Direct answer
What's the difference between a simulated AI answer and a real observed one?
A simulated AI answer is generated to model likely behavior, at no cost, without a live request to the AI system being tested. An observed answer comes from an actual request — either a live, search-grounded query or a captured response from a real session with ChatGPT, Claude, Gemini, Perplexity or Copilot. Simulation is a useful analysis tool. It is not evidence of what an AI system actually said, and shouldn't be reported as if it were.
01
Simulation estimates likely behavior; it doesn't prove what a real AI system would say to a real user.
02
A browser companion capturing an actual assistant session is a stronger evidence class than any simulation.
03
Any AI visibility report should label which of these evidence classes produced each finding.
A lot of "AI visibility scores" you'll come across are built on something that never actually asked an AI system anything.
That's not automatically dishonest — simulation is a legitimate way to model likely behavior, and it can be useful for exploring a lot of prompts cheaply. The problem starts when a simulated result gets presented with the same confidence as something a real AI system actually said. Those are two different kinds of evidence, and they deserve to be labeled differently.
Three different things people call "an AI answer"
The first is a live, search-grounded request: an actual call to an AI system with retrieval turned on, so the answer reflects information the system looked up at that moment. This is the strongest kind of evidence available, because it's as close as you can get to watching the system do the thing you're trying to measure.
The second is a captured observation from a real session — someone actually using ChatGPT, Claude, Gemini, Perplexity or Copilot, with a browser companion recording what came back. This isn't a synthetic test at all. It's a real interaction that happened, preserved as evidence.
The third is simulation: using an AI system to estimate what a model would probably say, without a live request to the actual target system. Growthract has a $0 AI-router-based simulation capability for exactly this — it's useful for exploring likely behavior across many prompts without incurring cost. But it is never a substitute for an observed result, and it should never be reported as one.
Why blurring these together causes real damage
If a report tells you "ChatGPT mentions your brand 40% of the time" and that number came entirely from simulation, you're not looking at ChatGPT's actual behavior. You're looking at a model's guess about ChatGPT's behavior. That guess might be reasonable, or it might be systematically wrong in ways that are hard to detect from the outside — and you have no way to tell which, because the report didn't say where the number came from.
This matters most when the finding is negative. A simulated "your brand doesn't come up" result is a much weaker claim than an observed one, and treating it as equally solid can send a team chasing a problem that may not exist in the real system at all — or missing one that does.
What an honest version of this looks like
The fix isn't to throw out simulation — it's genuinely useful for casting a wide net cheaply. The fix is labeling. Every finding should say plainly which of the three evidence classes produced it: grounded observation, captured real-session observation, or simulation. A reader should never have to guess.
This is closely related to a distinction we cover separately: even among genuine observed requests, some come back grounded and some come back without confirmed retrieval, and some don't complete at all. That's covered in Unavailable Isn't the Same as Not Mentioned. Both distinctions exist for the same reason: an AI visibility report is only as trustworthy as its willingness to tell you exactly how confident you should be in each individual finding.
If you want to run this kind of test yourself before relying on any tool's output, a practical starting point is a small, repeatable set of prompts across the real assistants your buyers actually use — covered in 12 ChatGPT & Perplexity Prompts to Audit Your Brand's AI Search Visibility.
Know which evidence you're looking at
Growthract labels grounded, captured and simulated results separately.
The browser companion can capture real observations from actual supported AI interfaces, distinct from anything simulated.
A couple of follow-up questions
Is simulation ever the right tool?
Yes, for exploring a wide range of prompts quickly and cheaply, or for supporting analysis alongside real evidence. The issue isn't using it — it's presenting it as something it isn't.
How do I know if a tool I'm using is simulating or observing?
Ask directly. If a vendor can't clearly explain whether a given number came from a live request, a captured session, or a simulation, treat the result with caution.
Continue exploring
Related insights
Case study
An illustrative strategy showing how a Series B fintech could improve citation coverage from 10% to 34% across a controlled set of high-intent ChatGPT and Perplexity prompts.
See the case study →
Insight
A failed AI visibility check and a genuine 'no mention' look identical on the surface but mean completely different things. Here's why that distinction matters and how to test for it.
Read the insight →
Insight
A copy-paste set of ChatGPT and Perplexity prompts you can run today to see whether — and how — your SaaS shows up when AI assistants answer category, comparison and vendor questions.
Read the insight →