# How to find customer pain points in webinars with Claude

> Claude is good at the second half of this job and cannot do the first half at all.

Canonical URL: https://vivu.ai/guide/how-to-find-customer-pain-points-in-webinars

Claude is good at the second half of this job and cannot do the first half at all. Give it the passages where customers described their problems and it will group them, name the patterns, and draft the messaging. It cannot watch six hours of webinar to find those passages. So the question is really how a described moment, rather than a keyword, gets turned into a list of places to look, and then who checks them.

## Why keyword search misses the thing you want

Customers almost never use the word you would search for. Someone describing a painful handoff says it took three weeks to get anyone to look at it, and nothing in that sentence matches a search for onboarding friction. Transcript search finds strings, and pain points are shapes. This is the gap that makes people fall back on rewatching at double speed, which works and does not scale past a couple of recordings.

## The routes

Rewatching with a notepad is the honest baseline. It is thorough, it is the only method where you definitely did not miss anything, and it costs an afternoon per recording.

Transcripts plus a text search get you to the recordings faster and inherit the keyword problem above. [Searching by what was said](https://vivu.ai/guide/search-video-by-spoken-words) is genuinely useful when you know the phrase, which is a different situation from this one.

Feeding whole transcripts to an assistant and asking for themes produces a plausible summary and loses the timing, and you cannot check it without going back to the recording anyway. It is also the point where a summary quietly invents an emphasis nobody made.

Building your own retrieval over chunked transcripts works and is a project. The fourth route is a hosted retrieval layer the assistant can query itself, so the finding and the analysis happen in one conversation.

## Ask for the passages, then let Claude group them

[Vivu](https://vivu.ai/mcp) can be added to Claude as a custom connector. With your webinar recordings uploaded into a project and indexed, you ask for the shape you are after rather than the word: the part where a customer describes what went wrong with their current setup, or the moment someone in the audience pushes back on the approach. What comes back is a set of openable time ranges, each with a line saying why it matched, and the result page lets you preview them one by one. Claude does the grouping and the write-up from the ones you confirm, which means the analysis you hand to a colleague has timecodes attached to it rather than a claim you would have to defend from memory.

Two habits make this work better. Write the subject into the query, because the words get read fairly literally: asking for a moment where someone disagrees can return a polite correction about the running order rather than a substantive objection, and adding what the disagreement is about narrows it. And expect the ranges to come back tight, sometimes only a few seconds, so open each one a little earlier than it starts if you want the context that makes the quote make sense.

## When the search comes back empty

An empty result does not tell you whether the thing never happened or your query was too specific. Those are different problems and the output looks identical. The move is to re-ask in plainer, broader language before concluding the recording does not contain it, and to accept that one honest no is worth more than ten matches you have not opened.

## When you do not need any of this

If you ran the webinar last week and you were on the call, you already know what the three interesting moments were, and writing them down beats any tooling. The same goes for a single recording with a short Q&A. This becomes worth the setup when you are mining a quarter's worth of sessions, or when the person doing the analysis was not in the room for any of them, which is the normal case for anyone working from an archive somebody else built.

## Where this route stops

Each search runs inside a single project, so recordings you want compared have to live together. Footage has to be uploaded and indexed in the cloud before it is searchable, and there is an allowance on searching, so vague queries have a cost. What comes back are ranges to open rather than a verbatim quote index, which matters if what you needed was the exact sentence with its exact wording: you still get that by watching the range and writing it down. Building a quote library from this is a manual last step, and there is no way around it that does not involve trusting a paraphrase.

## How to decide whether this is your problem

The signal is where the friction actually sits. If your messaging work is slow because nobody has decided what the story is, better retrieval changes nothing. If it is slow because the raw material exists, in recordings, and reaching it costs a day per question, then the bottleneck is the finding, and it will keep getting worse as the archive grows. A good test is to pick one question you already know the answer to and see how long it takes to produce the evidence for it. That number is what you are deciding about. [Finding footage you know exists](https://vivu.ai/guide/how-to-find-an-old-brand-video-we-already-made) is the same bottleneck with a lower stakes example.

## FAQ

### Can I search all of our webinar recordings at once?

You can search everything inside one project at a time, and recordings placed in the same project are searched together, so results can come back from several different videos in one pass. What you cannot do is ask one question that spans separate projects. If comparing sessions across a whole year matters to you, put them in the same project when you upload rather than splitting them by quarter.

### Why did my search return nothing?

Either the moment is not in the footage or the query was written too narrowly, and the empty result looks the same either way. Re-ask in plainer language, describing the situation rather than naming a concept, before you conclude the recording does not contain it. A query built out of internal vocabulary is the most common cause, because customers describe problems in their own words and not in yours.

### Does this give me the exact quote for a slide?

It gives you where to look. The range comes back with a short reason attached so you can tell whether it is worth opening, and the wording you put on a slide should come from watching it. Treat the reason text as a scanning aid rather than a transcript, particularly if the quote is going in front of a customer.

### Is a summary of the transcript good enough instead?

For getting a feel for a session, often yes. For pain points, a summary is the wrong shape, because it flattens the one uncomfortable thirty second answer into a bullet that sounds like everything else, and it strips the timing so you cannot go back and check. If the output is going to influence positioning, you want the passages themselves, watched by a person.