Social Listening Keyword Strategy: Queries That Catch the Right Conversations

The query is the product

Every social listening deployment is judged on its dashboards, and every one of those dashboards is only as good as the query behind it. A badly built query produces a chart with a confident line on it, drawn from data that is partly irrelevant and missing the conversations that mattered.

This is the least glamorous part of the discipline and the one that determines everything downstream. Sentiment analysis of the wrong mentions is precise nonsense. A share-of-voice figure built on an asymmetric query — thorough for your brand, casual for competitors — is worse than no figure at all.

What follows is a framework for building queries that hold up: what to include, what to exclude, how to test whether it works, and how to maintain it as the business and the language around it change.

The four layers of a query set

A working query set is not one search. It is several, each answering a different question, and mixing them into one is the most common structural mistake.

1. 1

   Layer 1: your brand, exhaustively

   Every way your name is written — official, short form, with and without diacritics, transliterated, common misspellings, the domain, product names that stand alone, the legal entity, and any former name after a rebrand. This layer should be over-inclusive and then narrowed by exclusions.

2. 2

   Layer 2: competitors, symmetrically

   The same treatment for each competitor you intend to compare against. Asymmetry here is what makes share-of-voice numbers wrong: a thorough brand query and a casual competitor query guarantees a flattering result that means nothing.

3. 3

   Layer 3: category and topic

   The language people use to discuss the problem you solve, without naming anyone. This is where prospective customers are found and where trends appear first, and it is the noisiest layer by a wide margin.

4. 4

   Layer 4: campaign and event

   Time-bounded queries for a launch, a campaign, a sponsorship or an incident. These have a start and an end date and should be retired rather than left running.

Building the brand layer

Start here, because it is the one you can verify — you know when a mention is genuinely about you.

• Write every spelling variant before touching the tool. Fifteen to twenty is normal for a brand of any age.
• Include diacritic-free forms. In markets using Azerbaijani, Turkish or similar orthographies, people routinely type without them and a query that requires them misses a large share.
• Include transliterations if any part of your audience writes in another script.
• Include plausible misspellings — doubled letters, dropped letters, joined and separated forms, keyboard-adjacent substitutions.
• Include the domain, which people type when recommending rather than tagging.
• Include product names that circulate independently of the brand.
• Include the previous name if you rebranded. It persists for years.
• Exclude aggressively: identically named companies in other sectors, place names, ordinary words, and any usage unrelated to you.
• Where the brand name is a common word, the exclusion list becomes the main body of work and the query is unusable without it.

The AZ-language treatment of this problem, including the orthographic variants in more detail, is covered in finding untagged mentions.

Building the category layer

The noisiest and most valuable layer. It needs more structure than the others.

1. 1

   Start from how customers describe the problem, not the solution

   People do not search or post using your product category name. They describe a situation. Collect this language from support tickets, sales calls and existing mentions rather than inventing it.

2. 2

   Require co-occurrence rather than single terms

   A single broad term returns everything. Requiring two related terms in the same post — the problem and a signal of intent, or the category and a location — narrows dramatically while keeping the relevant results.

3. 3

   Add intent markers where you want prospects

   Question forms, recommendation requests, comparison language. These separate people considering a purchase from people discussing the topic generally.

4. 4

   Exclude the obvious noise sources early

   Job postings, news aggregation, spam patterns, unrelated industries using the same words. These are identifiable within a day of running the query.

5. 5

   Bound it by language and region if that matters

   A category term may be common in one market and mean something else in another.

6. 6

   Accept that this layer will never be clean

   Category listening trades precision for discovery. The goal is a signal-to-noise ratio a person will keep reading, not perfection.

If the category layer is producing volumes nobody reviews, it is not working regardless of what the dashboard shows. Narrow it until someone reads it.

Testing whether the query works

Queries are testable, and most are deployed without being tested at all.

1. 1

   Measure precision by reading

   Take a random sample of a hundred results and count how many are genuinely relevant. This number is your precision, it takes an hour, and it is the single most informative thing you can do.

2. 2

   Measure recall by searching manually

   Search for things you know exist — a mention you saw, a review you know about, a discussion someone reported — and check whether the query caught them. Anything it missed tells you what to add.

3. 3

   Ask the team what they have seen

   Sales, support and staff encounter mentions the query misses. Asking them quarterly is a cheap recall test and it consistently finds gaps.

