How to Build an AI Agent for Media Monitoring Your Brand Mentions

Blueprint  ·  Monitoring & Intelligence  ·  Difficulty: Intermediate  ·  22 min read

A practical, no-code blueprint for PR, comms, and marketing operators who want an AI agent that watches the news for mentions of their company and executives, filters out the noise, and delivers a clean daily briefing – without paying $15,000 a year for a legacy monitoring suite.

Summary Card

• What it does: Pulls news and web mentions of your company name and a list of executive names, runs them through an AI classifier for relevance and sentiment, de-duplicates near-identical stories, and delivers a daily briefing to Slack, email, or a shared doc
• Who it’s for: PR and comms managers, marketing operators, investor relations leads, founders, and agency account leads who need reliable media coverage tracking without an enterprise tool budget
• Time to implement: 45 minutes for the manual version; 2-3 hours for the fully automated daily briefing
• Tools required: Google News + Bing News + a few RSS feeds (free), ChatGPT or Claude for classification and writing, Make or n8n for automation, Slack or Gmail for delivery
• Cost estimate: $20/month for the manual workflow; $30-60/month for the fully automated version; $0 if you already have an enterprise monitoring tool and just want the AI layer on top
• Difficulty: Intermediate – no code, but you will configure a few automation modules and refine your search queries over 2-3 iterations
• Last tested: April 2026 with Claude Sonnet 4, ChatGPT 5.3, Make, and n8n

What the market looks like in 2026

Legacy media monitoring suites like Meltwater, Cision, and Muck Rack still dominate large-enterprise PR budgets, with pricing that typically starts around $10,000 per year and climbs from there. For smaller comms teams and solo operators, that math has never worked. The gap used to be filled by Google Alerts and a lot of manual scrolling. In 2026, a capable AI agent stitched together from free news APIs, RSS, and a Claude or GPT classifier can cover 80 percent of what those enterprise tools do for coverage tracking – minus the licensed print and paywalled archive access, which matters for some teams and not for others. This blueprint shows you exactly how to build that agent.

The Five-Stage Media Monitoring Workflow

Entity List Template – Company Block

COMPANY: Acme Robotics

Primary name:
  "Acme Robotics"

Name variants (include all spellings and short forms):
  "Acme Robotics, Inc."
  "Acme Robotics Inc"
  "Acme"                  [short form - high false-positive risk]
  "ACME"                  [all caps - trade press often uses this]
  "$ACME"                 [ticker]

Disambiguators (must appear near the name for "Acme" alone to count):
  "robotics" OR "automation" OR "warehouse" OR "Dallas" OR "CEO Sarah Chen"

Exclusions (kill the match if any of these appear):
  "Acme Corporation" (the unrelated holding company in Ohio)
  "Wile E. Coyote"     (cartoon references)
  "Acme Brick"         (different company, Texas)
  "Acme Markets"       (grocery chain)

Products and initiatives to also track:
  "Acme Fleet OS"
  "Acme R-7 robot"
  "Project Beacon"     [internal codename that leaked in a TechCrunch story]

Entity List Template – Executive Block

EXECUTIVES:

- Name:           Sarah Chen
  Title:          CEO, Acme Robotics
  Name variants:  "Sarah Chen", "S. Chen" (in bylines only)
  Disambiguators: must appear with "Acme", "robotics",
                  "CEO", "founder", or "warehouse automation"
  Known aliases:  none
  LinkedIn URL:   linkedin.com/in/sarahchen-acme

- Name:           Marcus Oduya
  Title:          CTO, Acme Robotics
  Name variants:  "Marcus Oduya", "Marcus O."
  Disambiguators: must appear with "Acme", "CTO", "engineering",
                  or "fleet operating system"

- Name:           Priya Vasquez
  Title:          Chief Communications Officer, Acme Robotics
  Name variants:  "Priya Vasquez", "Priya V."
  Disambiguators: must appear with "Acme" or "communications"

OPTIONAL - Board members and investors to monitor:
- Name:           Daniel Park
  Role:           Board member, Horizon Capital
  Disambiguators: must appear with "Acme" or "Horizon Capital"

Why disambiguators matter more than exact-match rules

You might be tempted to just search for the full phrase “Sarah Chen CEO Acme” and call it a day. The problem is that real coverage rarely mentions all three in the same sentence. A reporter might write “Chen, whose company makes warehouse robots, said…” – a match you would miss with a rigid exact-match query. Disambiguators work as a soft filter in the classification step (Step 3), not as a hard search constraint. You cast a wider net during collection, then let the AI verify context.

