Your Keyword List Is Why Social Listening Failed — A Pruning Guide

type

Post

status

Published

date

Oct 3, 2026 21:00

summary

Most monitoring programs die from noise, not from missing data. A practical method for auditing, qualifying, and cutting keywords, with a false-positive taxonomy and a monthly pruning cadence.

slug

keyword-pruning-social-listening

Why lists rot

A monitoring keyword list degrades for reasons that have nothing to do with your setup quality.

Terms drift in meaning. A category term that was specific in January becomes generic by June as more vendors adopt it. Your match volume triples and relevance halves.

Your product changes. Keywords written for the old positioning keep firing. If you narrowed your scope — as we did when SignalMelo went from four channels to two — every keyword tied to the dropped scope is now pure noise.

Ambient volume rises. Reddit's daily active uniques reached 109 million in Q4 2025, up 39% year over year (Reddit Q4 2025 earnings). The same keyword catches more posts each quarter without you touching it.

Nobody owns the list. It was set up once during onboarding by someone who may not be on the team anymore.

The result is predictable: a list that was 70% useful at setup is 30% useful a year later, and the team's read rate tracks that decline exactly.

The false-positive taxonomy

Before you can prune, you need to name what's failing. Every irrelevant match falls into one of five buckets, and each has a different fix — which is why "just delete the bad keywords" is bad advice.

1. Homonym noise

Your brand or category term is also a common word. Mention, Signal, Melo, Pulse, Amplitude — any of these match thousands of unrelated posts.

Fix: qualify, don't delete. Require a second term in proximity, or exclude the contexts that dominate the noise.

2. Wrong-audience noise

The keyword is correct but the community isn't. "Best monitoring tool" in a sysadmin subreddit means server monitoring.

Fix: scope the keyword to communities rather than weakening the term. The keyword is fine; its range isn't.

3. News and aggregator noise

Your competitor's funding round generates hundreds of matches across news aggregators and bot accounts. High volume, zero intent.

Fix: exclude by account type and post format where your tool supports it, and treat competitor brand names as requiring a co-occurring intent word.

4. Stale-thread noise

The match is relevant but the thread is four months old and dormant. Replying achieves nothing.

Fix: this is a filter problem, not a keyword problem. Set a recency cutoff — for most categories, 72 hours for active reply, with a separate monthly pass for high-ranking evergreen threads.

5. Self-noise

Your own posts, your team's replies, your own press coverage syndicated across sites.

Fix: exclude your own domain and team accounts explicitly. This one is trivially fixable and shockingly common.

The audit

Run this quarterly. It takes about an hour.

Step 1 — Sample, don't scan. Pull 100 random matches from the last 30 days. Not the ones you replied to — random, including the ones nobody opened.

Step 2 — Tag each one. Relevant / irrelevant. For every irrelevant match, tag which of the five buckets it belongs to, and which keyword produced it.

Step 3 — Score each keyword.

Step 4 — Sort into three actions.

Precision	Action
Above 60%	Keep unchanged
25–60%	Qualify — add a required co-term or an exclusion
Below 25%	Cut, unless it's your brand name

Step 5 — Check the other direction. Look at the threads you did reply to this quarter. Which keyword surfaced each one? Keywords that surfaced zero replied-to threads in three months are carrying no weight regardless of their precision score.

That last step matters as much as the first four. A keyword can be 90% precise and still useless if it only ever matches threads nobody acts on.

What a pruned list looks like

Teams typically start with 40–80 keywords. Pruned lists tend to settle between 15 and 25.

The structure that survives:

Bucket	Count	Precision expectation
Brand + variants + common misspellings	3–5	Low precision acceptable, never cut
Competitor + intent qualifier	4–8	High — the qualifier does the work
Category terms, qualified	3–5	Medium
Request phrasings + category	3–5	Medium, high value
Symptom vocabulary from customers	2–4	Variable, worth the noise

Brand keywords are the exception to every precision rule. You keep them even at 10% precision, because the cost of missing a complaint about you outweighs the cost of scanning noise. Everything else earns its place.

The unbranded buckets in that table are the higher-value half of the list — see the threads that never mention your brand for how to build them.

Pruning as a habit, not a project

An audit you run once produces a list that rots again. Three things make it stick:

A one-click "not relevant" action. Whatever your tool, the person reviewing the queue needs a way to mark noise in the moment. Retrospective auditing catches less than in-flight tagging, because the reviewer already knows why a match was bad and won't remember in six weeks.

A named owner. One person owns the keyword list. Not the team, not the tool — a person, reviewed quarterly.

A change log. Record every keyword added or cut, with the date and reason. Without it you'll re-add a term you cut last quarter and rediscover the same noise.

If you're still evaluating tooling, the ability to qualify and exclude — not just match — is the feature that determines whether pruning is possible at all. Some free monitoring options only support exact keyword matching, which caps how precise your list can ever get.

Frequently asked questions

How many keywords should I monitor? For a focused product, 15–25 after pruning. The number matters less than the precision: 12 well-qualified keywords outperform 60 raw ones, because the queue actually gets read.

Should I ever monitor a broad term with low precision? Only your brand name, and only because missing a complaint is expensive. For everything else, low precision means the term isn't doing work — qualify it or cut it.

How do I qualify a keyword without losing real matches? Add a required co-occurring term rather than an exclusion where possible. Requiring "monitoring" near "signal" loses less than excluding every context you can think of, because exclusion lists are always incomplete.

What if my tool doesn't support boolean operators? Then your practical options are narrower keywords and community scoping. It's a real constraint — and it's one of the clearer reasons to move off exact-match-only tooling once volume grows.

How often should I prune? Quarterly for a full audit, continuously for in-flight "not relevant" tagging. Monthly is overkill for most teams; annually is too slow given how fast ambient volume grows.

Does pruning reduce how many opportunities I find? It reduces matches and usually increases replies. The constraint on most teams isn't detection — it's attention. A queue that gets opened produces more replies than a larger queue that doesn't.

The takeaway

Monitoring programs rarely fail loudly. They fail by slowly becoming unreadable, until checking the queue stops feeling worth the time.

The fix isn't better detection. It's the discipline to keep cutting — which is unglamorous, takes an hour a quarter, and is the difference between a tool your team uses and a subscription nobody cancels because nobody remembers it's there.

Worth asking at your next review: when was the last time anyone deleted a keyword? If the answer is "never," that's your diagnosis.

SignalMelo supports qualified keywords and exclusions across Reddit and X monitoring, with per-match relevance feedback so the list improves as your team works the queue.