with Search Console and referral traffic in GA4 |
| Self-Assessment | The model refines its answers after the fact; the list of sources is sometimes incomplete | Check the URLs provided; ignore the explanations of "reasoning" |
| One-time statement | No trend without regular repetition | Switch to a monthly or weekly tracking tool |
To move from a one-time test to ongoing monitoring, a tool that automatically queries multiple search engines using a fixed set of search terms is more reliable than manual data collection; the key metrics to track are detailed in GEO's KPIs.
How GreenRed Can Help You
Instead of juggling multiple tools, GreenRed's GEO Tracking module brings these metrics together in a single dashboard, compares them over time, and shows you which actions to prioritize. You can try it for free—no credit card required—on the page Rates.
Frequently Asked Questions
Is the free version of ChatGPT sufficient for this type of audit?
Yes, for brand and recommendation tests, provided that web search is enabled—which is the case for the free version with a limited quota. A paid subscription offers larger quotas, a choice of models, and unlimited temporary chats, which are useful for repeating each query three times without cross-contamination.
Why not just ask ChatGPT directly, “Am I properly indexed on your platform?”
Because the model doesn’t know its own statistics and will respond with a reassuring generalization. It has no way of knowing how often it mentions a particular site. The only valid approach is to ask prospects questions and observe their responses, without ever mentioning the brand in the question.
Should we use the API instead of the interface for auditing?
The API allows you to automate hundreds of requests and store the responses, but its results differ from those of the consumer interface: search tools, layout, and system instructions are not identical. To reflect what users see, the interface remains the standard; the API is used for regular, large-scale monitoring.
Can you correct false information provided by ChatGPT about your company?
Not directly. The model reproduces what it reads in its sources or what it has learned during training. The fix involves publishing the correct information on the site’s best-indexed pages, having the incorrect third-party pages corrected, and waiting for the next crawl; this can take anywhere from a few days to a few weeks.