How to Check Whether ChatGPT Mentions Your Brand: Which Prompts to Use and How Many Times to Ask
Five prompt types for checking your brand in ChatGPT by hand, margins of error by repeat count, a table of the odds your brand never shows up, temporary chat and search settings, a logging sheet and a seven-step routine, based on SparkToro and Evertune experiments and OpenAI documentation.
"Can I just ask ChatGPT whether it mentions our brand?" That is usually the first thing a marketer tries when they start looking at GEO. The CEO sends a screenshot from last night showing only competitors. Someone on the team asks again and this time your brand comes up third. Nobody has grounds to say which screenshot is right. Checking by hand is fine. Without a fixed set of prompts, a set number of runs and a controlled setup, though, all you have is an anecdote in one screenshot. This guide covers which prompts to ask, how many times, and under what settings so the results hold up, plus a logging sheet and a seven-step routine.
30-second definitions: five terms for manual checks
A manual check is a measurement method in which a person types a fixed set of prompts into ChatGPT and records by hand whether the brand appears.
A category prompt (non-branded prompt) is a question about a product category or situation that leaves out any brand name, which is how real customers ask for recommendations.
A branded prompt is a question that names the brand directly, used to check whether the description is accurate rather than whether the brand appears.
A repeated run is a collection method in which the same prompt is asked several times in new conversations and the varying results are read as an average.
Mention rate is the share of all responses in which the brand appears at least once, and it is the first metric to calculate in a manual check.
SEOcrawl lists the metrics for AI answer visibility as prompt coverage, mention rate, citation rate, share of voice and AI sessions in GA4.[6] Of these, a manual check can realistically produce mention rate and cited sources.
Can you check by asking ChatGPT yourself? The short answer
Yes, but the results only count as evidence if you meet three conditions. Your prompts should be mostly category prompts without the brand name, you need to ask each one several times, and every run has to happen in the same setup.
The reason is that ChatGPT gives a different answer almost every time. In an experiment SparkToro published in January 2026, 600 volunteers ran 12 prompts through ChatGPT, Claude and Google AI a combined 2,961 times. Ask ChatGPT or Google AI 100 times, and the chance that two responses contained the same list of brands was under 1%; the same list in the same order came up fewer than once in 1,000 runs.[1] So when the CEO's screenshot and the team's screenshot disagree, neither one is wrong. Each is a single sample drawn from the same distribution.
Collect enough runs and the numbers settle down. In the same experiment, a prompt about cancer care hospitals on the US West Coast returned City of Hope in 69 of 71 ChatGPT responses (97%), but it was named first in only 25 of them.[1] Rankings shuffle, while the rate of appearance becomes readable once you repeat. That is why a manual check records "how many times out of how many," not "what rank." Why AI brand recommendations change with every answer breaks the variation down into five causes.
Which prompts to use: five prompt types
Each prompt type tells you something different. Use category and situation prompts to judge visibility, and save branded prompts for checking whether the description is right.
| Prompt type | Example | What it tells you | What it can't tell you | Suggested share of the set |
|---|---|---|---|---|
| Category recommendation prompt | "Recommend leather tote bag brands for a professional in their 30s to carry to work" | Whether you make the shortlist, and which competitors appear alongside you | Whether your brand is described accurately | 40-50% |
| Situation or problem prompt | "What running shoes should I buy if I have wide feet?" | Whether you appear when customers ask in their own words | Recognition when someone searches your brand name | 30-40% |
| Comparison prompt | "Between Company A and Company B, which suits a small online store?" | How you're described next to a competitor | Visibility when the user doesn't know who to compare | 10-20% |
| Branded prompt | "What kind of company is XX? What are its main products?" | Errors in founding, product and price range descriptions | Actual recommendation visibility | Separate set |
| Source-check prompt | A category prompt with search turned on | The pages and domains the answer cites | Results when ChatGPT answers without search | Doubles as a category prompt |
These shares are not survey data. They are this guide's suggested starting point, so adjust them for your industry and goals.
