You don't know what AI is saying about you right now — and you've almost certainly never checked. I'm not talking about knowing whether you show up: I'm talking about knowing whether you show up with the right location, the right services, the correct information. Every error the models silently repeat is read by potential clients who never call you to flag it. Thirty minutes a month is enough to know what AI says about you — and to step in before the error reaches the wrong client.
Open ChatGPT today and ask the 3 questions a prospective client would ask about your brand: “what does [brand] do?”, “where is it based?”, “who does it work with?”. Note down what it answers. Repeat in a month. If you don’t, you have no idea how AI sees you.
It seems like a trivial exercise, but it’s the only serious way to understand whether the factual data about your brand is being reported correctly in AI answers. I’m not talking about visibility (you measure that with other metrics), I’m talking about accuracy: location, founder, year of founding, clients, products. If AI attributes the wrong location or a service you don’t offer to you, the problem isn’t that you don’t show up, it’s that you show up badly. And correcting an error already rooted in the models is upstream work, on the structured sources.
In this article I explain how to build a system for monitoring AI answers focused on the brand entity, why it’s different from classic visibility tracking, and what to do when you find errors. At the end you’ll find a test you can run in 30 minutes.
What it means to monitor the entity, not visibility
When SEOs talk about AI tracking they almost always think of “is my brand being cited or not?”. It’s a legitimate but partial question. There’s another, more subtle one, that concerns who you are to the model: your brand as an entity has a factual record in the knowledge graph that AI engines consult, and that record can be right, wrong or incomplete.
In the world of research on fact-checking language models, Rahman and colleagues have shown why this is a critical topic.
“Consequently, LLMs can generate misinformation, making robust fact-checking essential.” — Rahman et al., 2025
Translated: language models can generate wrong information, and for this reason systematic verification work is needed. It follows that, if your brand doesn’t have a periodic control process on what AI says about you, you’re delegating your factual reputation to chance. The business implication is direct: a prospect who asks ChatGPT “what does [your brand] do?” receives an answer you’ve never seen or validated, and that may contain errors that are hard to correct after the fact.
Entity-specific monitoring connects to the foundational work you’ll find in my previous articles: first you have to exist as a recognized entity (I talked about this in named entity recognition and in the periodic entity audit), then you have to check that the data associated with that entity is correct. Without the first step the second is impossible; without the second, the first leaves you at the mercy of silent errors.
The entity-specific prompt set
The heart of the system is a set of prompts focused on the verifiable facts about your brand. Not market questions (“who are the best producers of X?”), but precise questions about your entity.
The base set is this one, adapt it to your context:
- Where is [brand] located?
- Who founded [brand]?
- What does [brand] do?
- When was [brand] founded?
- Who does [brand] work with?
- What products does [brand] make?
- Which countries does [brand] export to?
- Who is the CEO of [brand]?
- What is the history of [brand]?
- What certifications does [brand] have?
Ten prompts, four AI engines (ChatGPT, Claude, Gemini, Perplexity), monthly cadence. Total: 40 answers a month to archive and compare. It’s 30 minutes of work if you do it by hand, less if you automate it with professional tools.
If AI attributes the wrong location or a service you don’t offer to you, the problem isn’t that you don’t show up, it’s that you show up badly.
The test I ran on 10 food export producers from Modena
To see how relevant this monitoring is in practice I ran an indicative test on a small but targeted sample: 10 Modena-based producers of balsamic vinegar, aged cheeses and cured meats, all with export business toward Europe and North America. I asked the 3 base questions (what it does, where it’s based, who it works with) to ChatGPT, Claude, Gemini and Perplexity. Total: 120 answers.
The results, on an admittedly small sample:
- Location errors: 3 brands out of 10 received from at least one engine a wrong location (Modena / Reggio Emilia mix-up, or generic attribution to “Emilia-Romagna” when the designation of origin requires municipal precision)
- Service errors: 4 brands out of 10 had services/products described inaccurately (a producer of traditional DOP balsamic only, described as a “wine vinegar producer”)
- Client/distributor errors: 6 brands out of 10 had at least one piece of wrong information about distribution partners, often based on years-old data
- Total silence: 2 brands out of 10 received answers like “I don’t have specific information” from 2 engines out of 4
It’s a small test and not a structured study, but the pattern is clear: even brands with good market positioning have a factual record in the AI engines’ knowledge graph that is inaccurate. And the most frequent problem isn’t absence, it’s error.
