Build a Social-Listening Agent for Brand Monitoring

A social-listening agent — a scoped, always-on system that watches the open web for your brand name, scores each mention for sentiment, and routes the ones that matter to the right person — is one of the most practical agentic builds a marketing team can own in 2026. It is also one of the most misrepresented: most tutorials pretend every platform is one friendly API away, and it simply is not.

The 2026 reality is a split map. X killed cheap firehose access when it moved the X API to pay-per-use pricing in February 2026. Reddit’s Data API is free only for non-commercial use, and a commercial agreement is not self-serve. TikTok’s only keyword-search API is closed to commercial users entirely. Anyone selling you a weekend build that covers “all social media” is skipping those pages of the documentation.

This guide takes the opposite approach. We map every source honestly — what is accessible, what it costs, and what each API explicitly cannot do, with the platform’s own documentation cited for every constraint. Then we assemble the pipeline from the sources that genuinely work: owned-account APIs, YouTube, RSS, review sites, and Google Alerts, with LLM sentiment scoring, tiered alert routing, and a weekly digest on top.

First, scope. This is a build for monitoring your own brand — mentions, reviews, complaints, praise — across social and review surfaces. It is the companion piece to our competitor-monitoring agent build, and the two are deliberately different machines. The competitor agent watches other companies’ websites, pricing pages, and ad-transparency libraries through MCP tooling; its alerts fire on competitor moves. This agent watches conversations about you, scores sentiment, and fires on reputation signals — a different source map, different triggers, same philosophy: own the system instead of renting it.

The rental alternative is the enterprise social-listening suite — Brandwatch, Sprinklr, and their peers. Those platforms are genuinely broader: they license data access an independent build cannot match. But neither publishes self-serve pricing for full deployments — both run custom-quote sales cycles, and Sprinklr announced it would retire its self-serve tier in April 2026, removing the one transparent low-end price point it had. For a small team that mostly needs to know when someone talks about us, and whether it’s good or bad, a scoped agent you own typically runs at a fraction of enterprise-tool cost — without the procurement cycle.

The prize is real. With 5.24 billion people on social media in 2026 — 64.8% of the global population — the conversation about your brand is happening whether you listen or not. And AI-powered listening is now table stakes across the analytics tooling landscape; the question is no longer whether the capability exists but who owns the pipeline and the data it produces. If your team is still doing this by manually checking apps, our social media service runs exactly this kind of monitoring as operational delivery.

Start with the three platforms most tutorials hand-wave, because they define what your agent realistically covers.

X (Twitter). On February 6, 2026, X launched pay-per-use API pricing and closed its flat monthly Basic and Pro tiers to new developers; existing subscribers are being migrated to the new model. The current rate card on docs.x.com prices a post read at $0.005 per resource, a user read at $0.010, and a post create at $0.015 per request ($0.200 if the post contains a URL). Reads of your own account’s data — “owned reads” covering your own posts, mentions, and followers — are discounted to $0.001 per resource, and every resource is deduplicated within a rolling 24-hour UTC window, so re-reading the same post in one day is not double-billed.

Run the arithmetic on a hypothetical before you architect around X. An agent pulling 1,000 brand-mention posts a day at $0.005 per read spends $5 a day — roughly $150 a month — before a single user lookup. That is workable for targeted queries on your brand terms; it is ruinous for broad firehose polling. And with X organic engagement averaging just 0.035% per post in our 2026 statistics roundup, broad polling buys you very little signal per dollar anyway. For higher volumes, X’s Enterprise tier exists — but X publishes no price for it. Access runs through a custom-quote interest form, and the specific dollar figures circulating on third-party blogs are unverified estimates we deliberately won’t repeat.

Reddit. The Reddit Data API is free only for non-commercial use, capped at 100 queries per minute per OAuth client id, averaged over a rolling 10-minute window. Unauthenticated traffic does not fall back to a slower lane — it is blocked outright, which closes the “just scrape the JSON endpoints” loophole older tutorials still recommend. Commercial or production use requires contacting Reddit through a support-ticket process for an approved agreement, and Reddit’s own Data API Wiki publishes no per-call commercial price. The per-call and per-year figures you may have seen on trade blogs are not confirmed against any Reddit-published price sheet, so treat them as noise: the operative fact is that commercial Reddit access is approval-gated, not self-serve.

TikTok. This is the distinction most content in this space blurs, and it is the crux of the honesty argument. TikTok’s Research API — the only TikTok surface that can query videos and comments by keyword or hashtag — is restricted to qualifying academic institutions and non-profit research bodies, with roughly four-week approval times and a 1,000-requests-per-day quota for those who get in. TikTok’s Display API, the one commercial apps can use, is built — per TikTok’s own product description — for re-displaying a user’s own authorized content in third-party apps; it is understood to offer no keyword or hashtag search for arbitrary brand mentions. Two different products; only one searches, and that one excludes you.

The matrix below is the planning artifact we wish existed when we scoped our first listening build: each candidate source mapped to its commercial eligibility, cost model, quota ceiling, and — the column most content omits — what it explicitly cannot do. Every cell traces to the platform’s own developer documentation, retrieved July 2026.

Read as a whole, the matrix carries the strategic point: the two platforms with the loudest brand conversations — X and TikTok — are precisely the two a small-team agent cannot cover at meaningful scale in 2026, one because of cost, one because of eligibility. That is not a reason to abandon the build. It is the reason to scope it honestly: cover X with targeted, budget-capped queries and owned reads, treat TikTok as an owned-account and manual-review channel, and let the genuinely open sources carry the automated volume.

Now the constructive half. Four source families are genuinely open to a commercial build in 2026, and together they cover more of the brand conversation than you might expect — especially for businesses whose customers talk in reviews, comments, and news coverage rather than viral posts.