Recommended Source Stack – Free and Near-Free

Tier 1 – Start here (free, covers 70 percent of English-language news):

• Google News RSS – The most overlooked free source in the entire monitoring space. Google News exposes an RSS feed for any search query. Build one feed per entity using a query like https://news.google.com/rss/search?q=%22Acme+Robotics%22&hl=en-US&gl=US&ceid=US:en. This pulls structured headlines, snippets, timestamps, and source URLs. Free. No API key.
• Bing News Search API – Microsoft’s news search has better coverage of mid-tier trade publications than Google in my testing. Free tier gives you 1,000 calls per month, which is plenty for a daily monitoring workflow. Requires an Azure account.
• GDELT Project – A free global news monitoring database that indexes news in 100+ languages. It is overkill for a small brand but excellent for companies with international exposure. API returns JSON.

Tier 2 – Add for fuller coverage:

• NewsAPI.org – Aggregates 80,000+ news sources. Free tier: 100 requests per day, 24-hour delay on results. Paid tier starts at $449/month and removes the delay. For most monitoring workflows, the 24-hour delay is acceptable.
• RSS feeds from the specific trade publications your industry cares about – Every major trade pub still publishes RSS even if they bury the link. Ten or twelve targeted RSS feeds often outperform a generic news aggregator.
• Reddit via the Reddit API – If your industry has active subreddits, pull mentions from there. Reddit often breaks stories days before trade press catches them.
• YouTube Data API – Free tier. Search for your entity names in video titles and descriptions. Catches podcast episodes and creator coverage.

Tier 3 – Add if social coverage matters:

• X (Twitter) API – The paid Basic tier is $200/month and gives you search access. The free tier no longer supports search. If you need X coverage and do not want to pay, use a tool like Typefully or a third-party monitoring dashboard that includes X.
• LinkedIn – No public API for search. The workaround is Google searching with a site:linkedin.com filter, which picks up public posts. Imperfect but better than nothing.

Quick-start recommendation

Day one, set up Google News RSS for your company name and each executive name, plus 5-10 targeted trade publication RSS feeds. Nothing else. Run the workflow for a week. You will quickly see which sources produce signal and which produce noise, and you can expand from there. Do not start with seven APIs and three scrapers. That is the fastest way to a workflow that breaks in two weeks and you never fix.

Google News RSS – URL Recipes

# Basic company name search, US English
https://news.google.com/rss/search?q=%22Acme+Robotics%22&hl=en-US&gl=US&ceid=US:en

# Company OR product name (any match)
https://news.google.com/rss/search?q=%22Acme+Robotics%22+OR+%22Acme+Fleet+OS%22&hl=en-US&gl=US&ceid=US:en

# Executive with disambiguator
https://news.google.com/rss/search?q=%22Sarah+Chen%22+AND+(Acme+OR+robotics)&hl=en-US&gl=US&ceid=US:en

# Only results from the last 24 hours
https://news.google.com/rss/search?q=%22Acme+Robotics%22&hl=en-US&gl=US&ceid=US:en&when=1d

# Exclude a known false positive
https://news.google.com/rss/search?q=%22Acme+Robotics%22+-%22Acme+Brick%22&hl=en-US&gl=US&ceid=US:en

Four Legitimate Ways to Cover Paywalled Content

What not to do

Do not build anything that scrapes paywalled content, routes traffic through an archive.today-style bypass, or uses a headless browser to defeat access controls. Setting aside the copyright and terms-of-service problems, it breaks constantly and puts your brand in an awkward legal position if anyone asks how you got the article. A clean workflow uses approved access paths – licensed tools, publisher alerts, or public snippets – and that is the end of the list.

If the article matters enough to read in full, ask a team member who has a personal subscription to screenshot it, or pay for the individual article. Some outlets (WSJ, NYT) sell one-off article access for a few dollars.

Media Mention Classifier – System Prompt

You are a media monitoring analyst reviewing news mentions for a US-based company.

Your job is to review each item in the input list and return a structured JSON array with one object per item.

