How to Check Whether ChatGPT Can See You: Entity Diagnostics in One Evening

Author: Дністер
Published: 2026-07-07T03:02:20.000Z
Language: en
URL: https://neurodrift.org/en/blog/yak-pereviryty-chy-bachyt-tebe-chatgpt/
Original (Ukrainian): https://neurodrift.org/blog/yak-pereviryty-chy-bachyt-tebe-chatgpt/

Seven steps with screenshots, three common gaps, and one question your radar should answer — not you.

-----
An air traffic controller in Boryspil sector stares at the screen at two in the morning. Almost empty. In the lower-right corner a blip drifts — a bright dot with no label. This is a primary return: the radar has bounced a signal off some aircraft's fuselage, so the controller knows something is flying there, knows the speed and heading — and nothing else. No tail number, no altitude, no idea who it is (Wikipedia, Secondary surveillance radar)). "All primary returns look alike," FAA documentation says in so many words, "and there is no way to tell which blip belongs to which aircraft except by its position on the map."

To go from blip to identified object, an aircraft needs a transponder and must squawk its assigned four-digit code — in controller slang, "squawk," from 0000 to 7777 (Pilot Institute). Without that code the pilot physically exists, engines roaring, fuel burning — yet for the system he doesn't exist as an identified entity. He is an anonymous dot that gets a vector of attention last.

Now imagine the screen is ChatGPT. The blip is you.

This article is about one specific mechanism: how ChatGPT decides you exist at all before saying a single word about you. The industry calls it "entity recognition." In client conversations I more often hear it put plainly: "AI doesn't see us." It passes as a convenient alibi — the black box is capricious, nothing we can do. In reality it is diagnosable, and you can do it yourself in one evening. Three lenses: how an aircraft becomes visible to the radar, what the system reads instead of your name, and why the first week of the right actions matters more than a year of honest SEO. The main point: invisibility in ChatGPT is not a verdict and not magic. It is a missing code you can check tonight and switch on within a week.

So-what: why this suddenly became your problem

Two years ago the phrase "ask ChatGPT who this contractor is" sounded like a joke. As of October 2025, it describes the behavior of roughly 800 million people per week — Sam Altman announced the number at Dev Day, which is nearly 10% of the world's adult population (TechCrunch, October 6 2025). Adults deciding whom to sign a contract with, which doctor to hire, whom to give an interview to — they increasingly ask the machine first, and only then open Google.

And here is the problem that concerns you specifically: the machine answers confidently even when it doesn't know. It doesn't say "no data" — it invents and delivers the invention as fact. For you this means silence won't save you: if the model lacks reliable data about you, it won't stay quiet — it will fill in the blanks. Researchers in 2025 stated this without euphemism: when asked about a less-known person, a model "confidently attributes incorrect achievements, birthplaces, or career histories, effectively mixing details from multiple individuals" (hallucination research review, 2025)). The model doesn't stumble or admit ignorance — "it simply presents the error as an undeniable fact." Anthropic's interpretability research the same year described the internal mechanism: when Claude recognizes a name but lacks sufficient data about the person, a false-brake fires — and the model generates a plausible fabrication (ibid.)).

Translation from machine-speak: if you are a primary return with no code, ChatGPT will not stay quiet about you. It will assemble your biography from pieces of other people's. And it will do so with the confidence of a notary.

Mechanism: how the system decides who you are

Let's name it plainly, because without a name the mechanism turns into mysticism. I work with the framework of "two visibility gates" — well articulated by AI-visibility practitioners, and it prevents the confusion between "nobody knows me" and "nobody cites me."

Gate one — parametric memory. What the model "learned" during training and holds in its weights. One binary question is decided here: does an entity under your name exist for the model, and is it anchored to your industry. How is it built? "Models recognize entities assembled from repeated, consistent descriptions across independent sources" — and a brand (or person) that exists only in its own marketing materials has not built such an entity (ALLMO). The key word is independent. Your website talking about you is an aircraft assigning itself its own code. The radar doesn't read that.

