What we can actually know about AI chatbot prompts: public proxies, not private feeds

No public source shows that anyone has comprehensive access to platform-wide AI chatbot prompt logs. Query fan-out and prompt rewriting are real and partly observable. Public search demand is the strongest documented proxy for likely chatbot intent. The defensible workflow is a proxy system built from public signals, usage research, and controlled API runs. It should be labelled as a proxy, not sold as a feed. This article separates what is provable from what is asserted, and treats the gap as a measurement-ethics question.

The category's prompt-tracking claims need disambiguation, not denial

AI visibility vendors say they track the prompts that matter for a brand. The implicit promise is visibility into what real users type into ChatGPT, Perplexity, and Gemini. The actual mechanism behind that promise is rarely disclosed in the pitch.

Some vendors are explicit about the limit in their own documentation. Otterly's help content recommends building a prompt set from brand terms, domains, industries, URLs, and SEO keywords Otterly AI · 2024B. That is a construction method, not a feed of observed user prompts. Peec's onboarding documentation describes the same shape of workflow, where the customer supplies the inputs the tool then runs Peec AI · 2024B.

The question is not whether prompt tracking is fraudulent. It is what the word "track" is doing. A vendor that builds a prompt set from public inputs is doing useful work. A vendor that implies it reads live user conversations is making a claim public evidence does not support.

No public source shows comprehensive access to real chatbot prompts

The strongest defensible statement here is bounded, not absolute. No public first-party source from OpenAI, Anthropic, Google, Microsoft, or Perplexity discloses a browsable feed of real user prompts. The phrasing matters: no public evidence reviewed here shows such a feed exists, which is different from claiming no non-public data exists anywhere.

The first-party documentation that does exist describes mechanisms, not dumps. OpenAI's help content explains how ChatGPT Search behaves for a user, without exposing other users' queries OpenAI · 2024A. The web-search API documentation describes what a caller can observe from their own requests, not a global query stream OpenAI · 2025A. Otterly confirms the gap from the other side. Its help content states there is no way to learn which prompts are most asked at ChatGPT or Perplexity Otterly AI · 2024B.

Any vendor implying platform-wide prompt visibility is asserting something past the public record. The honest position is the bounded one.

Four first-party findings are actually proven

Four findings from primary sources support a real but limited workflow. Each is documented by the platform itself, which puts it on firmer ground than vendor inference.

The first is search expansion. Google documents that the Trends Explore page uses Gemini to expand a seed topic into up to eight related terms and rising queries Google · 2025A. This is first-party proof that public search behaviour can be expanded into an adjacent-intent neighbourhood. The second is intent classification. OpenAI's usage research classifies consumer conversations into roughly 49% Asking, 40% Doing, and 11% Expressing Chatterji et al. · NBER / OpenAI / Harvard University · 2025A. The taxonomy is published in both the research paper and OpenAI's summary OpenAI · 2025A.

The third finding is prompt rewriting. OpenAI documents that ChatGPT Search rewrites a user query into one or more targeted queries. It may send additional, more specific queries after reviewing initial results OpenAI · 2024A. Fan-out behaviour is real inside a production chatbot search system. The fourth is partial observability. The web-search API states that the call output will usually, but not always, include the search queries that were run OpenAI · 2025A. The sources field can reveal the URLs consulted. Some fan-out is visible to an API caller; not all of it is.

Fan-out is a measurable mechanism, even when the queries stay hidden

Query fan-out is the step where one user question becomes several targeted sub-queries. A user types a single question. The system rewrites it into more specific queries, runs them, and may issue further queries after seeing early results OpenAI · 2024A. Some of those sub-queries are observable through the API output; many are not OpenAI · 2025A.

The mechanism has measurable consequences even where the specific queries stay hidden. Pages ranking for fan-out sub-queries are 161% more likely to be cited than pages ranking only for the main query Search Engine Land · 2025B. Fan-out lives inside the wider retrieval pipeline, and the citation effect is the part a brand team can act on.

Fan-out also has documented quirks that affect any reconstruction. ChatGPT Search has been observed switching to English in its fan-out queries even when the user prompts in another language Search Engine Journal · 2025B. A tool that surfaces fan-out queries does so from observable signals, not from a privileged feed Ahrefs · 2025B.

Public search data is the best proxy, not proof

Google Trends measures public web-search behaviour, not chatbot conversations. Its Gemini-assisted expansion is useful because it reveals adjacent demand and semantic neighbourhoods around a seed topic Google · 2025A. It should never be described as a direct measurement of what users typed into a ChatGPT or Gemini chat.

Search intent and chatbot intent are not identical, but they are close enough to be useful. Web search in the age of generative AI has its own documented characteristics that differ from conversational queries Kirsten et al. · Ruhr University Bochum / Max Planck Institute for Software Systems · 2025A. The shift in how attention concentrates under AI search is also documented Aral et al. · MIT IDE · 2026A. The proxy has to be read with that difference in mind. Search intent is the strongest publicly documented neighbourhood for likely chatbot intent, and that is the honest claim.

Calibrated proxy is honest research. Pretending the proxy is the thing is not. The distinction is the whole point of the method.

How vendors likely build prompt sets in practice

The most defensible public reconstruction of the vendor workflow has three steps. Vendors begin with customer-provided inputs: prompts, SEO keywords, URLs, brand terms, and public-demand proxies. They then run those prompts through targeted chatbot or search workflows. They then measure visibility, citations, and sometimes the query rewrites where the platform exposes them.

Vendor documentation supports this shape directly. Otterly's help content openly states the prompt set is built from brand terms, domains, industries, and SEO keywords Otterly AI · 2024B. Its guidance on adding prompts describes manual and assisted input, not a captured feed Otterly AI · 2024B. Peec's documentation describes the same customer-supplied input model Peec AI · 2024B. Some tools then surface the fan-out queries they can observe from those runs Ahrefs · 2025B.

Treat each vendor's help centre as evidence about that vendor's stated method, not as proof of the whole category. A vendor that admits the workflow is a proxy is more credible, not less.

The caveats an honest researcher keeps in every claim

Four bounded statements hold without overreaching. Each is the calibrated version of a claim the category often overstates.

Public search behaviour is the best available proxy for likely chatbot prompts, not a measurement identical to chatbot behaviour Google · 2025A. OpenAI's usage research is useful for classifying likely chatbot intent, not a public prompt dump Chatterji et al. · NBER / OpenAI / Harvard University · 2025A. Some query rewrites are observable to API callers, but API responses do not return every internal fan-out query OpenAI · 2025A. No public first-party source reviewed here proves comprehensive vendor access to prompt logs, which is narrower than claiming no vendor holds any non-public data anywhere.

Honest caveats strengthen the article's authority rather than weakening it. The bounded claim is the one that survives scrutiny.

A buyer's framework, not a vendor takedown

The best publicly documented workflow approximates intent from public search signals and generates prompt hypotheses from them. It classifies those hypotheses against OpenAI's published taxonomy, then inspects whatever rewrites or sources the platform returns on controlled runs OpenAI · 2025A. That is a useful proxy system. It is not direct access to global chatbot prompt demand.

The reader leaves with a usable method, not a list of vendors to distrust. Use Google Trends Explore with Gemini expansion as a seed-discovery layer Google · 2025A. Classify the resulting hypotheses against the Asking, Doing, and Expressing taxonomy OpenAI · 2025A. Run a controlled set of prompts through ChatGPT, Perplexity, Gemini, and Google AI Overviews, and observe the rewrites where available OpenAI · 2025A. When evaluating a vendor, ask one question: what is the actual source of the prompts you say you track? The answer is the test of the vendor's editorial honesty.