Automate RSS News Monitor with AI — Smart Alerts by Topic & Urgency in n8n — Step by Step
Your competitors announced a pricing change at 9:14 AM. A regulator published new compliance guidance before lunch. A key vendor got acquired. By the time you saw any of it — buried in a Slack thread
Your competitors announced a pricing change at 9:14 AM. A regulator published new compliance guidance before lunch. A key vendor got acquired. By the time you saw any of it — buried in a Slack thread or a forwarded email three days later — the window to react had closed. This is the quiet tax of information latency, and for technical founders and ops teams it compounds fast.
The problem: you're drowning in feeds, starving for signal
RSS is not dead. Every serious source — TechCrunch, Hacker News, industry blogs, regulator bulletins, competitor changelogs, Google News queries — still exposes a feed. The problem was never access. It's triage. A single active industry produces hundreds of headlines per day, and 95% of them are noise for your specific situation. Nobody on your team has time to read them all, so one of two things happens: they stop checking entirely, or they skim and miss the one item that actually mattered.
Traditional RSS readers don't fix this. They aggregate — they don't judge. A folder with 40 feeds and 2,000 unread items is not monitoring; it's a second inbox you'll declare bankruptcy on by Friday. What you actually need is a system that reads everything, understands your context, and only interrupts you when something is genuinely relevant and time-sensitive.
The solution: an AI-filtered monitor that only pings you when it matters
The RSS News Monitor with AI turns the triage problem over to GPT-4o. On a schedule you set, it pulls every configured feed, deduplicates against what it has already seen, and passes each new item through a language model that scores it against your topics and assigns an urgency level. Only items that clear your relevance threshold generate an alert — delivered straight to Telegram, formatted so you can act in seconds.
The shift is from "here is everything, good luck" to "here are the three things you needed to know today, ranked by how fast you should react." Instead of a folder you dread, you get a channel that stays quiet until it shouldn't. GPT-4o handles the nuance a keyword filter can't: it understands that "Series B" from your direct competitor is high urgency, while the same phrase in an unrelated vertical is noise.
Step-by-step setup in n8n
The workflow is built from standard n8n nodes. Here is how the pieces fit together.
1. Schedule Trigger. Start with a Schedule Trigger node set to your cadence — every 30 minutes for fast-moving markets, or hourly if you want to be gentle on API costs. Cron expression */30 * * * * works for the half-hour interval.
2. RSS Read (one per feed, or looped). Add an RSS Read node pointing at your feed URL. To monitor many sources, feed a list of URLs from a Set or Code node into a Split In Batches (Loop Over Items) node, then call RSS Read inside the loop. Google News is a powerful trick here: a query like https://news.google.com/rss/search?q=YOUR+TOPIC gives you a topic feed without needing the publisher to expose one.
3. Deduplication. This is the step most people skip and then regret. Without it you'll re-alert on the same article every run. Use an n8n Data Store (or a small database / Google Sheets node) keyed on the item guid or link. A Code node compares incoming links against stored ones and drops anything already seen. Only genuinely new items proceed.
4. AI classification with GPT-4o. Route each new item into an OpenAI (or AI Agent) node running gpt-4o. The system prompt is where the intelligence lives. Instruct it to return strict JSON — for example:
{"relevant": true/false, "topic": "one of [your topic list]", "urgency": "critical|high|normal", "reason": "one sentence"}
Give it your business context explicitly: "You monitor news for a B2B fintech ops team. Topics we care about: competitor funding, payment regulation, API outages at Stripe/Plaid, security breaches. Mark urgency 'critical' only if we should react within the hour." Set the node's temperature low (0–0.2) so classification stays consistent, and enable JSON response mode so parsing never breaks.
5. Parse and filter. A Code or Edit Fields node parses the model's JSON output. Follow it with an IF or Filter node: continue only when relevant === true. Everything the model rejected dies here silently — that silence is the whole product.
6. Telegram alert. Add a Telegram node with the "Send Message" operation, connected to a bot you create via BotFather. Build the message body from the model's output so it's scannable: urgency emoji, topic tag, headline, one-line reason, and the link. Something like:
🔴 CRITICAL · [Payments] Stripe announces API v2 breaking changes — affects our checkout by Q3. → {{link}}
Route critical and high to your main channel and let normal collect in a digest, or drop it entirely. Turn on disable_web_page_preview if you want tighter messages.
The benefits: quiet channels, faster reactions, zero manual reading
Once this runs, three things change. First, your team stops reading feeds — the monitor does it, tirelessly, every 30 minutes, without fatigue or bias. Second, your reaction time collapses: the gap between "something happened" and "we know about it" goes from days to minutes. Third, and least obvious, the alert channel earns trust. Because it only fires on genuinely relevant items, people actually read it — unlike the muted, ignored channels most alerting turns into.
The cost is trivial. GPT-4o classification on a few hundred headlines a day runs to a handful of dollars a month, and n8n self-hosted is free. Compared to the cost of missing one regulatory deadline or one competitor move, the ROI isn't close.
Common pitfalls and how to avoid them
Skipping deduplication. The number one failure. Test your dedup logic before going live, or your first day will be a flood of repeats that trains everyone to mute the channel. Store links persistently — an in-memory variable resets every execution.
Vague AI instructions. "Is this relevant?" with no context produces garbage. The model needs your specific topics, your industry, and a concrete definition of each urgency level. Spend your time on the system prompt — it's the difference between a useful monitor and a random one. Iterate: review a week of classifications and tighten the prompt where it over- or under-fires.
Non-deterministic output. If you don't force JSON mode and low temperature, the model will occasionally return prose that breaks your parser and silently kills the run. Enable structured output and wrap the parse step in error handling so one malformed response doesn't stop the batch.
Alert overload from a loose threshold. If everything is "critical," nothing is. Start conservative — only truly urgent items ping in real time — and loosen from there. A monitor that pings twice a day and is always right beats one that pings twenty times and is usually noise.
Polling too aggressively. A 5-minute schedule across 40 feeds hammers sources and your API bill for near-zero benefit. Most industries move on hour-scale, not minute-scale. Match your interval to how fast you actually need to react.
Build it once, tune the prompt over a week, and you have a permanent early-warning system that reads the entire internet on your behalf and only speaks when it counts.
Ja construimos isso pra voce
Nao comece do zero. O RSS News Monitor with AI — Smart Alerts by Topic & Urgency e um workflow n8n pronto para instalar — conecta suas ferramentas em minutos, sem codigo.
Instalar por $199 →