COMPANY AND EXECUTIVE CONTEXT:
[PASTE YOUR ENTITY LIST FROM STEP 1 HERE]

FOR EACH ITEM, DETERMINE:

1. is_relevant (boolean)
   - TRUE if the item is actually about the company or a listed executive
   - FALSE if it is a false positive (different company, different person, unrelated context)
   - Use the disambiguators in the entity list to decide borderline cases

2. relevance_confidence ("high", "medium", "low")
   - "high" = the entity is clearly named with matching context
   - "medium" = the entity is named but context is ambiguous
   - "low" = the entity might be mentioned but the match is uncertain

3. mention_type (pick one)
   - "feature"        = the company or exec is the main subject of the article
   - "quoted"         = an exec is quoted as a source
   - "passing"        = briefly mentioned in passing
   - "list_inclusion" = included in a list or roundup
   - "announcement"   = press release or funding/product news
   - "competitor"     = article is about a competitor and mentions our company for context

4. sentiment ("positive", "neutral", "negative", "mixed")
   - Base this on the tone toward OUR company specifically, not the article overall

5. topic_tags (array of 1-3 short tags)
   - Examples: "funding", "product launch", "leadership", "regulation",
     "partnership", "controversy", "earnings", "hiring", "industry trend"

6. executive_mentioned (string or null)
   - Name of the specific executive mentioned, if any

7. one_line_summary (string, under 25 words)
   - Write a neutral, factual summary of what the article says about our company
   - Do NOT copy sentences from the source. Paraphrase.

8. urgency ("high", "normal", "low")
   - "high" = anything negative, any crisis signal, any exec quoted, any major announcement
   - "normal" = standard coverage, routine mentions
   - "low" = passing references, old news resurfacing, list inclusions

OUTPUT FORMAT:
Return ONLY a valid JSON array. No preamble, no explanation, no code fences. Example:

[
  {
    "source_url": "...",
    "headline": "...",
    "publication": "...",
    "published_at": "...",
    "is_relevant": true,
    "relevance_confidence": "high",
    "mention_type": "quoted",
    "sentiment": "positive",
    "topic_tags": ["product launch", "funding"],
    "executive_mentioned": "Sarah Chen",
    "one_line_summary": "...",
    "urgency": "normal"
  }
]

RULES:
- If you are unsure whether an item is relevant, set is_relevant to false and relevance_confidence to "low". Better to miss a borderline match than to flood the briefing with false positives.
- Never fabricate URLs, publications, or quotes.
- If a field is unknown, use null. Do not guess.

INPUT ITEMS:
[PASTE THE RAW MENTIONS FROM STEP 2 HERE]

Model choice: Claude Sonnet 4 vs ChatGPT for this step

I tested both. For structured classification tasks like this, Claude Sonnet 4 is more consistent – it follows the JSON output format more reliably and is better at admitting uncertainty (which is exactly what you want for a filter). ChatGPT 5.3 is faster and cheaper but occasionally returns extra commentary around the JSON, which breaks downstream parsing. If you use ChatGPT, add “Return only the JSON array. Do not include any text before or after the JSON.” as the very first line of the prompt. That usually fixes it.

Daily Briefing Writer – System Prompt

You are writing a daily media briefing for the communications team at [COMPANY NAME].

INPUT: A JSON array of classified news mentions from the last 24 hours.

YOUR TASK:

1. Group items that cover the same underlying story. Items are the same story if they share a subject, a central fact, and are published within 48 hours of each other. Keep the earliest source as the primary link and list the others as additional coverage.

2. Drop items where is_relevant is false or relevance_confidence is "low".

3. Organize the remaining items into this structure:

## Daily Media Briefing - [Date]

### At a Glance
Three bullets that a busy CEO can read in 20 seconds:
- Total relevant mentions today
- The single most important story
- Anything urgent that needs a response

### High Priority
Items marked urgency: high. For each: headline, publication, one-line summary, link, sentiment. Group by story if there are duplicates.

### Standard Coverage
Items marked urgency: normal, organized by topic tag. Brief - one line per story.

### Passing Mentions
Items marked urgency: low. List format only - headline and publication. No summary.

### Executive Coverage
A separate section listing every item where executive_mentioned is not null, grouped by executive name. This makes it easy for each exec to see their own coverage.

### Sentiment Snapshot
A one-line summary: "Today: 12 positive, 4 neutral, 1 negative. The negative item is [brief description and link]."

RULES:
- Write in a neutral, factual tone. No hype, no alarm.
- Never copy sentences from source articles. Always paraphrase.
- If sentiment is negative or there is a crisis signal, flag it clearly at the top.
- Keep the total briefing under 800 words unless there is unusually heavy coverage.
- Format in clean Markdown that renders in Slack, email, and Google Docs.
- Use US date format (April 7, 2026) and US spelling.

