How to track and improve your brand inside ChatGPT

A client asked me last month what ChatGPT says about them. I typed their brand name in, got a glowing paragraph, screenshotted it and sent it over. Everyone was happy.

Then I opened a fresh session, asked the same thing a slightly different way, and ChatGPT recommended a competitor first and mentioned my client third, with a caveat about pricing that was two years out of date.

That is the whole problem in one story. ChatGPT does not have one opinion about your brand. It has a distribution of opinions, and you only ever see one sample at a time.

Before we go further, disclosure: I build LLM Pulse, one of the tools that measures this. So take the tooling section with that in mind. I have tried to be straight about where other tools beat mine, because they do in specific cases.

Let me walk through how to actually track your brand in ChatGPT, what is worth measuring, and where manual checking falls apart.

What you are actually trying to measure

"Track my brand in ChatGPT" sounds simple until you try to write it down as a number. There are really four separate things, and people mush them together.

Presence. Does ChatGPT mention you at all for the prompts that matter? Not "does it know you exist" but "when someone asks the buying question, do you show up". A cybersecurity vendor does not care whether ChatGPT can define their company. They care whether "best endpoint protection for a 200 person company" surfaces their name.

Position and share of voice. When you do show up, are you first or fifth? Are you named alongside four competitors or twelve? Share of voice is the percentage of relevant answers where you appear versus where your rivals do. It is the closest thing to a ranking that exists here.

Sentiment and framing. This is the one SEO people underrate. ChatGPT does not just list you, it characterises you. "Affordable but limited." "The enterprise option, expensive." "Great support, dated interface." Those adjectives travel. If the model has decided you are the cheap-and-cheerful choice, that framing shows up every time someone asks.

Citations. Which sources did ChatGPT lean on to say what it said? When it browses, it cites pages. Those pages are your leverage. Change what they say and you change the answer.

If you only measure one, measure presence for your money prompts. If you can measure two, add sentiment, because it is the one you can least afford to get wrong and the hardest to notice by eye.

How ChatGPT picks what to say about you

You cannot improve a system you treat as a black box, so here is the honest mechanical version, without pretending I have inside knowledge of OpenAI's stack.

ChatGPT answers about your brand from two places. The first is the model's trained-in memory, which is a compressed blur of everything it read up to its cutoff. This is where stale pricing and old positioning live. You cannot edit it directly. You can only influence the next training round by changing what the wider web says about you over months.

The second is retrieval. When ChatGPT browses, uses search, or pulls from its shopping and product data, it fetches live pages and summarises them. This is the part you can move in weeks, not years. It is also the part that produces citations you can inspect.

So the practical model is: the base model sets the default story, and retrieval either confirms it or corrects it. Your job is to make sure that when ChatGPT does go and look, the pages it finds are current, clear and say the thing you want said in plain language a model can lift.

One detail that matters. ChatGPT tends to reach for pages that read like reference material: comparison articles, review roundups, structured docs, Reddit and forum threads where real people describe you. A polished homepage full of slogans is close to useless here. A boring, specific, factual page is gold.

Setting up your prompts, properly

The mistake everyone makes is testing their brand name. "What is Acme?" is the least useful prompt you can run, because the people who type that already know you. You want the prompts of people who do not know you yet and are about to choose.

Build your prompt set in three layers.

Category prompts. No brand named. "Best project management tool for a design agency." "Cheapest way to send transactional email." These tell you whether you exist in the model's consideration set at all. They are the harshest and most important test.

Comparison prompts. "Notion vs Acme for a small team." "Alternatives to Acme." These reveal your framing against specific rivals and often surface the sentiment language most clearly.

Branded prompts. "Is Acme any good?" "Acme pricing." These check accuracy. This is where you catch the two-year-old price and the feature you sunset last spring.

Aim for coverage, not volume. Twenty to fifty prompts that mirror how your actual buyers ask will teach you more than five hundred generic ones. Write them the way a real person types, half-formed and specific, not like keywords.

Then run each one several times, ideally on different days. ChatGPT's answers vary between sessions and shift as the model and its data update. A single check is a single sample. One screenshot proves nothing.

Checking citations, by hand

Once you have an answer that mentions you, or worse, one that mentions a competitor instead, look at the sources. When ChatGPT browses it shows the pages it used. Open them.

You are looking for three things.

Which domains keep appearing. If the same three review sites and one Reddit thread drive most answers in your category, those four URLs are your priority list. Getting mentioned accurately there beats almost anything you can do on your own site.

What those pages actually say about you. Often the answer is "nothing", and that is the problem. You are absent from the exact pages the model trusts. Sometimes it is worse: the page says something wrong, and ChatGPT is faithfully repeating it.

