Turning customer reviews into a useful product pulse.

From public reviews to a weekly product pulse.

ChatGPT Review Intelligence is an approval-gated workflow that turns public customer feedback into two operational outputs: an internal weekly product pulse and a reusable support clarification. The supplied project documents a working flow built with n8n, Gemini, Google Drive, Google Docs and Gmail. It connects review analysis to a practical handoff while keeping a person in control of the final write actions.

The input contains 500 ChatGPT Android Google Play reviews dated 7 July to 29 September 2026. The India store-market parameter describes how reviews were retrieved; it does not establish where each reviewer lives. The project is an independent exploration, not work commissioned by OpenAI, and the review sample should not be treated as representative of every user.

Reading feedback is only the first step.

Product and Support teams need different things from the same feedback. Product needs patterns and concrete opportunities. Support needs an accurate explanation that can be reused when a question returns. Manually moving from scattered reviews to those outputs can be slow, inconsistent and difficult to trace back to the original evidence.

The scope focuses on recent public reviews and one recurring payment-related confusion. It does not attempt to resolve individual customer accounts. Every useful output needs to preserve the distinction between a reported experience, an interpretation of that experience and an official policy explanation.

Keep the evidence attached to the theme.

The documented analysis surfaced three leading themes: positive general feedback, study and academic utility, and performance limitations or feature constraints. It also identified confusion between free-tier limits, paid advanced services and subscription prompts. These are signals from this particular review set, not prevalence estimates for the wider product population.

Theme counts were generated by AI rather than manually coded research. The workflow asks for verbatim quotes and review IDs so a reviewer can inspect the evidence. This supports a narrower product observation: when a limit appears, people may need clearer context about what happened and what they can do next.

Bound the output before connecting the tools.

The implementation splits 500 reviews into five batches of 100 to make context handling more manageable. Each analysis returns structured JSON with no more than five themes, estimated counts and supporting quotes. A second synthesis step merges those outputs into a maximum of five final themes and selects the top three.

The final contract requires exactly three supporting quotes, a product pulse under 250 words and five neutral clarification bullets. Official OpenAI Help Center sources constrain the policy explanation, accompanied by source links and a last-checked date. These constraints reduce ambiguity but still require human verification; a structured answer can remain factually wrong.

A deliberate pause before writing.

After generation, the n8n workflow pauses at an approval form. Approved and Declined are explicit choices. Approval appends the structured update to a Google Docs product-pulse log and creates a Gmail draft containing the pulse, clarification and sources. The project documents an approved execution that completed both actions without sending an email.

The generated action ideas included communicating free-tier limits earlier, explaining subscription requirements more clearly and surfacing relevant support links. They remain suggestions for a team to assess. Creating a document and a draft demonstrates the workflow connection, not evidence that the ideas improved the product or reduced support demand.

Evaluate the handoff as well as the answer.

The next evaluation pass should verify quote accuracy, review-ID traceability, policy-link validity and output length. It should also exercise both approval paths: declining must prevent downstream writes, while approving should create the intended outputs without sending mail. These are proposed checks, not a reported accuracy score.

Automation needs an inspectable trail.

The strongest lesson is that useful automation includes evidence, output constraints and a clear human decision point. Next steps in the study include weekly sentiment trends, a review-evidence table, a dashboard and routing different issue types to the appropriate team. Before extending the workflow, I would make its evidence checks and approval behaviour easier to inspect and repeat.