INTER.R, the newsletter from Intermajor, argues that the main GEO opportunity is showing up for customer intent, such as "gift ideas for a woman in her 20s," rather than for brand-name searches.[5] Pick manual-check prompts by the same standard. Evertune warned that leading prompts that include the brand name or mix in price conditions can skew results, and advised keeping prompts narrow and consistent.[2]
Bad and good examples side by side make the difference clear.
- Bad: "Is XX a good brand?" → The brand name is in the prompt, so ChatGPT almost always mentions XX. It tells you nothing about visibility.
- Good: "Recommend just 3 brands of lightweight commuter bags that fit a laptop" → This matches the conditions and phrasing a real customer would use.
- Bad: "Which is best: XX, Competitor A or Competitor B?" → You supplied the comparison set, so you can't tell whether XX would make the shortlist on its own.
- Good: Cap comparison prompts at 10-20% of the set, and judge visibility only on category prompt results.
People also phrase the same need in very different ways. When SparkToro asked people to write their own prompts for the same goal (headphones for a family that travels often), the 142 prompts had an average semantic similarity of 0.081.[1] Even so, across 994 responses from running those prompts repeatedly, Bose, Sony, Sennheiser and Apple appeared in 55-77% of them.[1] Covering an intent with several phrasings matters more than the wording of any single prompt. Write each intent 2 to 3 different ways and include all of them in the set.
How many times to ask: margin of error by repeat count
Evertune's published analysis is the benchmark for how many runs a single prompt needs. It ran an SUV recommendation prompt in ChatGPT repeatedly and tracked Acura's mention rate. At 5 runs the mention rate was about 10%, but the margin of error was 27pp, so the true value could have been anywhere from 0% to 37%.[2] The margin of error fell to 12pp at 12 runs and about 6pp at 100 runs, with little further gain after 100.[2]
| Item | Value (pp) | Source |
|---|---|---|
| 5 runs | 27pp | (Evertune, 2026) |
| 12 runs | 12pp | (Evertune, 2026) |
| 100 runs | 6pp | (Evertune, 2026) |
SparkToro wrote that getting a real picture of an AI's recommendation list usually takes at least 60 to 100 runs averaged together.[1] Asking one prompt 100 times by hand isn't realistic, so a manual check sets a lower bar. Instead of measuring mention rate to the decimal point, the goal is to tell "almost never shows up" apart from "shows up often."
You can work out how many runs that takes from the probability of never appearing. If a brand's true mention rate is p, the chance it fails to appear in any of n runs is (1-p) to the power of n. The table below uses a binomial calculation that assumes responses are independent. Real repeated responses tend to resemble one another, so the actual values can be higher.
| True mention rate | 1 run | 3 runs | 5 runs | 10 runs | 20 runs | 30 runs |
|---|---|---|---|---|---|---|
| 10% | 90.0% | 72.9% | 59.0% | 34.9% | 12.2% | 4.2% |
| 30% | 70.0% | 34.3% | 16.8% | 2.8% | 0.1% | 0.0% |
| 50% | 50.0% | 12.5% | 3.1% | 0.1% | 0.0% | 0.0% |
A brand with a 30% mention rate still has a 70% chance of not appearing if you ask only once. Ask 5 times and that drops to 16.8%; ask 10 times and it's 2.8%. A brand at 10% has a 4.2% chance of never appearing even after 30 runs. So zero mentions in 30 responses doesn't mean a 0% mention rate. It means the rate is probably below roughly 11% (upper bound of the Wilson 95% confidence interval is 11.4%, our calculation). For confidence intervals and the response counts needed to compare periods, see AI visibility score sample size guidelines.
These calculations set the scale of a manual check.
| Check stage | Prompts | Runs per prompt | Responses | What you can judge at this scale | What you can't judge |
|---|---|---|---|---|---|
| First diagnosis | 10 | 3 | 30 | Whether you almost never appear on key prompts | Gaps versus competitors, month-to-month change |
| Monthly manual check | 15-20 | 3-5 | 45-100 | Strengths and weaknesses by prompt type, competitors that often appear with you | Changes on the order of 5pp |
| Tool-based measurement | Dozens to hundreds | Automated repeats | Hundreds to thousands | Share versus competitors, change over time | Blended totals can mislead unless each engine is viewed separately |
These prompt and run counts are not an external standard either. They are this guide's suggested starting point, based on the probability table above.