Two Meta-specific constraints to design around. First, the Instagram Hashtag Search API — the closest thing Meta offers to hashtag-based brand tracking — caps queries at 30 unique hashtags per IG user per rolling 7-day period, requires Instagram Public Content Access approval, and the search call itself returns only a hashtag ID; pulling actual posts requires the separate recent_media and top_media edges. Second, the mentioned_media endpoint takes a specific media ID as input — in practice that ID arrives via a webhook notification, so this is a push-driven channel, not a retrospective “search everywhere we were mentioned” query. And per Meta’s own reference: “Mentions on Stories are not supported.” Design for webhooks-first and the Meta side of the agent stays comfortably inside its limits.

YouTube’s constraint is structural, not just quota. There is no global comment-search endpoint — commentThreads.list returns comments for one video or channel at a time, so no API call can search comment text for your brand across all of YouTube. The workable pattern is the two-step: use search.list (which now bills to its own dedicated bucket of roughly 100 calls per day) to find candidate videos by title and description keywords, then sweep comments per video at 1 unit per call within the 10,000-unit daily pool. Quota increases require a manual Google review with no self-service purchase path — and developer guides consistently report that bulk-analytics use cases are often rejected, so architect to live inside the default.

Google Alerts rounds out the free tier — still functional in 2026 as a zero-cost feed for indexed web and news mentions. Its two structural gaps: it does not meaningfully surface X, Reddit, Instagram, or Facebook content, and alert emails are commonly reported to lag 24–48 hours behind indexing. Use it as the safety net that catches what your primary feeds miss, never as the real-time channel.

Once mentions flow in, the agent’s core judgment call is sentiment — and 2026 gives you two genuinely different tools for it. Peer-reviewed comparative work published on IEEE Xplore found that LLMs can surpass traditional machine-learning and transfer-learning baselines on sentiment classification of social media reviews, notably in zero-shot settings — no labeled training data required. For a small team, that is the headline: an LLM prompt with your brand context handles sarcasm, mixed sentiment, and aspect-level nuance (“love the product, support was useless”) that a bag-of-words classifier fumbles, on day one.

The counterpoint comes from production vendors operating at firehose scale, where latency and unit cost dominate. Meltwater, for instance, states that its production sentiment pipeline processes 450 million documents a day at roughly 20 milliseconds per document with 83% accuracy on English — vendor-stated figures from its own engineering blog, not independently audited, but a useful signal of where classic lightweight models still win. Your agent is not processing 450 million documents. At hundreds or a few thousand mentions a day, per-mention LLM scoring costs cents and the latency is irrelevant — which flips the trade-off the vendors face.

The architecture is a five-stage pipeline, and none of it requires exotic infrastructure — a scheduled worker, a database, an LLM API key, and a Slack webhook cover it. Ingest pulls from each source on its natural rhythm: Meta webhooks push in real time; YouTube, RSS, and review feeds poll on a 15-to-60-minute cycle (our recommended default, not a platform mandate); Google Alerts arrives by email or feed. Normalize maps every mention to one schema — source, URL, author, text, timestamp — and deduplicates, since the same news story will arrive via RSS and Alerts both. Score runs the hybrid sentiment lane from the previous section, emitting sentiment, severity, topic, and a suggested action as structured output. Route applies the alert tiers below. Digest rolls everything into the weekly email.

The routing tiers are where an owned agent quietly beats a rented dashboard: alerts land where your team already works, thresholds are yours to tune, and every mention is a row in your own database — queryable, joinable, permanent. That last property matters most when a mention is not just a reputation signal but a lead or a support case. A prospect asking “has anyone used these guys?” in a comment thread deserves a CRM record, not just a Slack ping — the same loop-closing logic we walked through in connecting social signals to your CRM. Wire the route stage to create CRM records for lead-shaped and complaint-shaped mentions, and the listening agent becomes a pipeline source, not just a smoke detector — the pattern our CRM automation engagements build day in, day out.

A listening agent is a long-running system touching other platforms’ data under their terms — the boring disciplines are load-bearing. Four rules keep the build durable.

Retention is a feature, not an afterthought. Reddit’s Data API terms are the sharpest example: stored user content must be deleted promptly after it is deleted upstream — Reddit’s own guidance recommends routine deletion within 48 hours. An agent that caches mentions for a weekly digest therefore needs a purge job that re-checks upstream state, not just a fetch loop. Build the deletion path on day one; retrofitting it after a terms-compliance question is far more expensive.

Budget-cap the metered sources. X’s pay-per-use model has no monthly ceiling unless you build one. Hard-cap daily read counts in the agent itself, lean on the $0.001 owned-reads rate for your own mentions and replies, and exploit the 24-hour deduplication window by consolidating queries instead of re-polling. On YouTube, treat the ~100 daily searches as the scarce resource: rotate query terms across days rather than burning the bucket every morning.

Label your data quality. Some feeds are real-time (webhooks), some are near-real-time (polled APIs), and some are slow (Google Alerts, commonly a day or two behind). Surface the source and its latency class on every alert so nobody treats a two-day-old indexed mention as a breaking crisis — or dismisses a webhook-fresh complaint as old news.

Design for source churn. The clear trajectory of 2023–2026 is enclosure: platforms discovered their conversational data is AI training gold, and free access has ratcheted toward metered, gated, or closed at each terms revision. We expect more of that, not less. The architecture hedge is to keep every source behind a thin adapter interface so any single feed can be added, re-priced, or dropped without touching the scoring and routing core. The agent you own outlives any one platform’s pricing mood — that portability is precisely what the enterprise-suite subscription never gives you.