Introducing Watchdog: Your Brand's AI Fact-Checker

ChatGPT has over 400 million weekly active users. Perplexity, Gemini, and Copilot are growing fast. When these tools answer questions about your industry, they make factual claims about your company. Sometimes those claims are wrong.

In February 2024, a Canadian tribunal ordered Air Canada to pay damages after its AI chatbot fabricated a bereavement discount policy that did not exist. The company argued the chatbot was a separate entity. The tribunal disagreed. Google lost $100 billion in market value when Bard hallucinated a false claim in a promotional demo. In October 2025, Deloitte issued a partial refund on a A$440,000 government report after AI-generated content was found to contain fabricated academic sources and a fake court quote.

These are high-profile cases. The smaller ones go undetected. A SaaS company discovers ChatGPT is telling prospects it has features that only exist in a competitor's product. An automotive brand finds AI models citing safety recalls that never happened. A financial services firm sees fabricated executive quotes attributed to its CEO.

According to MIT research published in January 2025, AI models use more confident language when hallucinating than when stating facts. The wrong information sounds more authoritative than the right information. That makes these errors harder for consumers to catch on their own.

The Vectara hallucination leaderboard tracks factual accuracy across major LLMs. Even the best-performing models hallucinate at measurable rates. The average general knowledge hallucination rate sits around 9.2%. Some models reach nearly 30%.

These rates apply to general knowledge. Brand-specific facts, which are less represented in training data, are likely hallucinated at higher rates. If your company changed its pricing six months ago, raised a funding round, or launched a new product, AI models may not reflect those changes for months or years.

A KPMG global study from 2025 found that less than half of respondents (46%) are willing to trust AI. Trust is declining even as adoption grows. Every hallucinated fact erodes the trust your customers place in the information AI delivers about you.

The AI visibility monitoring space has grown significantly. Profound, Otterly.ai, Peec AI, Semrush, and others now track brand mentions across AI platforms. They answer a valid question: is my brand showing up in AI responses?

That question matters. But it is only half the picture. Being mentioned is useless if the mention contains a fabricated pricing claim, an outdated product description, or a nonexistent partnership with a competitor.

Most tools in this space, including Profound (which raised $96M at a $1B valuation in February 2026), Otterly.ai (starting at $29/month), and Peec AI (backed by a EUR 21M Series A), focus on visibility analytics and sentiment. They are useful tools for understanding reach. They do not tell you whether what AI says about you is true.

Watchdog starts by building a fact profile of your company. It researches your brand across the web, gathering verified facts about your identity, leadership, financials, market position, products, and recent news. Each fact is tagged with a confidence level and source freshness. You can review, edit, and add your own facts to the profile.

Every time Sill runs a monitoring cycle, Watchdog classifies every AI response where your brand was mentioned. It extracts two things: caveats (qualifying statements that limit the recommendation) and factual claims (concrete assertions about your company). Each claim is compared against your fact profile to determine whether it is accurate, unverified, or a contradiction.

Watchdog uses semantic deduplication to group similar claims and caveats together, so you see patterns rather than noise. If three different AI platforms phrase the same pricing error in slightly different ways, Watchdog recognizes it as one issue, not three.

Watchdog generates alerts across seven categories. Each one represents a different kind of risk to your brand in AI outputs.

Each alert includes the exact AI response that triggered it, the platforms involved, and how many times the issue has recurred. For factual contradictions, Watchdog shows both what the AI claimed and what your verified fact profile says.

Watchdog does not fire alerts on day one. The first three monitoring runs are a learning period. During this phase, Watchdog classifies every AI response and builds baselines: what caveats appear consistently, which claims recur, and how different platforms talk about your brand. No alerts are shown during this phase to avoid false positives from incomplete data.

Starting from the fourth run, alerts begin. But the learning does not stop. Every action you take teaches Watchdog what matters to you.

When you dismiss an alert, Watchdog uses embedding similarity to suppress semantically equivalent alerts in the future. If you dismiss "pricing is expensive," it will also suppress "the cost is on the higher end" and "not the most affordable option" without you having to dismiss each variant individually. You can set suppression to be permanent or temporary (one week, one month, or three months).

When you confirm a factual contradiction, Watchdog updates your fact profile and strengthens its ground truth for future comparisons. When you tell it "this is actually correct," it adjusts its records to match what the AI got right. Over time, the fact profile becomes a living document that reflects the most up-to-date truth about your company.

This feedback loop means that after a few weeks of use, Watchdog surfaces fewer false positives and catches more genuine issues. The signal-to-noise ratio improves with every review cycle.

The fact profile is the foundation of everything Watchdog does. It contains structured, verified facts about your company organized into four categories: identity, leadership, financials, and market position. Each fact includes a source URL, a confidence rating, and a freshness indicator.

Watchdog builds the initial profile automatically by researching your brand across the web. You can then review every fact, correct errors, add information the research missed, and set how quickly facts should be considered outdated. A fast-moving startup might set "current" to 30 days and "recent" to 90 days. A more established company might use 90 and 365 days.

When the classifier encounters a claim that contradicts a fact marked "outdated," it treats the conflict with more nuance. A company that changed its CEO two years ago might legitimately have outdated references. A pricing change from last week is a different kind of problem. Freshness metadata makes this distinction possible.

The World Economic Forum's 2025 Global Risk Report ranked AI misinformation as a top concern for the second consecutive year. G2's 2025 Buyer Behavior report found that AI-generated recommendations are the number one source influencing B2B vendor shortlists. A Resolver survey found that 35% of brands report inaccurate AI responses have damaged their reputation.

These numbers will grow. As AI search grows from 2-6% of organic traffic toward the 25% migration Gartner predicted, the cost of unchecked hallucinations grows proportionally. A wrong fact in a Google result at least has a link you can click to verify. A wrong fact in a ChatGPT response is presented as authoritative truth with no obvious source to check.

RAG (retrieval-augmented generation) reduces hallucination rates by roughly 71% when used properly. But brand-specific information often falls outside RAG knowledge bases, especially for mid-market companies without Wikipedia pages or extensive press coverage. The brands most vulnerable to hallucinations are the ones least likely to have robust information available for AI retrieval.

Watchdog is available to all Sill monitoring subscribers. If you already have monitoring set up, Watchdog activates automatically. Your fact profile begins building as soon as your monitoring configuration is saved.

The first step is reviewing your fact profile. Watchdog's automated research is thorough, but you know your company better than any web scraper. Add any facts it missed, correct any it got wrong, and set your freshness preferences. The more complete and accurate your fact profile, the more useful Watchdog's alerts will be.

After three monitoring runs (typically three days), alerts start appearing. Review them in the Watchdog dashboard, where you can filter by alert type, platform, and status. Confirm contradictions, dismiss known caveats, and watch the system learn your preferences.