Social listening with AI agents: track mentions, trends and competitors

Social listening with AI agents means an agent fetches the posts, comments and mentions itself at the moment you ask, reads them, and writes the summary. There is no dashboard to open and no crawler running all week against a query you wrote in month one. A small brand can cover X, Instagram, TikTok, Reddit and YouTube this way for about $12.47 a week in tool calls, or roughly $54 a month, well inside the Scale plan's included credit.

What is agent-driven social listening?

It is monitoring where the collection step happens inside the same run as the analysis. The agent decides which sources the question needs, calls one tool per source, merges and dedupes what comes back, then writes a brief with real quotes and links. Keyword alerting tells you that a term appeared. An agent tells you what people meant by it, which five posts are worth answering, and what changed since last week.

The mechanical difference is when the query gets written. A classic listening suite needs the query before the data exists, so it is always answering last month's question. An agent writes the query at ask time, which means a competitor you first heard of on Thursday is in Monday's brief without anyone editing a saved search.

Why do listening dashboards go stale?

Because a dashboard has to commit to what it collects before you know what you will want to ask. The boolean query, the tracked handles and the sentiment model are all set on day one, and the world they were tuned for lasts about a quarter. New slang, a new competitor, a product rename, a meme that spells your brand wrong on purpose: none of it is captured until a person notices and edits the config.

The second failure is quieter. A dashboard collects whether or not anyone reads it, so the bill arrives every month while the tab stays closed. Most small teams open a listening tool hard for two weeks after a launch and then not at all, which is exactly the usage shape that per-seat pricing punishes. Listening is the cleanest case for paying by the pull rather than by the seat.

What does the monitoring loop look like?

Five steps, run on a schedule you pick: watch, pull, cluster, rank, brief. Only step two costs money. Everything after it happens inside the model, on data you have already paid for, which is why adding a sixth question to the brief is free and adding a sixth platform is not.

Step 1. Fix the watchlist

Write down four things and nothing else: your brand and the two spellings people get wrong, three or four competitor handles, two hashtags, and three communities. A watchlist longer than that produces a brief nobody finishes reading. This is the step an agent cannot do for you, because knowing which three competitors matter is a judgement about your market, not a query.

Step 2. Pull one source per call

Each platform gets the cheapest tool that answers the question you are asking of it. For a brand watch that is search on X, a hashtag feed on Instagram, keyword and hashtag search on TikTok, community search on Reddit, and comments under specific videos on YouTube.

The prices above are per unit and billed on what the call actually returns, with a floor of $0.0020 per call and a hard ceiling per call so a bad query cannot run away with your balance. On Twitter search a hundred results costs $0.12. On Instagram hashtag thirty posts costs $0.21. TikTok is priced per request rather than per row, so filling in a keyword search and a hashtag feed in one call is $0.12 whether it returns forty posts or four hundred.

Tool choice inside a platform matters more than platform choice. Reddit API searchreads through Reddit's own endpoints and has no start fee, while the broader Reddit posts tool pays a fixed fee before it returns anything, which on a 25-item daily pull costs more than the results do. Same platform, same answer, different bill. For competitor content, TikTok profile postsis the cheapest per-item tool in the catalog at $0.0009 per result, so a rival's last hundred posts costs nine cents.

Step 3. Cluster before you count

Raw mention volume is the least useful number in listening, because one reposted screenshot can triple it overnight. Have the agent drop reposts and anything under three engagements, then group what is left into three to five themes and report a count per theme. Five clusters with a real quote each is a brief a founder reads. Four hundred rows with a sentiment score is a spreadsheet nobody opens.

Step 4. Rank by who is asking

Sort by reach times recency, then pull out the posts that are actually questions. A three-day-old comment asking whether your product does X is worth more than a viral post that mentions you in passing, because one of them has an obvious next action and the other does not.

Step 5. Write the brief to a file

The agent starts every run with no memory of the last one, so the previous brief has to be an input. Save each week as a file, hand the last one back in the prompt, and the "what changed" line writes itself. Skip this and you have a snapshot tool, not a listening system.

What should a weekly brief track?

Four things, in this order: mentions, hashtags, competitor posts and sentiment themes. Everything else is a metric you will read once. A good brief fits on one screen and every line either names something that changed or something to answer.

• Mentions, split by platform. This week against last week, as a count and a direction. The split matters because a jump on Reddit and a jump on TikTok mean completely different things and call for different replies.
• Hashtags and their traction. Post count on your two tracked tags, plus any tag that shows up in more than five posts and was not on the list. New tags are how you find out what people call your category when you are not in the room.
• Competitor posts that worked.Each rival's single best post of the week with its engagement, and one line on why it landed. Pull their follower count in the same run with Instagram profile or X profile so growth sits next to content instead of in another tab.
• Sentiment as themes, not a score. A single number between minus one and one has never told anyone what to fix. Ask for clusters of complaints and praise, each with a count and a verbatim quote, and every negative one with a link so you can go read the thread yourself.
• The reply list. Up to five posts worth answering, ranked, with links. A human writes and sends every one of them.