The underlying mechanism is the one described in the work of Rahman and colleagues: existing fact-checking benchmarks have known limits.
“The majority of benchmarks either lack the realistic complexity of real-world claims and are domain-independent or are too narrow to be generalized.” — Rahman et al., 2025
Translated: fact-checking systems for language models suffer from benchmarks that are either too generic or too narrow. The operational consequence for you is that you can’t rely on the models to self-correct: you have to be the one to monitor the errors and correct the structured sources upstream.
After each correction, wait 4-8 weeks and rerun the test: AI models don’t update in real time, but re-indexing and re-training cycles pick up the structured corrections.
The errors I see most often
Working with Italian SMEs I’ve seen 4 error patterns recur:
Location confused with region. A Gragnano producer is often described as a “Campania company” without the reference municipality. For a food brand with DOP/IGP this imprecision cuts communicative value. The source of the problem is often an empty Wikidata record on the `P131` field (administrative entity) or a Google Business Profile with a generic address.
Wrong founding year. It happens when the year is cited in conflicting contexts on the site (e.g. “since 1952” in the claim, but “founded in 1947” in the about us section). AI takes the most frequent value, not the correct one.
Confusion with namesake brands. A balsamic vinegar producer can inherit attributes of a wine brand or a farm with a similar name. It’s an entity disambiguation problem that I covered in entity disambiguation.
Obsolete partnerships. “Works with [distribution chain]” when that relationship has been over for 2 years. Models keep dated information if the web isn’t updated.
How to correct errors at the source
When monitoring detects an error, the correction doesn’t happen “in the model” but in the structured sources the model uses to build the factual record. In the world of research on model grounding, this is the key point: retrieval mechanisms draw on web sources, and if the sources are wrong the answer will be wrong.
The concrete levers:
- Wikidata: open wikidata.org, search for your brand’s entity, check the location fields (P131, P17), founder (P112), founding date (P571). If they’re missing or wrong, update them with a reliable source
- Organization schema on the site: open Google’s Rich Results Test, paste your homepage, verify that the Organization block is present and complete (name, address, founder, foundingDate, sameAs)
- Google Business Profile: go to business.google.com, verify address, category, description. It’s one of the sources most weighted by AI engines for local data
- The site’s “about us” page: check consistency between founding year, founder, location. There must be ONE version, not variations
After each correction, wait 4-8 weeks and rerun the test: AI models don’t update in real time, but re-indexing and re-training cycles pick up the structured corrections.
Your 30-minute audit
Before wrapping up, the bare operational minimum:
- Write your 10 entity-specific prompts (use the set above, adapted to your business)
- Submit them to ChatGPT, Claude, Gemini, Perplexity. Save the answers in a spreadsheet
- Note every factual error (location, service, clients, year, founder)
- For each error, identify the structured source to correct (Wikidata, schema, GBP, about us page)
- Repeat the test in 30 days and compare
It’s an entry-level audit: the real analysis, on larger samples and with automated tracking, requires professional tools. But this first step gives you immediate visibility into what AI says about your brand today, and puts you in a position to act on the sources before a prospect receives the wrong information.
Why this work impacts your visibility in AI answers
Entity-specific monitoring isn’t a vanity tracking activity. It’s the quality control of the data that AI engines use to decide whether to include you in an answer, how to describe you, and in what context to cite you. A brand with a clean and consistent factual record is selected with more confidence by AI answer systems; a brand with contradictory data across Wikidata, schema and site gets downgraded or cited badly.
If you want to continue along this line, in the next articles of the series I’ll show you how to structure a competitor entity graph tracking, how to manage reconciliation between conflicting sources (you’ll already find something in knowledge graph reconciliation), and how to connect this work to the broader one on E-E-A-T for AI.
How visible is your brand to AI?
Find out in 30 seconds with our free tool. 11 automatic checks, instant results.