Whether your own pages get cited at all. If ChatGPT never pulls from your site for your own branded questions, your pages are either hard to retrieve, hard to parse, or not saying anything quotable.

This manual pass is genuinely useful the first time. Do it yourself before you buy any tool. You will understand your own situation better than any dashboard can explain it.

Why doing this by hand stops working fast

Here is where I will make the case for tooling, and it is a practical case, not a sales one.

Manual checking breaks on four fronts at once.

Sampling. To know your real share of voice you need each prompt run many times across days. Thirty prompts run daily is nine hundred checks a month. Nobody does that by hand for more than a week.

Memory. You cannot see a trend from screenshots in a folder. Did your sentiment improve after you fixed that comparison page? You need the before, and you did not save it consistently, because humans do not.

Blind spots. You will unconsciously test the prompts where you already do well. The prompts that matter are the ones where you lose, and those are exactly the ones you avoid typing.

No line to revenue. A screenshot cannot tell you whether the ChatGPT mention sent anyone to your site or sold anything. That is the question your boss or client actually asks, and manual checking has no answer.

At that point you either hire someone to run checks all day, which is expensive and boring, or you get a tool that does the sampling and the memory for you.

Where a tool earns its keep

A proper AI visibility tool runs your prompts on a schedule across the models, stores every answer, and turns thousands of raw responses into presence, share of voice, sentiment and citation trends you can actually read. That is the baseline. Most tools in this space do that competently now.

What I care about, and what I built LLM Pulse around, are the parts that are still rare.

The first is sentiment down to the exact wording. It is not enough to know ChatGPT is "70% positive" about you. I want the literal phrases: the sentence where it calls you expensive, the line where it repeats a discontinued limitation, the adjective it reaches for when a competitor is in the room. You cannot fix framing you cannot see, and a score hides the words. LLM Pulse surfaces the actual language and flags the negative or plain wrong ones so you know which page to go fix.

The second is tying mentions to real traffic. Connect GA4 or Plausible and you see the visits, conversions and sales that ChatGPT, Perplexity and Gemini actually sent, not just where you got named. This is the difference between "we appear more" and "appearing more made us money". It is the number that ends the argument about whether any of this matters.

Around those two, the useful boring stuff: which pages ChatGPT and other bots crawl (GPTBot, ClaudeBot, PerplexityBot and the rest, per page), concrete next steps rather than a wall of metrics, and a GEO Writer to fix the pages that are letting you down.

Two things I will point out because they are unusually practical and I do not see them enough elsewhere. LLM Pulse tracks five core models on every plan (ChatGPT, Perplexity, Gemini, Google AI Mode, Google AI Overviews), with Claude, Copilot, Grok, DeepSeek and Meta AI addable, without an enterprise call to unlock them. And there is a hosted MCP server with 45+ tools over OAuth, included on every plan and even the trial, so you can pull your own data straight into Claude or ChatGPT and ask questions of it in plain English. That last one is still genuinely unusual in this category.

Pricing is public, which in this market is not a given. It starts at 49 €/month, and seats are unlimited on every plan, so you are not taxed for adding your team or a client.

Prices change, so check the site before you buy. There is a 14-day trial and the MCP server is in it.

The honest comparison

I am not going to tell you LLM Pulse wins for everyone, because it does not.

If you are a Fortune 500 brand that needs SOC 2, SSO, on-prem deployment and the deepest data set in the category, look at Profound first. It is the enterprise heavyweight, its prompt volume and agent analytics data go deeper than anyone's, and if you have the budget and the procurement process to match, it earns the spend. Pricing is demo-led with no public numbers, which tells you who it is for.

If you run an agency or a team and want strong regional and multi-language tracking with per-country competitor comparison, Peec AI is a serious option from around $89/month with unlimited seats. Watch that extra models are paid add-ons priced per model, which changes the maths if you want broad coverage.

If you just want the cheapest honest way in and you like seeing real screenshots of AI outputs, Otterly.ai starts at $29/month and ships an MCP server and a public API too.

And if you already pay for Ahrefs or Semrush, both now bundle AI-visibility features. Shallower than the specialists, but already in your subscription, and for a lot of readers that is the right answer this quarter.

Where I think LLM Pulse is the honest pick is the middle: you are past manual checking, you want exact-wording sentiment and a real line to GA4 traffic, you do not want a per-seat bill or a per-model surcharge, and you would rather read public pricing than sit through a demo to learn what it costs.

Start here on Monday

Run your ten hardest category prompts by hand, the ones where you are afraid of the answer. Open every source ChatGPT cites and write down the four domains that keep appearing. Fix the one page with the most wrong thing on it. Then, if the checking is eating your week, put a tool on the sampling so you can go do the fixing.

ChatGPT is already telling thousands of people what you are. The only real question is whether you have read the reviews it is writing about you.