Playbook: Trimming Brand Monitoring False Positives Using AI Context

The limitation of boolean logic

Every operator who handles brand reputation knows the drill. You set up a search for your company name. You wait for the data to roll in. And then you spend the next three hours manually deleting tweets, articles, and forum posts that happen to share a word with your brand but have absolutely nothing to do with your product. Traditional keyword monitoring buries operators in noise. When relying strictly on boolean logic, a brand name like "Apple" or "Coach" requires an intricate web of exclusions. Even then, the false positives leak through. The workflow we keep coming back to relies on replacing those massive boolean queries with an AI context layer. Sentia replaces exact string matching with semantic understanding. Here is exactly how to set that up, step by step, so your team can stop deleting spam and start acting on real customer signals. We routinely see teams transform their daily triage process from a painful chore into a streamlined, high-signal operation.

The scenario: Horizon Luggage

To show how this works in practice, we will use a realistic scenario. Imagine you are managing brand operations for "Horizon", a premium luggage and travel accessories brand. If you track the keyword "Horizon", your inbox will immediately flood. You will capture discussions about the video game Horizon Zero Dawn. You will capture posts about Animal Crossing New Horizons. You will capture real estate listings featuring a "broad horizon" view. You will see posts about event horizons in physics. Standard boolean logic forces you to create exclusions: NOT "Zero Dawn" NOT "Animal Crossing" NOT "physics". But gamers use abbreviations. Real estate agents use creative phrasing. Scientists write papers with unexpected vocabulary. The operators we work with find that maintaining these exclusion lists becomes a full-time job. Every time a new video game or movie is released with your brand name in the title, your metrics are ruined for weeks. Sentia approaches this differently. Instead of guessing every possible wrong context, you describe the right context once.

Step 1: Establish the baseline monitor

Before applying the AI classifier, it helps to understand what the baseline data looks like. In Sentia, navigate to your Monitor Settings and create a standard keyword rule for "Horizon". Let it run for twenty-four hours. If you query the brand_mentions_raw table, you will see the full firehose of data. You can also view this directly in the Sentia Inbox by selecting the unfiltered view. The metric to watch here is the sheer volume of incoming rows compared to the items you actually tag for your marketing team to review. We routinely see substantial volumes of daily mentions where only a fraction are legitimate brand discussions. This creates alert fatigue. If every notification on your phone is a false positive, you stop checking the notifications entirely. Operator trust in the data degrades. When the data is dirty, reporting becomes a manual process of exporting to a spreadsheet and deleting rows one by one. This is exactly the busywork Sentia is built to eliminate.

Step 2: Configure the Custom AI Classifier

Now we apply the filter. Go back to your Monitor Settings. Under the Data Processing section, toggle on the Custom AI Classifier. This feature sits between the raw data ingestion and your notification routing. It reads the text of the mention, evaluates it against a plain-language prompt, and assigns a mention_relevance_score. Instead of writing boolean exclusions, you write an identity statement. For our luggage brand, the prompt looks like this: "You are evaluating social media posts and articles for a premium luggage brand named Horizon. The brand makes suitcases, travel bags, and packing cubes. Mark the mention as relevant only if the text is specifically discussing the luggage brand, travel gear, or a related customer service experience. Discard mentions of video games, real estate, natural landscapes, or other companies with Horizon in their name." Sentia saves this definition and applies it to every incoming mention in real time. The AI classifier looks at the surrounding sentence structure, the images attached, and the overall context of the post to make a deterministic ruling on relevance. It understands that a user complaining about a broken zipper on their Horizon is talking about luggage, while a user complaining about a broken quest in Horizon is talking about a game.

Step 3: Route verified mentions to Telegram

Filtering the data is only useful if it improves your daily operational workflow. Most operators do not want to sit inside a dashboard all day. They want the critical alerts pushed to the surfaces where their team already communicates. Once your classifier is running, navigate to the Routing tab in Monitor Settings. Set a rule that triggers when mention_relevance_score is categorized as High. Connect this rule to your brand management Telegram channel. Now, instead of a constant stream of irrelevant alerts, your Telegram channel becomes a high-signal feed. When a notification chimes, your team knows it is an actual customer talking about your luggage. They can click the link in the Telegram message, jump straight into the Sentia Inbox, and respond to the customer or escalate the feedback to the product team. This routing step is crucial for maintaining response times. We routinely see customer service teams improve their initial response metrics simply because they no longer have to dig through the noise to find the actual complaints.

Step 4: Audit the results in the Sentia Inbox

No automated system should run entirely without oversight. The final step in this playbook is establishing a weekly audit routine. Because Sentia does not permanently delete the mentions that fail the AI check, you can always review the discarded data. Open the Sentia Inbox and filter by the status discarded by classifier. Spend ten minutes scrolling through these rows. You are looking for false negatives · instances where someone was actually talking about your luggage, but the AI missed it. Perhaps they used a new slang term, or they were discussing a specific product line you forgot to include in the prompt. If you spot a trend, you simply open Monitor Settings, append a new sentence to your AI prompt, and save it. The system will adapt immediately. The operators we work with typically refine their prompt two or three times in the first month. After that, the system runs with remarkable stability, freeing the team to focus on strategy rather than triage. The data in your brand_mentions_raw table remains intact, but the data you actually interact with is perfectly clean.