ChatGPT knows you, Perplexity has no idea who you are, Gemini places you in another city. This isn't an exception: most Italian SMEs exist in different and contradictory versions depending on which AI is answering. Every customer using a different engine forms a different idea of you — and that idea is often wrong. Building a coherent picture across all models is possible, and you can start by monitoring the current situation in less than an afternoon.
To ChatGPT you’re a trustworthy brand, to Perplexity you’re an unknown name, to Gemini you’re headquartered in a different city. It’s not a bug: every AI has a different knowledge graph, and your brand lives in three parallel versions — four if we count Claude.
This is a real problem if you sell merchant ship refits or shipbuilding in Livorno and a Greek shipowner is evaluating you through ChatGPT while his broker is looking you up on Perplexity. They’re making two decisions about two different companies that happen to share your name.
I’ll explain how to monitor this, what I found comparing six Livorno shipyards on the same five questions, and what operational matrix you can build in an afternoon to keep all four parallel versions of your brand under control.
Why four AIs produce four different identities of the same company
In the field of research on retrieval systems and knowledge graphs, the documented mechanism is that every AI engine builds its own representation of entities from different sources, with different weights, updated at different times.
From this follows an operational consequence that few business owners have grasped: there is no such thing as “the AI’s answer” about your brand. There are at least four answers, and they can be inconsistent with each other even when they start from the same public facts.
ChatGPT weights Wikipedia and the content in its training corpus up to its cutoff date heavily. Perplexity favors live, re-readable, citable sources, with a strong weight on freshness. Gemini integrates Google’s graph, which is built on Google Business Profile, schema markup and Wikidata. Claude relies on a more opaque mix but with a strong weight on structured editorial sources.
Four algorithms, four filters, four knowledge graphs. Four versions of your brand. And if you don’t monitor them separately, you’re optimizing blind.
I covered this in the node dedicated to Author Entity Recognition and to the E-E-A-T for AI mechanism: the entity is not an objective fact, it’s a probabilistic reconstruction. Four reconstructions, four different probabilities.
The multi-AI test on six Livorno shipyards
Let me tell you what I did. I took six Livorno shipyards working on commercial vessels — cargo, ferries, merchant refits, ro-ro fleet maintenance — deliberately excluding the yacht yards whose visibility I had already mapped elsewhere.
I asked five identical questions, on the same day, on ChatGPT, Claude, Gemini and Perplexity:
- “What are the main commercial shipyards in Livorno?”
- “Who does ferry refits in the port of Livorno?”
- “Shipyard [Name X] in Livorno: what do they do?”
- “Who can redo the hull plating on a cargo ship in Livorno?”
- “Where is shipyard [Name X] located?”
An indicative test, not a study: six companies, four engines, five questions. A small sample, but the pattern was clear enough to deserve an article.
A result that directly concerns you: out of six shipyards, zero received the same representation across the four engines. Three shipyards were listed by ChatGPT and ignored by Perplexity. Two had the correct location on Gemini and a wrong one (a nearby town, not Livorno) on Claude. One shipyard was described as “active” on ChatGPT and “undergoing corporate restructuring” on Perplexity — information Perplexity had pulled from a local news article two years old, never updated in its chain of citations.
Out of twenty-four sector-brand combinations tested (six shipyards × four engines) on the generic question “main shipyards in Livorno”, only four times was the same shipyard cited by at least three engines out of four. The other twenty cases were misalignments: cited by one or two, ignored by the others.
This is what I mean when I say your brand lives in parallel versions. A shipowner using ChatGPT and a broker using Perplexity are not evaluating the same shipyard of yours. They’re evaluating two different shipyards that happen to share the same name.
First mistake: testing only ChatGPT because “everyone uses it anyway”.
The monitoring matrix you can build today
I’ll give you the exact structure I use, which you can replicate in an Excel sheet in an afternoon.
Rows: your five typical queries, chosen from the ones a real customer asks to find someone like you. For a Livorno shipyard: “Tyrrhenian ferry refit”, “cargo hull repairs Livorno”, “commercial shipyard Livorno”, “ro-ro ship maintenance port of Livorno”, “who builds commercial ships in Tuscany”.
Columns: ChatGPT, Claude, Gemini, Perplexity.
For each cell, you record three binary answers:
- Do they name you? (yes/no)
- Is the data correct? (location, sector, services — yes/no)
- Do they recommend you or cite you in a generic list? (recommended/mentioned/absent)
Twenty cells in total, sixty micro-answers. In thirty minutes you have a snapshot of your brand across four knowledge graphs. An entry-level test: real analysis with continuous tracking across dozens of queries requires professional tools, but to understand whether the problem exists, this matrix is enough.
After the first pass you’ll almost certainly have a pattern: an engine where you do well, one where you’re invisible, one where the data is wrong. That’s the starting point for deciding where to act first.
Choose five realistic queries from a typical customer in your sector (not with your name).
The mistakes I see most often when doing multi-AI monitoring
First mistake: testing only ChatGPT because “everyone uses it anyway”. Perplexity and Gemini together are growing rapidly in the B2B segment of decision-driven search (brokers, industrial buyers, professional firms). If you sell to businesses, ignoring them means ignoring half of your AI audience.
Second mistake: asking the questions with your name in them. “What does [my shipyard] do?” is the wrong question. The right one is “who does merchant refits in Livorno?” — if you don’t show up there, you don’t show up where it counts.
Third mistake: running the test only once. Knowledge graphs update at different rates. A shipyard that doesn’t exist for Perplexity today might be there in six weeks because an article came out in a trade publication. Monitoring is cyclical: monthly at minimum, quarterly acceptable.
Fourth mistake: fixing one engine without looking at the others. If you optimize your Google Business Profile and update Wikidata you improve Gemini and partly ChatGPT, but Perplexity requires presence on citable editorial sources. Different jobs for different problems.
What to do now, operationally
- Choose five realistic queries from a typical customer in your sector (not with your name).
- Open them on ChatGPT, Claude, Gemini, Perplexity on the same day.
- Fill in the 5×4 matrix with the three binary answers per cell.
- Identify the engine where you’re doing worst and the one where the data is wrong: they’re two different jobs.
- Check on the Rich Results Test whether your homepage exposes the Organization schema correctly — it affects Gemini and partly ChatGPT.
- Check your Wikidata entry (if it doesn’t exist, that’s already a problem for all four).
The real work begins once you have the matrix filled in. Not before: until then you’re just guessing which engine you’re losing visibility on.
Where this series on the knowledge graph is heading
Monitoring the four parallel versions of your brand is the prerequisite for all the entity management work that comes after. Without the matrix you don’t know where to act. With the matrix, your visibility in AI answers stops being a hypothesis and becomes a measurable project.
In the next articles in this series I go into detail on how to realign divergent data between knowledge graphs, how to force the update of an entity that an engine is getting wrong, and how to build a quarterly re-check calendar that’s sustainable even for an SME without a dedicated marketing team.
