Niche research

Turn one focused niche dataset into pain points, objections, the market's tone of voice, volume and engagement, and per-network market share — then keep it live with a listening agent.

A single, well-tuned niche dataset is the richest thing you can pull in buzzabout. For an e-bike DTC brand, one focused collection of electric bike commuting conversation tells you the market's pain points and objections, the tone of voice buyers actually use, where the volume and engagement sit, and how attention is split across social networks — before you spend a cent on ads or product. This page is about what you can extract from a niche, not just how to collect it: a short collect step, then the techniques that turn that dataset into market intelligence.

What you can extract from a niche dataset

Once a niche dataset is collected, it answers questions a survey never could, because it is real unprompted conversation:

Pain points — the recurring frustrations buyers raise (range anxiety, service friction, theft worries) and how often each comes up.
Objections — the reasons people give for not buying, in their own words ("too heavy to carry upstairs", "won't survive a wet commute").
Tone of voice — the actual vocabulary, slang, and register the market uses, so your copy sounds native rather than corporate.
Dynamics, volume, and engagement — how much conversation exists, whether it is growing, and the average engagement rate per post (a proxy for what the market cares about).
Market share across networks — which social networks own the conversation, and which trending topics dominate on each.

The rest of this tutorial collects a tight niche dataset, then shows the extraction techniques on top of it.

Prerequisites

An account that can run research, with credits available. Research depth options are gated by plan: Starter is single-source (you pick any one of Reddit, X, Instagram, TikTok, YouTube, or LinkedIn), Pro and Business are multi-source up to 1,000 mentions, and Enterprise goes up to 10,000.
A niche worth tuning a boolean query for. A meaningful read on pain points and market share needs roughly 100+ mentions — for a thin niche, widen the date range or broaden the query rather than reading too much into a handful of posts.

How to collect a tight niche dataset

The quality of everything you extract depends on the dataset matching your audience. A generic ebike search drowns you in mountain bikes, e-scooters, and regulation debates. Tune the boolean query until the conversation is actually your market.

Search a topic or paste a URL

On the home page you'll see Discover what people buzz about with a Search a topic or paste a URL box. Type a starting topic for your niche — for the e-bike brand, something like electric bike commuting. Or pick a starter prompt from a category tab (Trends, Audience, Content ideas, Market needs, Competitors, Briefs) using Show suggestions. Optionally set source, time, language, and country filters in the chip row, then send.

Keyword query

In the chat the assistant returns a research preview card titled Keyword query. It shows your topic as editable AND / OR / NOT chips, an estimated mention count, and a volume badge — Too narrow, Sweet spot, or Too broad. For a broad ebike query you'll often see Too broad, with the note High volume. It's recommended to narrow the intent to avoid generic results.

Tune the chips to Sweet spot

This is the core of niche research. The operators available in the chip editor are AND, OR, NOT plus parentheses for grouping and "double quotes" for exact phrases. The default join between two keywords is OR (broadens reach) — add AND to require co-occurrence, NOT to exclude. Operators are case-insensitive but render uppercase.

A reliable way to find a niche-specific Sweet spot is to combine a top brand or category keyword with an intent keyword:

Brand + pain point: "electric bike" AND (range OR battery OR theft) — surfaces the niche's frustrations.
Brand + feature: "electric bike" AND (torque OR "throttle" OR "pannier") — surfaces what buyers actually evaluate.
Brand + expressive language: "electric bike" AND (love OR hate OR "wish it" OR annoying) — surfaces tone of voice and strong opinions.
Competitor set (implicit OR): rivian polestar lucid style — "rad power" "aventon" "lectric" returns any post mentioning any of the three.

Narrow with AND / NOT (NOT scooter, NOT "mountain bike"), broaden with OR-synonyms (ebike OR "electric bike" OR "e-bike"). Nested groups are supported. Each edit re-runs the estimate — you'll briefly see Calculating… — and the badge updates. Aim for Sweet spot (the green badge). Sanity-check complex queries against the live count; if you over-narrow you'll see Too narrow or Insufficient volume with Low volume. It's recommended to broaden the search query or widen the date range.

Research depth

Pick how many mentions to collect under Research depth (How many mentions to collect when you run research.). The options are Overview (100 mentions), Landscape screening (200 mentions), Qualitative research (500 mentions), and Quantitative analysis (1,000 mentions). The credits line updates as you change depth — collection costs 1 credit per mention, so 500 mentions is 500 credits. Options above your plan are disabled. For a confident read on pain points and per-network market share, collect at least Landscape screening (200 mentions).

