How to check it properly, in six steps
Do this before you touch your website, your content or your budget. It costs nothing but a few minutes a day.
- Fix the question. Write the exact question a buyer would type, the one where your competitor showed up. Leave your own name out of it. Use the identical wording every run, because rewording it means you are testing a different question.
- Take your own history out of the picture. Start a new Temporary Chat and choose Unpersonalized before you begin. OpenAI's Memory FAQ says that choice stops existing memories, custom instructions and plugins being used, and that temporary chats do not create or update memories.
- Run it at least ten times, over at least three days. That number is our rule of thumb, not a published threshold. It gives you a rough read, not proof. SparkToro ran each of its prompts 60 to 100 times.
- Write down every brand named, every run. Not just the first one. Record the date as well.
- Open Sources when your rival appears. OpenAI's ChatGPT search help page says you can select Sources, when available, to view cited sources and other relevant links, and warns that results can be incomplete or incorrect. Note the domains. They are a clue to where the rival is being mentioned, not a full list.
- Work out two numbers. Divide the runs where your rival appeared by your total runs. Do the same for yourself.
Here is how a log might look. These rows are invented to show the layout, not real results.
| Run | Day | Brands named | You | Rival |
|---|---|---|---|---|
| 1 | Mon | Brand B, Rival, Brand C | No | Yes |
| 2 | Mon | Brand D, Brand B, Brand E | No | No |
| 3 | Wed | Rival, Brand C, You | Yes | Yes |
| 4 | Thu | Brand C, Brand B, Brand F | No | No |
| 5 | Fri | Brand B, Brand E, Rival | No | Yes |
Then read the result against this table. The right-hand column is our reading of the evidence, not a rule from the study.
| What your log shows | What it probably means | What to do |
|---|---|---|
| Rival in most runs, you in almost none, same pattern on different days | A possible gap in how consistently ChatGPT links you to this question | Work through the section on what makes a gap real below |
| You and the rival both appear, but in different runs | You are both names in a rotating set | Keep logging, change nothing yet |
| You appear often but rarely first | Presence is fine, and position on a single run is close to noise | Do not chase rank |
| A different rival turns up each time | Many credible options, and you saw one draw | Nothing to fix yet |
| You appear more often than the rival | The sighting was the exception | Stop worrying about it |
Why one look proves so little
The study behind the headline number is SparkToro's research on AI brand recommendations, by Rand Fishkin with Patrick O'Donnell, published on 28 January 2026 and updated on 28 June 2026. 600 volunteers ran 12 prompts through ChatGPT, Claude and Google's AI a combined 2,961 times. The prompts spanned consumer products, services and local categories. The headline result: a less than 1 in 100 chance that ChatGPT or Google's AI, asked 100 times, gives the same list of brands in any two responses. Order was worse. It was more like 1 in 1,000 runs before two lists came back in the same order.
The study also separates two things that feel like one. A hospital appeared in 69 of 71 ChatGPT answers to a cancer-care prompt, a 97% visibility rate, yet was the top mention in only 25 of them. High presence with low rank is a different situation from never appearing, and it needs a different response.
Category size matters too. SparkToro found that categories with few real options, such as cloud providers for SaaS startups, produced more consistent answers than categories with hundreds of options, such as new science fiction novels. A specialist market behaves more like the first case, so a competitor that keeps appearing there deserves more attention than it would in a crowded one. No study has tested a market like yours, so treat that as a reason to look closely, not a prediction.
A separate Washington State University study, reported in March 2026, found ChatGPT gave consistent results in only about 73% of cases when the same prompt was submitted 10 times. It tested true or false judgements on 719 research hypotheses, not brand recommendations, so it is not a number about your market. It does show the same question producing different answers is not unique to brand prompts.
Do not read any of this as "AI tracking is pointless". SparkToro's author started out expecting exactly that and changed his mind. He concluded that rank on a single run is close to random, but that measuring how often a brand appears across dozens or hundreds of runs is probably a reasonable way to see how prominent it is in the AI's consideration set. Those are two different claims. Only the first one is unreliable.
What makes a gap real, and what to do next
If the log comes back the other way, with your rival in most runs and you in few or none, over several days, in our reading that is a signal and not noise, though ten runs is a screen, not proof. SparkToro's data has an example close to our own world: when Google's AI was asked to recommend digital marketing consultants with e-commerce expertise, one agency appeared in 85 of 95 responses. That was a single prompt on a single engine, so it shows what a lopsided result looks like, not how common one is.
The Sources list from step 5 is your first piece of evidence. Domains that keep appearing next to your rival are clues to pages the engine may have drawn on. Our inference, which no engine has confirmed, is that being named on pages like those matters more than anything you change on your own site. The fixes break into two jobs, and we have written both up in full. Whether your site is readable enough for an engine to cite is covered in Make Your Website Readable to AI. Whether other trusted sources corroborate your business is covered in How to Get Cited by ChatGPT. If you want to keep measuring once you have made changes, How to Check Whether AI Is Recommending Your Brand covers what tracking tools actually measure.