4. 4

   Re-test after any change

   To the query, to your product names, or after a campaign introduces new language.

5. 5

   Track precision over time

   It degrades as the world changes around a static query — a new company with a similar name, a term that acquires a new meaning.

Maintaining the query set

Queries decay. Treating them as a one-time configuration is why listening programmes quietly stop being accurate.

• Version the queries with dates and reasons for each change. Without this, a shift in volume cannot be distinguished from a query edit.
• Never change a query mid-analysis without noting it. A trend line that spans a query change is not a trend line.
• Review quarterly: new products, new campaign language, new competitors, new misspellings, terms that have acquired other meanings.
• Retire campaign queries on schedule rather than leaving them accumulating.
• Keep the exclusion list growing. It is the part that does most of the work and it is never finished.
• Document who owns the query set. Unowned queries are unmaintained queries.
• Keep a note of what you deliberately chose not to monitor, so the decision is not silently reversed later.

The versioning discipline matters more than it appears. Most listening programmes cannot explain their own historical charts because nobody recorded when the query changed.

Data and tooling requirements

What the query layer needs from the platform underneath it.

• Boolean logic with grouping, negation and proximity — the minimum needed to express co-occurrence and exclusions.
• Language and region filtering.
• The ability to read raw results, not only aggregates. A platform that shows charts and hides posts cannot be validated.
• Historical backfill where available, so a new query can be tested against the past rather than only forward.
• Query versioning, or a disciplined external record if the tool lacks it.
• Separate reporting per query layer.
• Multilingual handling where your audience writes in several languages — the same term set rarely works across them.
• Automatic grouping at volume, since the category layer produces more than a person can read — which a consolidated monitoring platform handles as part of the workflow.

The ability to read raw results is non-negotiable. Our automation services page covers building the collection and triage layer where a platform does not provide what is needed.

What a query cannot fix

Some limits are structural and no amount of query craft addresses them.

• Anything not publicly accessible — private groups, direct messages, content behind logins.
• Coverage that varies by platform and changes over time, which makes historical comparison less reliable than it looks.
• Mentions in images and video without text.
• Conversations that happen offline, which for many businesses is the majority.
• The systematic bias of who posts: people with strong opinions, not a representative sample of customers.
• Sarcasm, context and meaning — the query decides what is collected, not what it means.

State these when presenting findings. An analysis that is explicit about coverage is trusted where it is valid; one that implies completeness gets discounted entirely when a gap is discovered.

Decision framework and next step

Four questions before building a query set.

1. 1

   Have you written every spelling variant of your brand?

   This is the first hour of work and it is usually longer than expected.

2. 2

   Is your brand name also an ordinary word?

   If so, the exclusion list is the project and the query will not work without it.

3. 3

   Are competitor queries as thorough as your own?

   If not, do not publish share of voice.

4. 4

   Who will read a hundred raw results, and when?

   Schedule it. An unvalidated query produces charts nobody should rely on.

Build the brand layer first and validate it by reading. Add competitors symmetrically only if you intend to compare. Add the category layer last and narrow it until someone reads it consistently. Keep the four layers reported separately. The distinction between the queue and the analysis is covered in monitoring versus listening; our AI solutions overview covers staging.

Frequently asked questions

1. 1

   What should a social listening keyword strategy contain?

   Four separate layers: your brand with every spelling variant and a substantial exclusion list; competitors treated symmetrically; category and topic language drawn from how customers describe the problem; and time-bounded campaign queries that are retired on schedule.

2. 2

   How do you know whether a query works?

   Read a random sample of a hundred results and count how many are genuinely relevant. Then test recall by searching for mentions you know exist and checking whether the query caught them. Both take about an hour.

3. 3

   Why keep the layers separate?

   Because merging them produces a volume number that rises when the category gets busier and is read as brand growth. Each layer answers a different question and needs its own reporting.

4. 4

   What is the most common mistake?

   An asymmetric competitor query — thorough for your own brand, casual for everyone else — which guarantees a flattering share-of-voice figure that means nothing.

5. 5

   How often should queries be reviewed?

   Quarterly, for new products, campaign language, competitors, misspellings and terms that have acquired new meanings. Version every change with a date and a reason, or historical charts become uninterpretable.

The query is the least visible part of a listening programme and the one that determines whether anything built on top of it is worth reading.