Capture what AI models search for, cite, and reason about, directly from the platforms where AI answers live.
How to Open
- Popup nav →
AI Insights. - Works on chatgpt.com, gemini.google.com, and Google Search when an AI Overview appears.
- Extraction starts by itself when you open the tab. There is no button to press.
- On unsupported pages you’ll see a prompt to navigate to a supported platform. The Bulk analysis button stays available from that screen.
Supported Platforms
| Platform | What’s Captured |
|---|---|
| ChatGPT | Fan-out queries, sources and their categories, sourcing pipelines, products and merchant offers, image/news/place carousels, entities, reasoning telemetry, turn intent, model type |
| Gemini | Fan-out and related queries, source attributions |
| Google AI Overviews | The Overview answer text and its citations |
Where Google AI Overview capture works. It runs on 17 Google country domains: google.com, .co.uk, .nl, .de, .fr, .es, .it, .ca, .com.au, .co.jp, .com.br, .co.in, .be, .pl, .se, .at and .ch. On any other Google domain the tab will not pick up an Overview, which is the single most common reason it looks empty on a SERP that clearly has one. This is being widened to every Google country domain in the next release.
What You See
Not every section appears for every answer. Sections with no data stay hidden rather than showing you an empty box.
Original Input Prompt
- The query you typed into the AI.
Sourcing Pipelines
- Which retrieval path each source arrived through: Bright Data, Oxylabs, Labrador (licensed content), or the open web.
- Shown as a doughnut chart, with a tooltip on each segment explaining what that pipeline is.
- This is the closest thing available to a look at how ChatGPT actually sources an answer. See Sourcing Pipelines for what each one means.
Search / Model Queries (fan-out)
- The internal queries the model ran behind your prompt.
- Typed by route, so you can tell a plain web search from a shopping lookup or an image lookup.
- Numbered, with individual copy buttons.
AI Reasoning / Thoughts
- Chain-of-thought output when the model exposes it.
- Plus reasoning telemetry: thinking duration, effort level, and the titles of each step.
Products
- Merchant offers with seller, price, rating, review count, and the product’s position in the carousel.
Images, News and Places
- Carousel items pulled out of the answer, with their source attribution.
- These depend on the model returning carousel data. Many answers come back as prose with inline citations instead, in which case these sections stay hidden.
Tool Calls and RAG Sources
- The tools the model invoked and the retrieval-augmented sources behind the answer.
Citations
- Domain, title, and full URL for each source, as clickable links.
- Categorised as citation, news, image, academia, product, hidden, or other, with a tooltip on each category badge.
Entities
- Named entities the model surfaced in the answer.
Analytics Panel
- Query count, citation count.
- Platform badge (ChatGPT / Gemini / Google AIO), the model slug, and whether the answer was a Thinking or an Instant response.
- Timestamps and conversation IDs.
Working Across Conversations
History
- Your last 30 days of captures, kept on your own machine.
- Filter by platform, model, or project, and search across the lot.
- Aggregate stats across the set: total queries, unique queries, conversations.
- Load more to page back, or delete any single conversation.
Projects
- Group conversations by client or campaign so a month of research stays sorted.
- Assign the conversation you’re looking at from the panel, or manage the full list (create, rename, delete) from the Manage projects panel. See Projects.
Brand tracking
- Type a brand or competitor name and track it.
- Tracked terms appear as mention-count tiles across your captures. Remove a tile to stop tracking it. See Brands and Competitors.
Bulk analysis
- Opens a full dashboard in its own tab that runs an entire prompt list through ChatGPT unattended, then rolls every capture into one view: top domains and URLs, a fan-out tag cloud, the sourcing-pipeline split, products, entities, places, and brand versus competitor share of voice. See Bulk Analysis Dashboard and Bulk Prompt Analysis.
Actions
- Export CSV: every section exports on its own, so you can take just the citations, just the fan-out queries, or just the products.
- Copy All Queries: clipboard all fan-out queries at once.
- Copy All Citation URLs: a quick list of source URLs.
- Refresh Data: re-extract the current conversation. The panel tells you how old the cached copy is.
Settings
One setting lives in this tab: Enhanced Google AI Overview extraction. It is off by default. Turning it on injects a network interceptor on Google Search pages that catches Overview data the page-scrape can miss. Leave it off unless AI Overview capture is coming back thin.
Workflow Wins
- See which sources AI cites without manually clicking through.