INPUT:
[PASTE THE CLASSIFIED JSON FROM STEP 3 HERE]

n8n Workflow – Node by Node

Node 1: Schedule Trigger

Set it to run once daily at 7:00 AM in your local time zone. For high-stakes brands (public companies, crisis-prone industries), set it to every 2-3 hours instead.

Nodes 2-N: RSS Feed Reader (one per feed)

Add one “RSS Feed Read” node per Google News query and per trade publication feed. n8n makes each feed a parallel branch. Configure each node with the feed URL you built in Step 2. Set the filter to only pull items newer than the last run.

Node M: Merge

Add a “Merge” node to combine all the RSS branches into a single stream of items. Use “Append” mode.

Node M+1: Deduplicate by URL

Add a “Remove Duplicates” node that matches on the link field. This strips obvious duplicates before you spend AI tokens on them. Rough deduplication at this stage saves real money.

Node M+2: Format for Classifier

Add a “Code” or “Set” node that shapes each RSS item into the format your classifier prompt expects: { source_url, headline, publication, published_at, snippet }. Aggregate all items into a single JSON array to pass to the AI.

Node M+3: AI Classifier

Add an “Anthropic (Claude)” or “OpenAI” node. Paste the classifier prompt from Step 3 in the system field. Map the formatted JSON array into the user message. Set output to JSON.

Node M+4: Filter Relevant Items

Add a “Filter” node that passes through only items where is_relevant == true and relevance_confidence != "low". Everything else gets dropped.

Node M+5: AI Briefing Writer

Add a second AI node with the briefing prompt from Step 4. This one deduplicates by story, groups by urgency, and produces the final Markdown briefing.

Node M+6: Deliver

Add a “Slack – Send Message” node or “Gmail – Send Email” node. Map the briefing text as the body. For Slack, use a dedicated #media-coverage channel. For email, send to your comms team distribution list.

For the tech-curious: the API-level alternative

If you are comfortable with Python or Node, you can build this workflow as a single script running on a cron schedule instead of using Make or n8n. The core loop is: fetch RSS feeds with feedparser (Python) or rss-parser (Node), deduplicate by URL hash, call the Anthropic or OpenAI API with the classifier prompt, parse the JSON response, call the API a second time with the briefing prompt, and post the output to Slack via webhook. The whole thing runs in about 150 lines of Python and costs under $5 per month to run on a free-tier VM. The no-code version in this blueprint is the faster path for most people, but the script path gives you more control over rate limits, error handling, and multi-language coverage.

Failure Points and Fixes

Problem: The briefing is flooded with false positives about a company that shares your name.

Fix: Update your entity list exclusions. Add the specific competing names as exclusion terms. Re-run one day of mentions and verify the classifier now drops them. If false positives keep appearing after that, the problem is usually that your disambiguators are too weak – add more.

Problem: The classifier returns invalid JSON and the automation breaks.

Fix: Two fixes in order. First, switch to Claude Sonnet 4 if you are on ChatGPT. Second, add a “Code” node after the AI step that uses JSON.parse() inside a try/catch and routes invalid responses to a retry branch. In n8n, you can also force JSON output mode on the AI node, which mostly eliminates the problem.

Problem: The briefing misses obvious major stories.

Fix: This is usually a source coverage problem, not an AI problem. Check whether the missed story appeared in any of your RSS feeds at all. If not, add a feed that covers that publication. Google News RSS is wide but not comprehensive – add direct RSS feeds for the 5-10 publications that matter most to your industry.

Problem: Deduplication is too aggressive and drops distinct stories.

Fix: The briefing prompt groups stories that share “a subject, a central fact, and a 48-hour window.” If it is merging things that should stay separate, add an instruction: “Treat stories with different central claims as distinct, even if they share a subject.”

Problem: Your Anthropic or OpenAI costs spike unexpectedly.

Fix: Add the URL-level deduplication node earlier in the pipeline (before the AI call). Also check whether any of your RSS feeds are returning hundreds of items per run – usually a bad filter. A reasonable daily volume is 20-100 raw items; anything over 300 means one of your queries is too broad.

Problem: The Google News RSS feed suddenly returns zero results.

Fix: Google occasionally rate-limits or rotates its RSS endpoints. Test the feed URL directly in a browser. If it shows results in the browser but returns empty in automation, add a realistic user-agent string to your HTTP request. n8n and Make both let you set custom headers on RSS nodes.

Monthly Cost by Setup

Blueprint in the AI Automation Blueprints series.

Every blueprint is co-authored with AI and tested by me. Last updated April 2026.