How to Check If ChatGPT Recommends Your Business (Repeatable Method)

To check if ChatGPT recommends your business, run a fixed set of 15 to 25 unbranded buyer questions, repeat each one at least five times in fresh chats, and log whether your brand and each competitor is named. Report the share of answers that name you, not what one chat said. Single spot checks mislead because the answers change on every run.

Key takeaways

• One manual check proves nothing. Search Engine Journal, January 2026 reports SparkToro found less than a 1-in-100 chance of ChatGPT returning the same brand list twice.
• Measure a rate, not an answer: mentions divided by runs, for you and for each competitor, across a prompt set that never changes.
• Use unbranded questions ("best invoicing software for freelancers"). Asking "is Acme good?" only tests whether ChatGPT has heard of you.
• Run each prompt 5 times or more. With 50 answers, a 28% mention rate carries a margin of roughly 12 points, so small moves are noise.
• Log mentions and citations separately. Ahrefs, October 2026 found 54% of ChatGPT citations in its sample were for sites not named in the answer text.
• Automate once the sheet gets tedious. The method stays the same; a tool just runs it daily.

How can I check if ChatGPT recommends my business?

Treat it as a small experiment with a fixed instrument. The steps below take about two hours for the first baseline and are the same whether you do them by hand or with software. For the wider programme around this, see our guide to running your first AEO baseline.

Step 1: Write 15 to 25 prompts a buyer would type

Pull them from sales calls, support tickets and search queries. Cover five kinds: category discovery ("best CRM for a 10-person agency"), comparison ("X vs Y"), use case ("how do I automate invoice chasing"), problem ("why are my emails landing in spam") and brand ("what is Acme"). Keep most of them unbranded, since that is where recommendations happen. Which prompts to track goes deeper on selection, and the free AI Prompt Finder suggests buyer questions for your category.

Freeze the wording. If you change a prompt, you have started a new series.

Step 2: Fix the conditions

• Open a new chat for every run, so earlier answers cannot colour later ones.
• Log out, or use a clean profile with memory and custom instructions off, so your own history does not bias the result.
• Pick one mode and stick to it: with web search on, or off. Note which.
• Record the date, because results drift.

Step 3: Run each prompt at least five times

Five runs of 20 prompts gives 100 answers per engine. That is enough to see a pattern and few enough to do by hand in an afternoon. Paste each full answer into a sheet; do not summarise from memory.

Step 4: Log mentions against competitors

Use one row per answer. This layout is enough:

Column	What to record	Why
Prompt ID and type	Discovery, comparison, use case, problem, brand	Lets you split results by buying stage
Run number and date	1 to 5, plus the date	Separates noise from change over time
You named?	Yes or no	Numerator for your mention rate
Position	1st, 2nd, 3rd or "in a longer list"	Directional only; order is unstable
Competitors named	Each brand, comma separated	Builds share of voice
Your URL cited?	Yes or no, with the URL	Citation is separate from mention
Wording	The phrase used about you	Catches wrong or stale claims

Step 5: Turn rows into three numbers

Mention rate is answers naming you divided by total answers. Competitor share of voice is the same for each rival. Citation rate is answers linking to your domain divided by total answers. Our explainer on AI share of voice covers the formula and what it misses. The free Competitor AI Share of Voice tool compares you with up to three rivals across four engines, which is a quick cross-check on your sheet.

Why do ChatGPT answers about my brand change every time I ask?

Because the answer is generated, not looked up. Two identical prompts can produce different lists, different lengths and different orders.

SparkToro's study, written up by Search Engine Journal, January 2026, ran 2,961 prompts across ChatGPT, Claude and Google's AI features, using 12 brand-recommendation prompts and hundreds of volunteers. It found less than a 1-in-100 chance that ChatGPT or Google's AI would return the same list of brands across repeated runs of a prompt, and the same list in the same order less than 0.1% of the time.

The same research found the problem is not hopeless. MediaPost, 2026 reports that when the engines were asked to recommend digital marketing consultants with ecommerce expertise, one agency, Smartsites, appeared in 85 of 95 responses, about 89%. Visibility across many runs was far steadier than position in any one answer. That is the case for measuring a rate.

Other things also move the answer:

• Web search on or off. With search, ChatGPT pulls live pages; without it, it leans on training data.
• Your own account. Memory, history and custom instructions tilt answers towards what you already use.
• Location and language. A "near me" or regional question changes the candidate set.
• Model updates. A new release can shift a whole category; see what a new model release changes for brand visibility.
• Source churn. Ahrefs, October 2026 tracked about 37,500 citations daily for 20 days and found 50 to 83% were gone the next day, with about 75% of those back within 20 days.

Our own piece on AI citation stability shows how to read that churn.

Why are manual spot checks unreliable?

A spot check fails on sample size, bias and memory. The table sets out the specific failure and the fix.