- Read the fan-out queries the model generated. They are the closest thing to a keyword list written by the retrieval system itself.
- Build fact-check lists and content briefs from citations.
- Monitor AI Overview presence on your target SERPs.
- Watch how a brand’s share of mentions moves against its competitors across a batch of prompts.
Recommended Plays
- Citation audit: Export a CSV of the sources an AI cited for a topic, then verify accuracy.
- Keyword mining: Copy the fan-out queries to seed keyword research.
- Sourcing analysis: Check the pipeline split. An answer sourced mostly from licensed content behaves very differently from one sourced from the open web, and it tells you whether being crawlable is even the lever.
- Competitive intel: Track your brand and two competitors, run a prompt set, and read the share of voice.
- Prompt engineering: Review reasoning and step titles to refine future prompts.
Tips & Troubleshooting
- Nothing captured on ChatGPT? You need to be signed in and on an actual conversation. The home screen has nothing to extract.
- Old ChatGPT conversations work. Extraction reads the conversation from ChatGPT’s own backend using your existing session, so you can open a chat from last month and capture it. Two things are the exception: turn intent and live merchant offers are only caught while the answer is streaming, so they will not appear on a conversation you reopen later.
- A reopened Gemini chat captures less. Gemini capture listens to live network traffic. Reopening an old conversation produces no traffic to listen to, so it falls back to reading the page, which finds fewer queries. Capture a Gemini conversation while you are having it.
- Google AI Overviews only appear on certain queries. Try informational searches, and check you are on one of the supported Google domains listed above.
- ChatGPT CSP console errors are normal. They come from ChatGPT’s own page, not from the extension.
- Data is cached per conversation. Click Refresh to pull a fresh copy after the answer changes.
- Your captures stay on your machine. Extraction talks to ChatGPT, Gemini, and Google directly. Nothing from your conversations is sent to a Sprout server. See Data and Privacy.
Frequently Asked Questions
AI Insights: Common Questions
The Brands card shows no mentions. What’s wrong?
Nothing is scored against your brand list until you press Reprocess all conversations in the Bulk Analysis dashboard.
Add your brands and competitors first, then Reprocess. It re-scans everything you have already captured against the current list, so you do not need to re-run any prompts. If counts still look low, add more aliases, the misspellings, the legal entity name, the bare domain, and Reprocess again.
Why did my bulk run stop part-way through?
It almost certainly auto-paused, which is deliberate. A run pauses after 3 consecutive hard failures, or after 5 consecutive answers that never got persisted. Both usually mean ChatGPT rate-limited you or your session expired. Sign back in and press Resume rather than restarting the batch.
Items marked Timed out exceeded the 180-second per-prompt limit. Use Retry unfinished to re-queue just those.
Keep first batches small. Twenty prompts will tell you whether your list is well-formed before you commit to two hundred.
What is a sourcing pipeline, and why does it matter?
It is the retrieval path ChatGPT used to fetch a given source. ChatGPT records one of four and the extension reads the label directly: Bright Data and Oxylabs are commercial web scrapers, Labrador is an allowlist of licensed publishers such as Reuters, WSJ, Wikipedia and arXiv, and Open web / SERP is the search baseline.
It matters because it tells you which lever would actually move you. If answers in your category come mostly through the scrapers, being fetchable and answering the question in the page body is the work. If they come mostly through Labrador, you will not optimise your way in and the lever is coverage by those publishers instead.
Can I capture an old ChatGPT conversation, or only a live one?
Old conversations work. Extraction reads the conversation from ChatGPT’s own backend using the session you are already signed in with, so you can open a chat from last month and capture it in full.
Two things are the exception. Turn intent and live merchant offers are only caught while the answer is streaming, so they will not appear on a conversation you reopen later.
Gemini is different: its capture listens to live network traffic. Reopening an old Gemini conversation produces no traffic to listen to, so it falls back to reading the page and finds fewer queries. Capture Gemini conversations while you are having them.
Which Google domains does AI Overview capture work on?
Currently 17: google.com, .co.uk, .nl, .de, .fr, .es, .it, .ca, .com.au, .co.jp, .com.br, .co.in, .be, .pl, .se, .at and .ch.
On any other Google domain the AI Insights tab will not pick up an Overview, which is the most common reason it looks empty on a SERP that clearly has one. This is being widened to every Google country domain in the next release.