If the log shows a rotating set instead, there is nothing to fix yet. Keep running the test once a week and do not spend money reacting to a single screenshot.
Mentioned is not the same as recommended
Your log counts mentions. A mention means your name appeared in the answer. A recommendation means ChatGPT put you forward as the answer to the buyer's question. Add one column to your log, "Recommended?", and mark Yes only when the answer singles you out as a good fit, not when you sit in a list of names. This is our own method, not something SparkToro measured. A brand that is often mentioned but rarely recommended has a different job from one that is never mentioned.
Why a competitor might keep appearing
Nobody outside the AI labs can say why one brand is named over another. These are our own working hypotheses, not statistics, and none is confirmed by OpenAI.
- Third-party mentions. Rivals that are named on comparison articles, directories and forum threads give an engine more places to find them than a brand that mostly talks about itself. Your Sources list from step 5 is where to look.
- Reviews. A competitor with many recent reviews on the sites buyers use leaves a larger trail than one with few.
- Comparison pages. A page that sets two or more options side by side may be an easy source for an answer that has to choose. If your rival has one and you do not, note it.
Treat each as a thing to check against your own log and Sources, not a fix to buy.
What OpenAI has and has not said about personalisation
This is the part most people guess at, so here is the documented edge of it. OpenAI says ChatGPT may use your IP address to estimate your general location, which can improve results for nearby restaurants, local news and weather. Its Memory FAQ says that when Memory is on, ChatGPT may use relevant details to formulate a more useful search query. Its Memory FAQ gives location and dietary preferences refining a restaurant search as the example, and its search help page shows a "near me" restaurant question being rewritten as "top restaurants San Francisco" once ChatGPT estimates your area from your IP address.
Neither page says memory or location changes which non-local business gets named in a recommendation answer. That link is a reasonable guess, not a documented fact. But it is why step 2 matters. If your own chat history includes your industry, your city or your competitors, your account may be running a different search from the one your prospect runs.
Guarantees no one can back
No engine publishes its selection algorithm, and nobody outside the AI labs can hand you a formula that works across categories, engines and time. Be wary of a guaranteed ChatGPT citation, a promised ranking position or an "AI SEO score" sold as settled science. None of it is backed by anything OpenAI or Google has confirmed.
Chasing rank on one run is the quickest way to waste budget, since the research above found ordering close to random on any single attempt. If you want the wider picture of where search and AI-answer budget pays off, read SEO vs AEO: Where to Invest First and How AI Overviews Choose Which Brands to Cite.
Run the test before deciding anything. If a pattern holds across a week, tell us about your business and we will help you work out what is happening in your category before you spend anything fixing it.
↳ Frequently asked
01How many times should I ask ChatGPT before I believe the answer?
There is no published threshold. Our rule of thumb is at least ten runs over three or more days, which gives a rough read of how often each brand appears. SparkToro ran each of its prompts 60 to 100 times and described appearance rate across dozens or hundreds of runs as probably a reasonable measure, so treat ten as a screen, not proof.
02Does asking ChatGPT the same question twice count as a real test?
It rules out the most obvious false alarm, but two answers cannot show a pattern. SparkToro found under a 1 in 100 chance of the same brand list coming back twice, so two different lists are the normal result and two matching lists would be unusual.
03Why would ChatGPT mention a competitor by name and never mention us at all, even once?
That could be a genuine gap in corroboration or site structure, or it could be a category with many credible answers and a naturally low chance of naming any one brand on a given run. The repeat test over several days is what tells those two apart.
04If we never show up, does that mean ChatGPT does not know we exist?
Not necessarily. Absence in a handful of manual runs is not proof of absence overall, especially in a category with many possible answers. It is a reason to test properly before concluding anything.
05Can my own ChatGPT history make a competitor appear more often?
It can shape the search ChatGPT runs. OpenAI says that with Memory on, ChatGPT may use relevant details to formulate a more useful search query. It has not said this changes which non-local brand gets named, so test in an Unpersonalized Temporary Chat to remove the variable.
06Should we pay for a service that promises we will be recommended by ChatGPT?
Be sceptical of any fixed guarantee, since neither OpenAI nor Google has published the mechanics a vendor would need to guarantee that outcome. In our view, a service that improves how readable and corroborated your site is can genuinely help. A service that promises a named outcome is promising something nobody can currently deliver on demand.
07Does this apply the same way to Google's AI Overviews as it does to ChatGPT?
The instability finding covers both, because SparkToro ran its prompts across ChatGPT, Claude and Google's AI Overview and AI Mode. The mechanics behind each differ, which is why we have written about Google's citation behaviour separately.