A weekly brief prompt you can paste

This is the whole system in one message. Fill in the watchlist, paste it into an agent that has the social tools connected, and run it every Monday. The rules block is doing more work than it looks: the price check stops a wide query from spending, and the verbatim rule stops the model from softening complaints into summary.

It is Monday. Cover the last 7 days only.

WATCHLIST
  brand        <brand name> + the two spellings people get wrong
  competitors  <handle 1>, <handle 2>, <handle 3>, <handle 4>
  hashtags     #<tag 1>, #<tag 2>
  communities  r/<sub 1>, r/<sub 2>, r/<sub 3>

PULLS  (one source per call, cheapest tool that answers it)
  twitter.search     brand terms, 100 results
  instagram.hashtag  each hashtag, 30 results
  tiktok.api         keyword search + hashtag feed
  reddit.api         each community, 25 items
  instagram.posts    each competitor, 20 posts
  tiktok.profile     each competitor, 100 posts

RULES
  Check the price of every call before you run it. Skip anything over $0.50.
  Drop reposts, and drop anything with fewer than 3 engagements.
  Quote people verbatim. Never paraphrase a complaint.

OUTPUT  (400 words, no preamble)
  1. Volume     mentions per platform, this week vs the numbers in last-brief.md
  2. Themes     3 to 5 clusters, each with a count and one real quote
  3. Negatives  every complaint, verbatim, with a link
  4. Rivals     each competitor's best post this week and why it worked
  5. Reply to   up to 5 posts worth answering, with links, ranked
  6. Changed    one line on what moved since last week

Write the result to brief-<today>.md so next week has something to diff.

Two things make this reliable rather than clever. The agent sees the price of a call before it runs it, so a $0.50 ceiling is a real constraint and not a suggestion. And every pull names a specific tool, which stops the agent inventing a search strategy that quietly costs twenty times more than the one you meant. If your agent connects over MCP, it discovers those tools and their prices on connect, so the names in the prompt are the same names it sees.

What does a week of listening cost?

For a small brand watching five platforms and four competitors, about $12.47 a week. That is 69 calls returning roughly 2,793 records, priced per unit with the price shown before each call. The same coverage on a mainstream listening suite is a seat, and entry seats generally start in the low hundreds per month whether you open the tab or not.

The lever with the most travel is cadence, not coverage. Dropping the Reddit watch from daily to three times a week takes $4.73 down to $2.03, which is more saving than deleting a whole platform. The second lever is result limits: most of these tools bill per row returned, so a limit of 30 instead of 100 on a query that only ever surfaces a dozen real mentions changes nothing about the brief.

Worth saying plainly: these are the tool costs, not the total. You still pay your model provider for the tokens spent reading a few thousand posts, which for a weekly brief on a mid-tier model lands in single dollars. Budget around $20 a week all in, and expect the model bill to grow faster than the tool bill if you widen the watchlist.

When do you still need a human?

For every judgement call and every outbound action. An agent is good at fetching, deduping, clustering and summarizing, and unreliable at the four things that actually carry risk: irony, intent, priority and replying. Treat the brief as evidence, never as a decision.

• Sarcasm and in-jokes."Love that it crashes every time I open it" is scored positive by most sentiment models and by plenty of agents reading fast. Community in-jokes look like praise or abuse depending on context the model does not have.
• Bug or preference. Ten people saying a flow is confusing might be one broken screen or ten people who wanted a different product. Only someone who knows the roadmap can tell those apart, and the difference decides a sprint.
• Anything that gets posted. Never let an agent send the reply. Automated posting, liking and DM sending are exactly what platform anti-spam systems are built to catch, and a banned brand account costs more than the listening ever saved. The agent drafts, a person sends.
• A real incident. When something is spreading and wrong, the brief tells you it is happening and a human decides what gets said. That is a legal and reputational call, not a summarization task.

Where Goro fits

You can run this loop with whatever tools you already have. The reason it usually stalls is not the agent, it is that five platforms means five scrapers to maintain or five vendor accounts to hold, and the maintenance outlives the enthusiasm. That is the same problem covered in web scraping for AI agents, applied to one narrow job.

Goro is one connection that exposes all of these as tools your agent can discover: 24 social tools across X, Instagram, TikTok, Reddit, YouTube, Facebook and LinkedIn, each with its price visible before the call and all of them drawing on a single balance. The prompt above runs against it as written, because those slugs are the real tool names. Every price in this article, including the weekly total, comes straight from the tools pages, where the per-unit rate and the per-call ceiling are listed for all 62 tools.