What setup to ask in: personalization and search settings
Even with the right prompts and run counts, you can't compare results if the setup changes each time. A manual check needs to control four variables.
Personalization. According to OpenAI's help center, before sending the first message in a temporary chat you choose Personalized or Unpersonalized, and Unpersonalized does not use memory, custom instructions or plugins.[3] Neither mode creates or updates memories while the chat is temporary, and you can't change the personalization setting once the conversation has started.[3] Run brand checks in an unpersonalized temporary chat. If your account has a memory like "Our company is XX," answers will lean toward it.
Whether search is used. ChatGPT can search the web automatically when current information would help, and users can also turn search on themselves by choosing Search from the tools menu.[4] Responses that use search include citations, and clicking Sources shows the cited sources and related links.[4] The same prompt rests on different evidence depending on whether search was used, so log it in its own column.
Location. OpenAI explains that ChatGPT may infer an approximate location from your IP address and share it with search providers, and that with memory on it can use saved memories to rewrite search queries.[4] For a brand selling in Korea, ask from Korea in Korean, and flag any run made over a VPN.
Login status and model. OpenAI's help center states that logged-out users can also use web search.[4] Login status, plan and the model used can vary from run to run, so record them every time.
Four common mistakes
Asking repeatedly in the same chat window. The previous question and answer become context for the next response. Open a new temporary chat for every run.
Sharing only the screenshots where your brand appears. Reporting the 1 run out of 5 that mentioned you as a single screenshot makes a 20% mention rate look like 100%. Log every run in the sheet, including the ones with zero mentions.
Using rank as the metric. When the same list in the same order comes up fewer than once in 1,000 runs, a single ranking is close to chance.[1] Count whether the brand appears, and note order only in a reference column.
Judging visibility from branded prompts. ChatGPT describing your brand well when asked "Do you know XX?" doesn't mean it recommends you on category prompts. If the branded answer gets your description wrong, that's an information error, not a visibility problem; Fixing wrong brand information in AI answers covers how to respond.
The seven-step manual check and logging sheet
| Step | What to do | Standard | Output |
|---|---|---|---|
| 1. Write the prompt set | Write category and situation prompts customers would actually use | 10-20 prompts, each intent phrased 2-3 ways | List of prompt IDs |
| 2. Separate branded prompts | Put company, product and price range questions in their own set | About 5 | Accuracy checklist |
| 3. Fix the setup | Unpersonalized temporary chat, search on or off, location, model | Identical every round | Setup log column |
| 4. Run repeatedly | Repeat each prompt in a new conversation | 3-5 runs per prompt, same day | Saved raw responses |
| 5. Log | Mention or not, order, competitors that appeared, cited URLs | Every run, including zero-mention runs | Logging sheet |
| 6. Tally | Calculate mention rate by prompt type and overall | Always report the response count | "12 of 60, 20%" format |
| 7. Compare monthly | Rerun next month with the same set and setup | New baseline if 10% or more of prompts change | Monthly trend table |
The logging sheet only needs the columns below. The values are a made-up example to show the format.
| Date | Prompt ID | Run | Setup | Brand mentioned | Mention order | Competitors that appeared | Cited URLs | Description error |
|---|---|---|---|---|---|---|---|---|
| e.g., 11-09 | C03 | 1 | Unpersonalized, search on, Seoul | Yes | 3 | Company A, Company B | 1 company blog post | None |
| e.g., 11-09 | C03 | 2 | Unpersonalized, search on, Seoul | No | None | Company A, Company C | None | N/A |
The cited URL column matters most. On runs where your brand didn't appear, look at which pages the competitors were recommended from, and you'll see what you need to create. Why competitors appear in AI search but your brand doesn't diagnoses that problem, and How to get your brand mentioned in ChatGPT covers how to improve ChatGPT visibility.
