AI answers
Buyer-question research
Choose AI answers in New niche, enter your product and public website, and add up to 20 buyer questions. Choose the available assistants and one to five repeats for a full scan. Quick scans use the first three questions once.
AI and social research share your existing account, niche limits and scan allowance. A cached AI result does not make new provider calls. A scan that cannot start releases its reservation; if provider execution has started, a failed run can still count toward your allowance. The scan history records its status.
The report distinguishes product mentions, own-site citations, other products named and cited versus consulted sources. It records questions, models, date and repeated observations. Missing results are shown as unavailable, never filled in. Compare the same questions and settings before interpreting a change; no scan guarantees recommendations.
Free research is public on Explore, including product, questions and a limited report. Paid research starts private and can be published by its owner. Copying public AI research copies the setup; it does not run a new scan or transfer another account’s private data. Do not put confidential or personal data in buyer questions.
Docs
How Outliers works
Baseline. For each account we take the 20 most recent posts that are at least 48 hours old (and under 90 days old), sort them by views, drop the top 2 and bottom 2, and take the median. Fewer than 8 eligible posts marks the baseline low confidence; you still see the outliers, badged.
Outlier. A post published 2–14 days ago whose views are at least 5× its account's baseline. We also flag engagement outliers: (saves + shares) / views at least 3× the account's median rate, the leading indicator for posts still climbing.
Instagram. Baseline = trimmed median of the last 20 posts of the same kind that are at least 48 hours old: views for reels, likes for photos and carousels (Instagram does not show their views). Engagement: (likes + comments) / views for reels. A post is only ever compared with the same kind of post from its own account.
Climbing. Posts under 48 hours old are shown with a provisional multiple but never scored, so a fresh post cannot distort the baseline and a slow starter is not written off.
Cache and budget. Account data is cached for 24 hours. A scan re-fetches only stale accounts; a scan that fetched nothing is free and does not count against the Free plan's 3 scans per month. Free niches hold up to 10 accounts, paid niches up to 30.
Public and private. Every niche on the Free plan is listed on Explore, where anyone can browse its outliers and copy its account set; that is how the catalog grows. Paid plans keep niches private by default and can publish any of them. Dropping back to Free publishes them again.
Video and carousel analysis. After each scan, the best new outliers (5 on Free, 15 on Creator, 30 on Pro) are analyzed from the actual video and audio: the hook as spoken or shown, every shot, the script, on-screen text, editing, products, and a recreation blueprint a video pipeline can consume. Photo carousels count toward the same number and are read from their slide images: every slide's role and text as shown, the visual style, and a slide-by-slide blueprint; their sound is not analyzed. Any other post can be analyzed on demand.
Teardown and adapt. A teardown comes from the video or carousel analysis when there is one, otherwise from a small model on the caption, stats and multiple. Adaptations use a larger model with your voice profile and reuse the structure and hook mechanism, never the words; a variant that repeats the source hook is rewritten before you see it.
Niche report. Every niche gets a "what is working" report computed from its analyzed outliers: format, hook technique and mechanism counts, video length, speech vs music, carousel slide counts, topics and products, plus a written summary with hooks to try and next post ideas. Public niches show theirs on Explore.
Set up in Claude or ChatGPT
- Sign in here and set a voice profile in Settings.
- Add the connector: URL
https://mcp.outliers.ninefold.ai/mcp. Claude: Settings → Connectors → Add custom connector. ChatGPT: Settings → Apps & Connectors → Create. - Approve the sign-in when the browser opens. Then ask: “Create a niche called plant care with these handles …”, “scan it”, “show me the outliers”, “tear down the top one”, “adapt it for me”.
Scripts and agents. A client that can't open a browser to sign in can use an API key instead (Creator and Pro). Make one in Settings and send it as Authorization: Bearer <key> to the same URL. A key acts as you until you revoke it.
Tool reference
| tool | kind | what it does |
|---|---|---|
| list_niches | read | Your niches with account counts, last scan time and your scan budget. |
| create_niche | write | Create a niche from 5–30 handles, or from hashtags/keywords alone: discovery auto-picks the best-fitting accounts so the niche is ready to scan. |
| browse_public_niches | read | Curated public niches, most-copied first, with their top outliers. |
| use_public_niche | write | Copy a public niche into your account (free until scanned). |
| publish_niche | write | List one of your niches on Explore with a one-line description, or unlist it (paid plans). Returns the public report URL. |
| approve_accounts | write | Approve proposed accounts for a niche. |
| scan_niche | write, costs a scan | Fetch the latest posts for every approved account (24h cache unless force_refresh), recompute baselines, flag outliers. |
| get_outliers | read | Ranked outliers: multiple, engagement multiple, hook, format, url, plus outlier_id and post_id. |
| teardown | read | Hook, format tag, 3–5 structure bullets, one sentence on why it worked. Cached per post; grounded in the video or slides once the post has been analyzed. |
| analyze_video | write, costs ~$0.02 | Gemini watches the actual video: every shot with timestamps, the spoken script, on-screen text, audio, editing, products shown, tags, and a recreation blueprint with per-shot generation prompts. For a photo carousel it reads the slides instead: every slide's role and text, the visual style, and a slide-by-slide blueprint. Cached per post. |
| analyze_niche_videos | write, costs ~$0.02/post | Batch-analyze the niche's outliers, videos and carousels, that have no analysis yet, best multiple first. |
| niche_report | read | What is working in the niche right now: formats, hook techniques, mechanisms, video length and carousel slide counts, topics, products, plus a written summary, hooks to try and next post ideas. Cached 24h. |
| public_niche_market_report | read | The latest market report of a public niche, without copying it. |
| adapt | write | Three script variants (hook, beats, CTA) in your voice from your voice profile. Never copies the source. |
| save_outlier | write | Keep an outlier in the saved tab. |
| dismiss_outlier | write | Hide an outlier from the feed (reversible). |
| weekly_digest | read | The same digest the Monday email sends: top outliers of the last 7 days per niche. |
No tool deletes anything. Errors tell you what to do next, including how to upgrade when a Free-plan limit is hit.
Privacy, terms, support
We fetch public TikTok and Instagram data through a third-party scraping API, store the posts we score, and store the teardowns and adaptations you generate. Full details: privacy policy and terms.
Support: support@ninefold.ai. A human answers within 48 hours.