An AI does not give you rankings, it gives you answers, and a different one every time. This explains how to measure that without fooling yourself, what we store from each answer, and what the published research says can actually be changed.
Measuring what an AI says is nothing like measuring Google rankings. On Google there is a list, your page holds a position, and that position is roughly the same whether you check today or tomorrow. In an AI there is no list: there is a written answer, and that answer changes every time.
Everything below follows from that. If the answer changes, a single data point means nothing and you have to measure differently.
Ask an AI assistant the same question twice and you get two different answers. Not slightly different. Researchers repeated the same questions many times over and compared which pages the engine used on each run: only three to five out of every ten repeated from one go to the next, depending on the engine.
That gives us the two rules we work by:
The same work calculated that the margin of error narrows by about 29% every time you double the number of measurements, and that the number of answers needed for a ranking to settle ranged from 40 to never, depending on the topic. So frequency is not a product extra: it is what turns noise into a trend you can read.
One signal is jumpier still: tone. The same large-scale study found that whether the AI speaks well or badly of you changes almost seven times more often than whether it names you at all. That is why one negative answer is not a crisis, and a month of them is.
We ask the engine your questions and store the whole answer. Everything below comes out of those stored answers. Nothing is estimated or extrapolated.
| What we look at | What you see | Why it is useful |
|---|---|---|
| Whether it names you | Yes or no, for each question, each day | How many days out of thirty you came up. That is your real number |
| Where it names you | First, third or last in the list | Being named last out of six is not the same as leading the list |
| How it talks about you | Good, neutral or bad | The jumpiest signal, so read it over weeks and never from one answer |
| Who it names instead | The competitors sharing the answer with you | You see exactly who takes the customer when you are not there |
| Which pages it read | The links behind each answer | Usually other people's pages: reviews, magazines, comparisons |
| What it searched for | The searches it ran before answering | The AI telling you what it went looking for and could not find on your site |
That is the question to settle before spending a penny on this, and somebody measured it. Similarweb followed what people did right after an AI recommended a company, and compared it with the competitor it did not recommend.
Out of 100 people who got a recommendation, how many visited the website
Blue: the company the AI recommended. Grey: the competitor people went to instead
Source: Similarweb, 2026 Generative AI Landscape. United States, computers, July to December 2025. Bars on a scale up to 15 people.
Now the honest part, which almost nobody includes. This shows the two things happening together; it does not prove the AI caused the visit, and plenty of other things push somebody to buy.
There is also a study that puts a number on the usual exaggeration. One website saw its visits from ChatGPT grow almost six times over, but compared against other pages on the same site, most of that growth was simply ChatGPT growing. What could honestly be credited to the work done was a little under double. When you read a case study with a huge number in it, take a good deal off.
Worth knowing before you start, because it sets expectations. Researchers went through more than 100,000 AI answers about more than 100 companies and found a fairly brutal ladder.
Out of every 100 relevant answers, how many name the brand
More than 100,000 answers about more than 100 companies, March to May 2026
Source: P. Kumar, Generative Engine Optimization at Scale, arXiv:2606.20065.
The good news comes from a different study. They showed an AI two products identical in every way except the name, one a real brand and one invented: the real brand won all 670 times. Then they gave the unknown one a tiny edge, less than a tenth of a star more in rating, and the result flipped. The famous name only wins while the AI has nothing else to go on.
Putting the available research together, the list is shorter and less exciting than most of what you will read:
And a warning we would rather give you ourselves: a review of 45 studies on this subject concluded that what has been rigorously proven is fairly modest. You can improve how the AI uses you once it has already found your page; reliably making it find you more often has not been proven yet. Anyone promising guaranteed results is ahead of the evidence.
Nothing to install and nothing to change on your site. Thirty seconds.
Real ones, the way a person would say them out loud. Or let our AI draft them from your sector and change whatever you want.
Every answer is stored whole, with its mentions, position, tone and the pages it read. Most engines daily; the most expensive to query, every 72 hours.
How many days you came up, how you compare with competitors, which pages feed the answers you lose and what the AI searched for before replying.
Each has its own audience and its own way of answering, which is why each has its own page with its own data:
| Assistant | Why it matters |
|---|---|
| ChatGPT | The most used assistant in the world, the one people open to ask for recommendations |
| Google AI Mode | Google's own conversational answer, in front of the biggest audience there is |
| Google AI Overviews | The summary above the normal results, read by people who never scroll down |
| Perplexity | Shows its sources for everything, popular with people researching before they buy |
| Gemini | Comes built into Google's apps and Android phones |
| Copilot | Comes built into Windows and Office, so it is everywhere at work |
| Claude | Very popular with professionals and university users |
| DeepSeek | Growing fast, especially outside Europe and the US |
| Qwen | The main assistant in Chinese-speaking markets |
Because the engine does not answer the same thing twice. Researchers repeated the same question many times over and only three to five out of every ten pages it used repeated from one go to the next. A monthly check gives you a number so noisy you cannot tell a real improvement from a coincidence.
Ten well chosen questions are enough to see the trend. More questions add precision, but the big jump comes from checking every day instead of once, not from going from 10 questions to 40.
Yes, and it is more reliable than looking at yours alone. Because it is a comparison, movements that affect the whole sector at once cancel themselves out.
No. There is no code or tag to add. We only need your brand name, your domain and your questions.
You will see it in the history, with the full answer stored and the pages it used to build it. That is what lets you go to the specific source causing it, which is almost never your own site.
Everything on this page comes from published research, not from us:
Free plan, no card: we ask your questions to ChatGPT every day and keep the history. Full plan detail on the pricing page.
Start tracking free