How the review-mining pipeline builds a pain ranking

Independent customer reviews in, ranked pains and page-level copy gaps out. Runs on the web only: there is no MCP tool for this yet.

Signature

This instrument has no MCP tool. It runs on the web only, so you drive it from the page rather than from Claude or ChatGPT.

What you ask

Prompts land with the captured session, so that what is printed here is what actually worked rather than what should have.

What runs

Nothing fires over MCP for this one. The pipeline runs server-side when you start it from the web page.

What comes back

No session to show

This instrument has no MCP tool, so there is no chat session to capture. What it produces is the web report itself, described above.

When to reach for this

Use this when you know the page fails and not what to say instead

Analytics tells you the page loses people. It cannot tell you which sentence is missing. This mines independent customer reviews in your category, ranks the pains buyers actually raise, then checks your page against that ranking and returns the gaps as copy to write.

Not this one if you have never run a page check

A page that fails on layout, speed or a broken call to action does not have a copy problem yet. Fix what is visibly broken first; wording is the layer underneath.

Not this one for your own reviews

It reads independent sources rather than testimonials you control, which is the point: the pains people write when they are not talking to you are the ones your page has to answer.

How to read it

Frequency is counted, not estimated

Earlier versions asked a model how often a pain came up, and got a number it invented. The pipeline now counts occurrences across the collected evidence and keeps the quotes, so a pain ranked first is first because it appeared most, and you can read the sentences it appeared in.

A low-confidence run is flagged rather than hidden

There are two quality gates: a minimum amount of evidence, and a fidelity check on how much of the corpus is vendor-written rather than customer-written. A run that fails either still completes, and is labelled low confidence. Read that label before you rewrite a page around the output.

The gap matters more than the ranking

The ranking tells you what customers care about. The value is the intersection: the pains ranked high that your page never mentions. A pain your copy already answers well is not a task, however often it appears.

Source mix is worth checking

The corpus is only as good as where it came from. A run drawn overwhelmingly from one platform reflects that platform's population, not your buyers. The report keeps its sources so you can see the mix rather than trusting the summary.

Limits

Web only: there is no MCP tool for this, so it cannot be run from Claude or ChatGPT. The reasons are runtime and cost rather than oversight — a run is deep research in background mode with a 25-minute ceiling and a real per-run price, and both are gated at the web trigger.
It runs on a fair-use cap during the free beta, per account and globally per day.
It reads what the internet says about your category. A product with almost no independent discussion will produce a thin corpus, and the run will say so rather than invent one.
It finds the gap between customer language and your page. It does not write the page, and it does not measure whether the rewrite worked. Log the change and verify it afterwards.
Reviews skew toward the extremes. Loud complaints are over-represented against quiet satisfaction, which is useful for finding objections and misleading as a satisfaction measure.

Page checkMeasures Core Web Vitals through PageSpeed Insights and reads a mobile screenshot against the conversion heuristics. Asynchronous: acknowledge, then poll.Visual proofCaptures real desktop and mobile screenshots through an anti-bot browser so the model looks at the page instead of guessing from HTML, then records what it observed.Full conversion auditOne prompt, one call: headline KPIs against the prior period, the biggest funnel leak, anomalies with significance testing, and ranked findings matched to the hypothesis library.

Go deeper

Run itOpen the voice of customer tool ReadReview Mining: Turn Customer Reviews into Landing Page Copy ProtocolEndpoint, OAuth scopes and client configs

Try it on your own data

Every instrument runs on your own GA4, read-only, inside Claude or ChatGPT. One connector covers all of them.

Add the connector →

Session captured 2026-08-03 against the ConvRadar demo property. Masked in the image: cropped to the conversation column, so the account sidebar and workspace name are out of frame, no connector URL is in frame, so no token can leak, no account email or avatar is in frame.