Where manual checks run out, and when to switch to a tool
A manual check tops out at one engine, ChatGPT, and around 100 responses. Telling a 30% mention rate apart from 35% at a 5% significance level and 80% power takes about 1,377 independent responses per period (two-proportion comparison formula; the full working is in AI visibility score sample size guidelines). That is beyond what you can process by hand. Add Gemini, Perplexity and Korean engines and the work multiplies by the number of engines.
In Korea, some services will run a manual diagnosis for you. One GPT search visibility diagnosis listed on Kmong bases its basic package on the brand name and up to 3 core keywords, with competitor comparison available on inquiry.[7] That works as a first look, but by this guide's standard, a check of about 3 keywords only tells you whether you almost never show up. Intermajor points readers to AI Score, its free self-diagnosis tool.[5]
Monitoring tools that automate repeat runs and tallying include Profound, Peec AI and Otterly.ai outside Korea, and BOIDA, which emphasizes support for Korean and domestic engines. When choosing one, check whether it runs each prompt repeatedly, whether it reports results as rates alongside response counts, and whether it discloses its measurement setup. Free tools to try first are compared in Free AI visibility audit tools in Korea, and metric design is covered in GEO KPIs and prompt monitoring. For the basics of GEO, start with What is GEO?
Key takeaways
- Checking by asking ChatGPT directly is fine. Just don't draw conclusions from asking one prompt once.
- Judge visibility with non-branded category and situation prompts, and use branded prompts only to check description accuracy.
- For one prompt, the margin of error was 27pp at 5 runs and about 6pp at 100 runs (Evertune, 2026). A manual check of 10 to 20 prompts at 3 to 5 runs each aims only to separate "almost never shows up" from "shows up often."
- Ask in an unpersonalized temporary chat with a new conversation every time, and log whether search was used, location and model (OpenAI Help Center, accessed 2026).
- Record "X of Y responses," not rank, and let the URLs cited for competitors tell you what to do next.
Related companies
- 보이다 (BOIDA)생성형 검색 최적화(GEO) 솔루션, AI 가시성 측정
- 인터메이저 (InterMajor, AI Score)GEO/AEO 컨설팅, 디지털 에이전시
- Otterly.aiAI 가시성 모니터링 툴
- Peec AIAI 가시성 모니터링 플랫폼
- ProfoundAI 가시성 모니터링 플랫폼
Frequently asked questions
- One answer can't tell you that. Even for the same prompt, the chance that ChatGPT returned the same brand list twice was under 1% (SparkToro, 2026). A brand with a true mention rate of 30% will still be missing from a single answer 70% of the time (binomial calculation, assuming independent responses). Ask each prompt at least 3 to 5 times, across 10 or more different prompts, before drawing a conclusion.
- It depends on the goal. For a first diagnosis, 3 to 5 runs per prompt is enough to separate 'almost never shows up' from 'shows up often.' Putting a number on one prompt's mention rate takes far more. In Evertune's analysis the margin of error was 27pp at 5 runs, 12pp at 12 runs and about 6pp at 100 runs (Evertune, 2026), and SparkToro recommended averaging over at least 60 to 100 runs in most cases (SparkToro, 2026).
- Little for checking visibility, but it is necessary for checking accuracy. Once your brand name is in the prompt, ChatGPT has to include it in the answer, so you learn nothing about whether it gets recommended when a real user asks a category question. Evertune also noted that leading prompts containing the brand name can skew results (Evertune, 2026). Use these prompts to check that the founding year, products and price range are described correctly.
- Memory, chat history and location all feed into the answer. OpenAI says that with memory on, ChatGPT can use saved memories when it rewrites search queries, and it may share an approximate IP-based location with search providers (OpenAI Help Center, accessed 2026). Use an unpersonalized temporary chat and log your location so runs can be compared.
