How do you check whether ChatGPT recommends your brand, and why is one answer not a measurement?

To check whether ChatGPT recommends your brand, ask it the unbranded questions your buyers type, repeat each question many times in fresh chats, and count how often your brand is recommended, with a 95% confidence interval. One ChatGPT answer is not a measurement, because the same question can return different brands on different runs.

• OpenAI says ChatGPT may search the web automatically when a question would benefit from current information, and that search results and citations can be incomplete, outdated, or incorrect (OpenAI, Help Center, read 30 October 2026).
• OpenAI's API documentation says model output is non-deterministic, and that even different snapshots within one model family could produce different results (OpenAI, API documentation, read 30 October 2026).
• Penlid's methodology prints a 95% Wilson interval of 10.8% to 60.3% for 3 of 10 samples and 27.2% to 32.9% for 300 of 1,000, although both read as 30% (Penlid, Methodology, read 30 October 2026).
• Penlid's free AI visibility check asks one engine, Claude with live web search, and not ChatGPT (Penlid, AI visibility check tool page, read 30 October 2026).

Why is one ChatGPT answer not a measurement?

One ChatGPT answer is a single draw from a process that varies, so it cannot tell you how often your brand is recommended. OpenAI's API documentation says the content generated from a model is non-deterministic, and Penlid's methodology draws the consequence: A single observation is therefore not a measurement. A preprint from April 2026 says answers can vary across runs, prompts, and time, making one-off observations unreliable.

ChatGPT adds steps of its own. OpenAI's help page says it may search the web automatically when a question would benefit from current information, that when it partners with other search providers it typically rewrites your query into one or more targeted queries that it sends to them, that with memory enabled it may use relevant saved memories when rewriting a search query, and that it may use an approximate location based on your IP address. That gives at least four places where two runs can differ: whether ChatGPT searches, how it rewrites the question, what the search returns and how the answer is worded (our reasoning, not OpenAI's list).

How do you check whether ChatGPT recommends your brand?

You check it by asking ChatGPT the questions your buyers ask, many times, and counting how often it recommends your brand, in six steps.

1. Write unbranded questions. A question that names your brand makes a mention near-certain and measures nothing, so Penlid's methodology asks the comparison, alternatives and buyer's-guide questions a real customer types (our guide to unbranded buying prompts explains why). Save every prompt word for word.
2. Fix one setup. Keep the signed-in state, model choice and country the same for every run, because OpenAI says ChatGPT may use saved memories and an approximate location. A signed-out session has no saved memories (our reasoning).
3. Ask each question many times, each in a new chat. A new chat keeps earlier turns from shaping the answer, and spreading the runs over several days covers variation over time (our reasoning). For a result near 30% to carry an interval about 20 points wide, plan on about 80 runs in total (see the table below).
4. Record every run. Save the date and time, your setup, the model if the interface names it, the full answer, whether your brand was recommended and whether your site appeared among the sources.
5. Count and report an interval. Divide the runs that recommended you (k) by all runs (n) and report k of n with a 95% Wilson score interval, not a bare percentage. Penlid's methodology prefers Wilson to the textbook normal approximation because the normal interval behaves badly at small n and near 0% or 100%.
6. Repeat on a schedule, and treat a model change as a break. Penlid's methodology says A rate that spans a model change is two measurements, not one. OpenAI's release notes list updates to GPT-5.5 Instant on 28 May and again on 24 June 2026, the same model name both times, and our guide to why AI visibility changes when you changed nothing covers the rest.

Is being cited the same as being recommended?

No: a citation and a recommendation are different outcomes, so count them in separate columns. Penlid's AI visibility check page says Being listed as a source and being recommended as an option are different outcomes. OpenAI says responses that use web search may include citations, and its 2024 launch post for ChatGPT search described a Sources button below the response that opens a sidebar with the references. If ChatGPT does name you, check what it says about you, because OpenAI warns Search results and citations can be incomplete, outdated, or incorrect.

How much do more runs narrow the uncertainty?

Slowly: around a result of 30%, a 95% Wilson interval is 60.3 − 10.8 = 49.5 points wide at 10 samples, 40.8 − 21.1 = 19.7 points wide at 80 and 32.9 − 27.2 = 5.7 points wide at 1,000. The table's first row is our own computation for the four samples that Penlid's free check takes: for 0 of n the upper end is z × z ÷ (n + z × z), which for z = 1.96 and n = 4 is 3.8416 ÷ 7.8416 = 0.490, so 0 of 4 allows a true rate of up to 49.0%.

Step 3 suggests about 80 runs because 24 of 80 is the smallest sample in which a whole count equals exactly 30% and the half-width, (40.8 − 21.1) ÷ 2 = 9.85 points, is at most 10 (our computation). The interval measures only sampling noise for the prompts you asked; it does not correct for prompt choice, personalization or model changes, which Penlid's methodology lists as separate limits.

What should you check if ChatGPT never names your brand?

Check first that you have not opted out of OpenAI's search crawler, because OpenAI says Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though can still appear as navigational links. OpenAI recommends allowing OAI-SearchBot in your robots.txt file and allowing requests from its published IP ranges. Each crawler setting is independent, so you can allow OAI-SearchBot while disallowing GPTBot, and OpenAI says a robots.txt change can take about 24 hours to take effect for search, so wait a day before you sample again. Our guide to which crawlers to allow for AI search compares the vendors. Allowing the crawler is not a promise of a mention: OpenAI's help page says ChatGPT ranks search results using multiple factors and that placement is not guaranteed, and the OpenAI pages we read do not say how ChatGPT chooses which brands to name.

Can a tool do the checking for you?

A tool can repeat the runs for you, but only if its report names ChatGPT among the engines it actually asked and shows how many runs sit behind each figure. Penlid's methodology puts the test this way: A tool that lists engines it did not ask is reporting a number nobody measured. Our free AI visibility check does not ask ChatGPT. Its page says it asks one engine, Claude with live web search, four of the questions your buyers type, so its numbers describe Claude. It shows a 95% Wilson interval on every figure because four samples is four samples, and it reports being recommended and being cited separately.