Customgpt

CustomGPT runs its own analysis over the conversations your agent has had and reports the common intents, the topics that keep coming up, and the sentiment behind them. Your assistant can pull that alongside the raw conversation counts and the day-by-day chart data, then read the actual messages from the conversations that stand out. That is the difference between a number that went up and knowing what people wanted.

What you end up with. A written account of what your users are asking for, with the real conversations behind each pattern.

Start with it bringing you the information and you deciding.

Never without. Read a handful of the underlying conversations yourself, not only the summary. The intelligence report is one model's reading of your users, and it will smooth over the odd request that is actually the interesting one.

Built on: Get Customer Intelligence Report / Get Conversation Analytics / Get Analytics Chart Data / List Conversation Messages

Your assistant can list every source feeding the agent, every document those sources produced, and the metadata on each one including its title, source URL, and word count. It can filter by crawl and index status, so the pages that silently failed to index show up as what they are. Pointed at an agent nobody has audited, it will tell you what the agent is actually answering from.

What you end up with. An inventory of the agent's sources and documents, with the failed and stale ones separated out.

Start with it bringing you the information and you deciding.

Never without. Check the index status and not just the document count. A source can be connected and listed while the pages under it failed to crawl, and the agent will answer confidently from whatever did make it in.

Built on: List Agent Sources / List Agent Documents / Get Page Metadata / Get CustomGPT Project

Leads captured in conversations can be exported with their email addresses, names, phone numbers and any custom fields, filtered to a date range. Your assistant can pull the ones from this week and hand you a list ready for whoever follows up, rather than the export sitting in a dashboard nobody opens. It reads the leads; it does not write them anywhere.

What you end up with. This period's captured leads as a list you can hand straight to sales, with the date range stated on it.

Start with it bringing you the information and you deciding.

Never without. Set the date range yourself and know where the output lands. These are contact details for real people who typed them into a chat window, and pulling the full history when you wanted the last seven days is a bigger disclosure than you intended to make.

Built on: Export Leads / Get Conversation Analytics / List CustomGPT Projects

There is a verification action that fact-checks a message against the agent's own source documents and reports which claims are supported, partly supported, or not supported at all. Your assistant can walk the recent conversations, read each answer with its citations and trust score, run verification on the ones that look thin, and record thumbs up or down as it goes. The feedback is a signal, not a change to the agent.

What you end up with. A reviewed set of recent answers with the unsupported claims named, and the weak ones flagged in the product.

Start with it doing the work and checking with you.

Never without. Spot-check the verification against the source document yourself. This is a model checking a model, and an answer marked as supported can still be supported by a page that is two years old and wrong.

Built on: List Conversation Messages / Get Message / Get Message Trust Score / Verify Message Accuracy / Submit Message Feedback

Documents that came from a URL can be re-crawled individually, so a page you just updated stops being answered from its old version. Your assistant can list the documents, find the ones whose source has moved on, and reindex them. Source-level settings such as auto-sync frequency and crawl depth can also be adjusted, which is the version that stops you doing this by hand every month.

What you end up with. The agent answering from the current version of pages you changed, plus a sync schedule that keeps it that way.

Start with it doing the work and checking with you.

Never without. Look at what the crawl settings control before changing them. Source settings decide which pages get added and which get dropped on a refresh, so a widened crawl depth can pull in pages you never meant the agent to quote.

Built on: List Agent Documents / Get Page Metadata / Reindex Page / Update Source Settings

New content goes in as a source, by sitemap URL, file upload, or a connected integration, and indexing starts as soon as it is added. Your assistant can check your account's remaining storage credits first, add the source, then watch the project until the processing finishes and tell you what actually got indexed. That last part matters, because a source that is added is not yet a source that answers questions.

What you end up with. New content live in the agent, with a count of the documents that indexed successfully and the ones that did not.

Start with it doing the work and checking with you.

Never without. Check your storage credits and your crawl scope before adding a sitemap. Indexing is charged against the characters you index, and pointing an agent at a whole domain when you meant one documentation section is how a month's quota goes in an afternoon.

Built on: Get Usage Limits / Add Source to Project / List Agent Sources / Get CustomGPT Project

The safe shape here is to clone the agent first, which copies the knowledge base, persona and settings, and change the copy. Settings can be read before and after, so you have a record of what moved. Persona changes are versioned rather than overwritten, and an earlier version can be made current again, though CustomGPT documents that rollback as needing a Custom plan.

What you end up with. A tested change to the agent's behavior, made on a copy first, with the old configuration written down and recoverable.

Start with it doing the work and checking with you.

Never without. Save the current settings before you change anything, and confirm which project ID it is writing to. Updating settings only touches the fields you name, but this agent is answering your customers right now, and a persona edit applied to the live project instead of the clone is live the moment it saves.

Built on: Clone CustomGPT Project / Get Project Settings / Update Project Settings / List Persona Versions / Activate Persona Version

Documents can be deleted from the knowledge base one at a time, and whole sources can be disconnected along with everything they brought in. The useful version of this is mostly the reading beforehand: the assistant lists the documents, shows you which ones are stale, and removes only what you marked. CustomGPT's own note on the document deletion action says it cannot be undone.

What you end up with. The outdated content out of the agent's reach, and a written list of exactly what was removed.

Start with it doing the work and checking with you.

Never without. Approve the document list yourself, and never delete a source when you meant to delete a page. CustomGPT states that removing a document cannot be undone, and deleting a source takes every document it produced with it, so restoring means re-adding the content and paying the indexing credits again.

Built on: List Agent Documents / Get Page Metadata / Delete Page from Agent / Delete CustomGPT Source