Preview mentions, then Start research

Click Preview mentions to inspect sample posts for free and sanity-check that the query is on-topic for the niche. Previews and estimates cost nothing. When the sample looks right, click Start research to collect the dataset — this sends the message Collect this dataset. and spawns the run. This step is what charges credits.

Now analyse it — extract the market intelligence

Collection just gives you the raw conversation. The extraction happens in chat: ask the assistant questions against the dataset, or apply a skill (open the book icon in the chat input → Browse all skills → open a skill → Use this skill). Each skill returns a hero chart, a detail table, and a few takeaways. Here is what to pull from a niche dataset and how:

Pain points and objections

Ask the assistant to find the recurring themes — it runs pattern detection, showing a Pattern detection block that reads Discovering recurring themes then Pattern detection complete, grouping mentions into named categories with the share of posts in each. For the e-bike niche you might see clusters like range anxiety, commuting-cost savings, and service or repair friction. To go further, ask it to separate genuine objections (reasons not to buy) from neutral discussion, and to quote the strongest examples in the buyers' own words.

Tone of voice

Ask the assistant to describe how this market talks: the vocabulary, slang, register, and the emotional language buyers reach for. Pulling the actual phrasing — not your internal jargon — is what makes downstream copy sound native to the niche. The expressive-keyword query you tuned earlier (AND (love OR hate OR "wish it" …)) makes this read sharper.

Volume, dynamics, and engagement

Ask for the volume and engagement profile of the niche: how many mentions, whether the conversation is trending up or down over the date range, and the average engagement rate per post. Engagement is a proxy for what the market actually cares about — high-engagement themes are the ones worth building product and content around.

Ask the assistant for share of voice by platform — the portion of the niche's total mention volume that sits on each network. It renders this as a horizontal bar comparison so you can see, at a glance, whether the conversation lives on Reddit, YouTube, or TikTok. Follow up by asking it to compare the trending topics per network: the e-bike commuting conversation on Reddit (long troubleshooting threads) reads very differently from TikTok (range tests, theft footage), and that tells you where each message belongs.

Share of voice isn't a button or a standalone page — you ask the assistant for it in chat, and it builds the comparison from the dataset's mention volume. The "Mentioned brands" skill produces a per-brand share-of-voice bar coloured by sentiment if you want to compare brands inside the niche instead of platforms.

Keep the niche live with a listening agent

A one-off niche dataset is a snapshot. Attach a listening agent to it and the same query re-runs on a schedule, so new pain points, product launches, and shifts in tone keep landing without you re-collecting by hand. From the sidebar choose New listening agent, reuse the boolean query you tuned above, and pick a cadence — it will surface new mentions as they appear and can push them to Slack or a webhook. This turns niche research from a project into a standing radar on your market.

Gotchas

Breadth is the boolean query, not a toggle. There is no control labelled broad or specific. You widen or narrow the niche by editing the boolean query (AND / OR / NOT, nested groups, "quoted phrases") and watching the volume badge. A space between two keywords means OR, not AND — add AND explicitly when you need co-occurrence.
Trust the volume badge. Insufficient volume (red) means too few posts for a reliable report; Too narrow (red) and Too broad (orange) bracket the usable range; Sweet spot (green) is the target. As a rough fallback, under 300 reads as Too narrow and over 30,000 as Too broad. For dependable pain-point and market-share reads, aim for ~100+ collected mentions.
Preview is free; collection costs credits. Preview mentions and the live estimates cost nothing. Start research collects at 1 credit per mention, so depth 100 / 200 / 500 / 1,000 costs that many credits.
Depth and multi-source are plan-gated. Over-plan depths are disabled with an upgrade prompt. Starter is single-source, so cross-network market-share comparison needs Pro or above.
Share of voice is conversational, not a feature. There is no dedicated Share of Voice page, button, or chart type. You ask the assistant, which builds a per-platform or per-brand mention-volume comparison. Don't expect a standing roster card.
Pattern detection is conversational and asynchronous. There is no Detect patterns button — you ask the assistant, which calls its pattern-detection tool, polls for a result, and reuses an existing pattern on a dataset rather than re-running it. The named categories and their per-cluster share are real; framing one as a "content gap" is your editorial reading, not a built-in metric.