- When your monthly response count passes a few hundred, or when you need to report share against competitors and change over time. Telling a 30% mention rate apart from 35% at a 5% significance level and 80% power takes about 1,377 independent responses per period (two-proportion comparison formula), which is not something you can do by hand. Until then, a monthly manual check with the logging sheet in this article is enough to see which direction you are heading.
Q.I asked ChatGPT about my category once and my brand didn't come up. Does that mean we have no visibility?
Q.How many times should I ask the same prompt?
Q.Is there any point in asking ChatGPT 'Do you know my brand?'
Q.Why do I get different results on my account than a coworker does?
Q.When should I move from manual checks to a paid AI visibility tool?
Sources
- [1] ↑AIs are highly inconsistent when recommending brands or products — SparkToro
- [2] ↑How many repetitions of a single prompt does it take to understand your AI visibility? — Evertune
- [3] ↑Temporary chat in ChatGPT — OpenAI Help Center
- [4] ↑Searching the web with ChatGPT — OpenAI Help Center
- [5] ↑구독자님의 브랜드, 구글에서는 상위인데 왜 ChatGPT에는 안 나오는지 궁금하셨나요? — INTER.R(인터메이저 뉴스레터)
- [6] ↑SEO vs AEO vs GEO: 차이점, 비교표, 데이터 총정리(2026) — SEOcrawl
- [7] ↑내 브랜드가 GPT 검색에 노출 되는지 진단해드립니다 — 크몽
Related documents
- How Many Prompts and Runs Make an AI Visibility Score Trustworthy? Sample Size and Repeat-Run GuidelinesHow many prompts, run how many times, before you can trust an AI visibility score. Covers confidence intervals by response count, the responses needed to compare two periods, the design effect of repeated runs, public experiments from SparkToro, Evertune and others, and a five-step sampling procedure.
- Why AI Recommends Different Brands Every Time You Ask the Same Question: 5 Causes of Response Variability and How to Measure Around ThemAsk ChatGPT, Claude or Google AI the same question twice and the recommended brands change. This guide splits the cause into five layers (sampling, inference nondeterminism, query fan-out search, personalization and prompt wording), backs each with public experiments from SparkToro and others, and lays out a five-step measurement response.
- Free GEO Audit Tools Compared 2026: Where to Get a Free AI Search Visibility Check, 9 Korean and Global OptionsFree GEO audits come in three shapes, self-serve scans, free trials of paid products, and assisted reports from agencies. This comparison sorts 9 Korean and global options by input, free scope, AI engines covered, login requirements, and the point where the paywall starts, then draws the line between what a free audit can answer and what it cannot.
- How to Get Your Brand Surfaced in ChatGPTChatGPT builds answers from pretraining data and live web search. This piece lays out how to surface your brand in those answers by allowing GPTBot, structuring content so it can be cited, and consolidating your entity.
- 5 Reasons Competitors Show Up in AI Search: but Your Brand Doesn'tA diagnostic breakdown of why ChatGPT, Perplexity, and Google AI Overviews cite competitors while your brand goes unmentioned: covering crawler access, machine readability, entity recognition, third-party mentions, and content structure, with a step-by-step fix for each.
- GEO KPI Performance Measurement: Citation Rate, AI SoV, and Prompt MonitoringA step-by-step guide to the core GEO KPIs: citation rate, AI Share of Voice, and prompt monitoring. Covers the arXiv-grounded statistical framework, brand-scale benchmarks, and a side-by-side comparison of domestic and global measurement tools for quantifying AI search visibility.
- How to Fix Wrong Brand Information in AI Answers: A Five-Step Evidence-Layer GuideChatGPT, Perplexity, and Google AI Overviews get company names, founding years, products, and prices wrong. The answers have no edit button, so the repair runs through the evidence layer: your own schema, Wikidata, and third-party pages. Five steps, with a verification method for each.
- What Is GEO: The Definition of Generative Engine Optimization and How It Differs From SEOGEO (Generative Engine Optimization) is the strategy of getting your content cited in answers produced by generative engines like ChatGPT and Perplexity. Here is the definition, how it differs from SEO, and how it works.