Simulating a buyer's three-turn conversation with ChatGPT, Gemini and Perplexity

Why one prompt is not the same test as a conversation

Most AI visibility measurement, including most of what a tool like this one does most of the time, asks one question and reads one answer. That is the right unit for building a statistically sound rate across many prompts: ask the same kind of question many times, across many topics, and count how often a brand comes up. It is also, by itself, not the same thing as how a real buyer actually uses an AI assistant.

A real buyer asks a broad question, reads the answer, and then asks a follow-up that depends on what the assistant just said. 'What's the best way to handle X' becomes 'what tools do that' becomes 'would you recommend this one specifically'. Each later question is shaped by the answer before it. Testing only the first, isolated question misses whatever happens once the conversation actually develops, which is often where a recommendation either holds up or falls apart.

What 'real conversation history' means, technically

Running three questions back to back is not the same as running one conversation. The difference is whether each later question is sent to the model together with the actual prior turns, formatted as real conversation history, the way the engine's own chat interface would send it. ChatGPT's and Perplexity's chat APIs accept a list of prior messages with their roles. Gemini's API accepts the same idea as a list of prior turns. A journey builds that list turn by turn: the first question goes out alone, then the second question is sent together with the first question and the engine's own first answer as history, then the third question is sent with the full two-turn history before it.

This matters because an engine's second answer, given the real history, is not necessarily the same as what it would say to that same second question asked cold, with no context. A model that recommended a brand in its first answer may double down, qualify, or reverse course once asked directly whether it would still recommend that brand, and that shift is only visible if the first answer was genuinely part of what the model was conditioned on when producing the second.

Three scripted starting points, not an improvised chatbot

Each journey follows one of three fixed templates. The three questions in a template are written in advance and do not change based on what the engine answers; only the history changes, because the conversation genuinely carries forward. This is a deliberate, honest simplification: generating a follow-up question dynamically from the previous answer is a harder, separate problem, and claiming to do that without actually doing it would be the kind of overclaim this whole approach tries to avoid.

Checking each turn without trusting the engine's word for it

Whether a brand was named at each turn is not something the engine is asked to self-report. Each answer is checked the same deterministic way a normal audit run is: the brand's own name and known aliases are matched against the literal answer text, independent of anything the model claims about itself. This keeps a journey's results comparable to everything else measured on the platform, rather than introducing a second, differently-trusted way of deciding whether a brand was mentioned.

Why only three engines, out of five

Of the five AI surfaces this platform observes, only three, ChatGPT, Gemini and Perplexity, are genuine conversational assistants with an actual multi-turn session to extend. Google's AI Overview and AI Mode are observed as a single search-results-page snapshot for one query at a time. There is no conversation thread underneath that snapshot to carry a second, context-aware question into, so a journey is not run on those two surfaces. This is a real, structural limitation of what those two surfaces are, not a product choice that could be lifted later without the underlying surfaces themselves changing.

What a journey costs, and why

Each turn in a journey is a real call to a real engine, exactly like a single audit prompt run, and is metered the same way: one credit per turn, three credits for a three-turn journey. There is no separate, cheaper tier for this kind of test, because there is no cheaper way to get a genuine, context-aware answer out of a real model.