Over the past year we have watched a pattern repeat across our own business and the clients of our agency, Grow and Convert. Customers arrive after long, heavily personalized conversations with AI tools. We know because we’ve asked them to share those chats.
That is good news on its own, but it creates a reproducibility problem. If a lead comes from a unique, personal conversation with ChatGPT, how do you get more of them? You can’t see these conversations in any AI visibility tool. Nothing can show you the three-page exchange your latest lead had with ChatGPT, because no one outside OpenAI has that data. The conversations are private and personalized, and the model often knows a great deal about the user’s situation.
Company after company tells us the same thing. The leads keep coming, but they don’t know what people are typing to get them. We call these invisible prompts.
This is a Traqer version of a framework first published on the Grow and Convert blog. You can read the original article here. Traqer is the AI visibility tool built by the Grow and Convert team.
What are invisible prompts?
By invisible prompts we do not mean small wording variants you forgot to track, like “what are some good content agencies?” against “find me the best content agencies.” Adding more phrasings to a tracker does not solve this. The issue runs deeper than wording.
Even if a real user types the exact prompt you’re tracking, ChatGPT, Claude, or Gemini will factor in a large amount of personal context you cannot predict, and that context changes the recommendation. We think of it as the difference between the literal prompt and the effective prompt. The literal prompt is what the user types. The effective prompt includes everything the model already knows about them.
Someone might type “find me a good content marketing agency.” What the model works with is everything it knows about that person: their company size, budget, industry, what they’ve tried before, who is already on their team, and more. The literal prompt is one line. The effective prompt is closer to a long brief about their situation that ends with “given all of that, what agency is right for me?”
That is not trackable in the way marketers are used to, where prompts behave like keywords with search volumes. We’ve moved to a setting where the search volume of most prompts is one.
Tools that claim to show “real” prompts, including ones drawing on data from millions of chats, are summarizing patterns from people who volunteered to have their browsing tracked. They still describe a world of keywords and volumes. What they do not capture is the personalization underneath each prompt, and that personalization helps decide whether you get recommended.
An example from the real world
One of our clients found Grow and Convert through ChatGPT. Her literal prompt was along the lines of “what are the best SEO companies that have driven strong results and steady leads for people in our space?” From the response, it was clear the model already understood her space, her goals, and her constraints. She had not typed those details into that prompt.
If we tracked that exact prompt in a visibility tool, we would not see ourselves mentioned, even though in her real session we were. We confirmed this. In a fresh incognito ChatGPT window, her prompt, even with her industry named explicitly, produced a different answer that did not mention us at all.
The real prompt, with all of the context shaping the recommendation, is longer and more personal than anything a tool will show you.
Why invisible prompts are only going to grow
The trend is moving in one direction as AI tools become more embedded in people’s work and personal lives. More people log in and stay logged in. Few use ChatGPT or Claude in incognito.
These tools are building longer histories of each user. People are giving agents access to their email and calendar. Context windows are growing, so the tools factor in more information, and memory features carry details from past chats into new ones. A model knows more about a user who has been chatting with it for a year than one who signed up yesterday, and that history shapes recommendations. The gap between the literal prompt and the effective prompt keeps widening.
How to earn more invisible prompt leads
The answer is the same one we have given since early 2025, when we first argued that AI search would reward ultra-specific content: create ultra-specific content that teaches LLMs the use cases and pain points where your product fits. You cannot predict the searches in advance, so you give the models the material they need to recommend you across many specific situations.
We didn’t know our financial sector client would ask ChatGPT for an SEO agency, or how she would phrase it. We had not written anything specifically for her prompt. We had spent years publishing about why most content marketing fails to generate leads and how we fix it. That body of content is partly what allowed the model to connect us to her situation.
Do the same across as many specific scenarios as you can. Start where you clearly perform well, and with real case studies. Then cover features, the pain points you solve, and the situations where your product is the right fit.
Why owned content comes first
This is why owned content sits at the top tier of our Prioritized GEO framework. Your own site gives you the room to describe where and when you are a fit, in detail. A mention in a single Reddit sentence does not give a model enough context to connect you to a specific user’s needs. You can build that content foundation deliberately on your own site; you cannot do that in someone else’s comment thread.
What not to do: chase single-prompt visibility
The same logic explains why on-site tweaks like schema or an llms.txt file do little here. They do not help you cover unique scenarios. It also means engineering your way to visibility for one very specific prompt is probably a poor use of effort, because few people will ever type that exact prompt. An isolated page for one narrow prompt might look good inside a visibility tracker, which can please a client or a manager, but if no one has that exact exchange it will not help in practice.
How to measure progress when the prompts are invisible
Measurement isn’t hopeless. You cannot know the exact prompts, but you can watch how often you appear at the topic level, across clients and over time. This is the basis of topic-based (personalized) GEO and our philosophy underpinning the way we designed Traqer.
The differences are visible in the data. One established brand with years of SEO presence showed topic visibility of 59% on ChatGPT, 71% on Perplexity, 78% on Google AI Overviews, and 85% on Gemini. A younger brand on a newer domain, without that history, showed 16%, 13%, 24%, and 8% on the same engines. The gap is real and worth tracking.
You can also see it across topics for a single brand. Grow and Convert’s own visibility for “B2B content marketing services” looks very different from “B2C content marketing services.” We have taken on both, and our process works for both, but we have had more B2B clients, written more about B2B problems, and published more B2B case studies, so the models know us better in that context.
That difference is actionable. To win more B2C conversations, we need to publish more about our B2C work so the models have the context to recommend us there. Tracking visibility per topic and per model, rather than as a single blended score, is how you turn invisible prompts into something you can manage. The prompts stay invisible, but your presence on a topic does not.
Invisible prompts are a form of word of mouth
This resembles word of mouth, the oldest marketing channel there is. The outcome every brand wants is to be the name that comes up whenever someone asks for a product or service in its space.
Think about what happens when you ask a friend for a recommendation. The question carries all the detail of your specific situation. If you ask a knowledgeable friend which surfboard to buy, your spoken question might be short, but the answer they give accounts for everything they know about you: that you are a beginner, your height, the board you ride now. The literal question is brief. The effective question they answer is much longer. The same holds whether someone is asking about software, a law firm, clothing, or a coffee shop.
AI search works in a similar way, with ChatGPT, Claude, or Gemini in the place of the friend who knows you. The difference is that a friend knows surfboards from experience, while a model knows only what it has read. If you want LLMs to recommend you the way a knowledgeable friend would, you have to publish the content that gives them the context to do it.
You cannot predict invisible prompts. You can build the content foundation that gives LLMs enough detail to recommend you when those conversations happen.
What this means for agencies
For agencies, invisible prompts reset what you can promise and how you report. You cannot guarantee a client a ranking for a named prompt, and single-prompt screenshots are a weak basis for a retainer. The defensible approach is to build each client’s owned content across their real use cases, then track topic-level visibility per client and per model so you can show movement over time. Traqer places no limit on brands or users per account, so a portfolio of clients fits without per-brand fees.
You can read the original, longer version of this framework on the Grow and Convert blog.