Before You Buy AI Visibility Tools, Know Where ChatGPT Gets Its Opinion of You

AI visibility tools measure how often AI assistants such as ChatGPT name or cite your brand, and the number reflects work done somewhere else. An AI model learns about a brand from its training data and from pages it pulls in while answering, mostly other people’s content, so fund that content and keep the tool cheap.

The pitch for AI visibility tools goes like this. Some buyers now ask ChatGPT which vendors to consider, so you need software that shows what the AI says about your brand and helps you change it. Dozens of these tools have launched in the past 2 years, from startups, SEO software companies, and ad agencies.

The first half of that pitch is fair. The second half assumes the AI’s opinion of you is something a dashboard can watch and a subscription can move. Follow that opinion back to where it comes from, and you find mostly content other people wrote, which a dashboard can only report on.

Where does an AI model get what it knows about your brand?

It has 2 sources. The first is everything the model read before it was released, an enormous pile of web pages, books, and other text (its training data). What it absorbed there stays fixed until the company ships a new version.

The second is the web search many AI assistants run while they answer, pulling in fresh pages to support the reply (grounding). Microsoft describes grounding as the step AI systems use to retrieve relevant content before they generate a response (1. Microsoft Clarity, 2026).

Both sources draw on the open web, which is why ordinary search work still matters. Cyrus Shepard, who runs the SEO consultancy Zyppy, reviewed the published experiments and studies on what gets a page cited by AI. Near the top of his list sat whether the AI can reach the page at all, during training or while answering, and how well the page ranks in ordinary search for the same question. His summary was “win SEO, win AI citations” (2. Shepard, 2026).

Can you change what ChatGPT says about your brand by talking to it?

Talking to the chatbot is the wrong lever. OpenAI’s help page says conversations from personal accounts may be used to train its models, that people can opt out, and that business accounts are excluded by default (3. OpenAI, n.d.). The only path it describes runs through training a future model.

So typing “Coca-Cola is terrible” into ChatGPT does about what shouting at a billboard does.

How do AI models decide which brands to recommend?

Once facts are in front of a model, they beat a famous name. Researchers at Trine University and Texas A&M gave 3 AI models a list of 10 skincare products with identical specs. Of those, 9 had invented brand names and 1 was real. The real brand won every trial. Then the researchers gave one invented brand a slightly better rating, and a rating edge of less than a tenth of a star was enough for it to win half the time (4. Chu & Hou, 2026). The product details sat in the question itself, so the test shows how models weigh information they can see.

The models defaulted to the name they knew and dropped that default as soon as they had a reason to. For a smaller brand, that’s the useful half of the finding. The open question is how your facts reach the model in the first place, and that’s where press coverage, reviews, and search rankings come in.

Whose content does the AI read about you?

Mostly other people’s. Omniscient Digital, an organic growth agency, ran prompts that named specific brands through ChatGPT, Perplexity, Gemini, and Google’s AI features, then sorted every source the answers cited. Coverage the brand didn’t pay for or publish, such as press, reviews, forums, and directories (earned media), made up 48% of citations. Other companies’ content, such as a competitor’s comparison post, made up 30%. The brand’s own site came in at 23% (5. Omniscient Digital, 2026).

When someone asks an AI about your company, about three quarters of the sources it cites were written by somebody else. A visibility dashboard reports on that pile. Changing what’s in it takes press coverage, reviews, and other people’s posts.

Why is an AI visibility score closer to a poll than a ranking?

The answers change from one ask to the next. Rand Fishkin of the audience research company SparkToro set out to prove AI tracking was pointless. Working with Gumshoe, an AI tracking startup, he had volunteers run the same prompts again and again through ChatGPT, Claude, and Google’s AI. In that test, if you asked ChatGPT or Google’s AI the same question 100 times, the chance that any 2 answers listed the same brands was under 1 in 100 (6. Fishkin, 2026).

He did change his mind on one point. How often a brand shows up across many runs turned out to mean something, and he concluded that share is probably a reasonable thing to measure. Rankings stayed close to useless.

The tools don’t agree with each other either. “If you use three different tools and give them the same prompts, you get three different answers,” Paul Dyer, CEO of the AI services agency /prompt, told Digiday (7. Digiday, 2026).

So read the score the way you’d read a poll. Ask the vendor which questions it asks the AI, how many times it asks each one, and in which AI assistants. A score without its sample is a number you can’t check. You can run a small version of that poll yourself by asking your buyers’ questions by hand each week .

What should you pay for instead?

Pay for what the AI reads. That means coverage in publications your buyers trust, reviews on the sites they check, and pages that rank in ordinary Google search for the questions they ask. That’s familiar work for most marketing teams, and it pays off with human readers too.

Then keep the measurement cheap. Microsoft made the AI citations report in Clarity, its free website analytics tool, generally available in May 2026. It shows which of your pages get cited in AI answers across the AI services Microsoft supports, and which queries the AI used to find them (1. Microsoft Clarity, 2026). Start with it before you sign a contract.

A paid tool earns its cost when you’re tracking a category against named competitors across several AI assistants and you’ll act on what it shows. The trade-off is that its findings will mostly send you back to press, reviews, and search.

In the demo, ask what your team would do differently because of the number. If the answer is more press coverage, more reviews, and better search rankings, you can start that work without the tool.