Spot-check habit	What goes wrong	Repeatable fix
Ask once, screenshot the result	You see one draw from a wide spread	5 or more runs per prompt
Ask "is [brand] good?"	Tests recognition, not recommendation	Unbranded buyer questions
Use your own logged-in account	Memory and history bias the answer	Clean profile, new chat per run
Reword prompts each time	Changes the instrument and the result	Frozen prompt wording
Check only ChatGPT	Engines disagree	Add Gemini, Perplexity and Claude; see why engines disagree
Check when it is convenient	No trend, no baseline	Fixed weekly or daily schedule
Count a link as a win	Cited is not the same as named	Log both; see citation vs brand mention

How many runs do I need before I trust the number?

Enough that the margin of error is smaller than the change you care about. Here is a worked example with invented figures, to show the arithmetic. Assume 10 prompts, 5 runs each: 50 answers.

• Your brand is named in 14 answers: 14 / 50 = 28%.
• Competitor A is named in 31 answers: 62%.
• Competitor B is named in 9 answers: 18%.

The standard error for 28% on 50 answers is the square root of (0.28 x 0.72 / 50), about 6.3 points, so the 95% margin is about 12 points: your true rate is plausibly 16% to 40%. If next month you read 34%, you cannot call that growth. Runs of the same prompt are not fully independent, so treat the margin as a floor, not a ceiling. Doubling to 100 answers narrows it to about 9 points.

The practical rule: call a change real only when it exceeds the margin, or when it holds across several consecutive weekly readings. Ahrefs, October 2026 gives a similar tip for citations: count how many of the last seven days you appeared, rather than judging from one check. The same logic is in our AI Visibility Measurement Standard.

What should I read in the results?

Sort every prompt into one of four outcomes. This decision table says what to do with each.

Outcome across 5 runs	Meaning	Next action
Named in 4 or 5 runs	Established in that answer	Watch for drops; check wording is accurate
Named in 1 to 3 runs	On the edge of the candidate set	Strengthen the page and third-party sources behind it
Named in 0 runs, rivals named	Outside the set	Find which pages the answers cite and what they say
Named, but wrong or dated claim	Accuracy problem	Follow what to do when AI gets your brand wrong

For the zero rows, open the cited sources. They are usually review sites, listicles, forums and rivals' own pages. That list is your content and outreach backlog, and it is why how to rank in ChatGPT starts with source analysis.

Before blaming content, rule out access. OpenAI's crawler documentation describes OAI-SearchBot as the crawler behind ChatGPT search, so a site that blocks it in robots.txt may not appear in those answers. The free AI Crawler Checker shows whether GPTBot, OAI-SearchBot, ClaudeBot and PerplexityBot can reach your site and generates a corrected robots.txt.

Which answer-quality checks catch more than a mention?

A mention is not always a recommendation. Read each answer where you appear and grade it on three questions: were you recommended or just listed, was the claim accurate, and who else was recommended alongside you? Record the phrase used about you. If ChatGPT calls a mid-market product "a budget option" or quotes an old price, that is a finding even when the mention rate looks fine. Our guide to brand sentiment in AI answers covers how to score this.

When should I automate the check?

Automate when the sheet becomes the job. At 20 prompts, 5 runs, 4 engines, one weekly pass is 400 answers to run, paste and tag. That is the point where most teams stop doing it consistently.

Three signs it is time: you need a trend by week, you want more than one engine, or someone outside marketing wants a number. ClickForest, published July 2026, updated October 2026, also advises a fixed list of 20 to 50 questions repeated daily or weekly, which is the same discipline at larger scale. Before spending, try the free Noma by Nurdd brand audit: it checks how AI engines describe your brand.

Where does Noma fit?

Noma by Nurdd is an AEO (answer engine optimisation) and AI visibility platform. Noma asks 13 AI engines, including ChatGPT, Gemini, Perplexity and Claude, the questions your buyers ask, every day, and shows where your brand is named, where competitors are named instead, and which pages the answers draw on. It automates the method above; it does not promise a result. Plans start at $79 a month on the pricing page, checked 3 November 2026.

See it on your own brand: book a demo or start the 3-day free trial. A card is taken at signup so billing can start on day 3; cancel before then and nothing is charged.

Common mistakes

• Changing prompts between months and comparing the results.
• Reporting position as a KPI when order is the least stable part of an answer.
• Declaring victory after one run where you appeared.
• Crediting a page edit for a single reappearance, which Ahrefs warns against.
• Tracking only branded prompts.
• Ignoring wrong descriptions because the mention count is up.

How we researched this

We reviewed OpenAI's crawler documentation, the SparkToro study as reported by Search Engine Journal and MediaPost, Ahrefs' citation stability study and other published how-to guides on checking ChatGPT recommendations. Facts, figures and prices were checked on 2026-11-03. The worked example is illustrative arithmetic, not measured data, and we did not run a test for this article.

What to do next

Write your 20 prompts today, run each five times tomorrow and log the rows. Then decide whether to keep the sheet or automate it. To see the same measurement running daily on your brand, book a demo or start the free trial.