Gate two — real-time retrieval (RAG). This is when the model, while generating a response, reaches into the live web and pulls pages. What is decided here: will you be cited for a specific query today.

Two gates — two different diagnoses, and that distinction is critical:

Strong first gate, weak second → you are mentioned but not cited (the model knows you exist, but doesn't pull your fresh content).
Weak first gate, strong second → you are cited on a narrow query, but as a random blip, with no understanding of who you actually are.
Both weak → you are a primary return. An invisible aircraft.

And here is why this matters in numbers. The Semrush AI Visibility Index 2025 study found a "Mention-Source Divide": in most industries fewer than one brand in five manages to be both frequently mentioned and consistently cited as a source (Semrush). Most fall into exactly this gap between the two gates — not because they are poor, but because they are working on only one half.

There is one more subtlety that makes people give up prematurely. Parametric memory is "frozen" at the training cut-off: models train on static snapshots of the web, and a brand or person that appeared after that date has zero presence in the weights — only the second gate, real-time retrieval, remains (LeadSources). This sounds like a life sentence, but it's actually good news: you can open the second gate in a week, not years of training cycles. Even if you are "physically" absent from the model's memory, you can become visible tomorrow — as long as you let crawlers in and give them consistent facts. The radar doesn't need to remember your aircraft from a past flight; it just needs the transponder transmitting now.

Evidence ladder: seven diagnostic steps in one evening

Here is the tool. No code, no contractor — just you, a browser, and about ninety minutes. Follow the order strictly: each step narrows the diagnosis.

Step 1. Direct run in ChatGPT. Open ChatGPT and ask three ways: "Who is [your name]?", "What do you know about [your company]?", and — most important — "Who would you recommend for [your service] in [your city/niche]?" The third question is the one that counts: the first two test memory, the third tests whether the entity is tied to a query that actually brings in money. Screenshot every response — this is your "zero point" to compare against in a month after making corrections. If you're absent in the third, your problem isn't obscurity — it's misalignment. One small but important detail: turn off chat memory and run again in a history-free session (or in someone else's account). Otherwise the model may "recognize" you only because you told it about yourself in earlier conversations — and you'd be testing your own echo, not the real entity.

**Step 2. Check whether the model is inventing you.** Reread the step 1 response for quiet lies: facts blended with someone else's, wrong city, another person's job title attributed to you. This is not cosmetic — if the model is already composing your biography, you don't have an empty blip; you have a blip with a false code, which is worse.

Step 3. Find your Google Knowledge Panel. Google your own name or brand and look at the right side of the screen on desktop (or at the top on mobile) (Google Knowledge Panel Help). Panel exists → you already have a recognized entity in Google, a strong signal for all machines. No panel → your entity is not yet "verified." This is the most common diagnosis for smaller players.

Step 4. Check Wikidata. Go to wikidata.org and search for yourself or your company. Every record has a unique identifier — a QID, the letter Q plus digits (Berlin is Q64, to give a sense of scale) (Wikidata:Identifiers). If you have a Wikipedia article, find the QID via Tools → Wikidata item (Wikipedia: Finding a Wikidata ID). Wikidata is the public database from which machines draw "official" facts. Not there → you are an entity without a passport.

Step 5. Check whether your site lets AI crawlers in at all. Open yoursite.com/robots.txt in a browser and look for Disallow lines targeting GPTBot (OpenAI), ClaudeBot, PerplexityBot (Search Engine Journal, AI crawler agents list, December 2025). Many companies block them accidentally — someone once copied another site's template. It's like taping over your own transponder and wondering why the radar is silent.

Step 6. Run robots.txt through a tester. Bing Webmaster Tools has a free robots.txt tester that will show exactly which line blocks which agent (Bing Webmaster Tools). Five minutes, and you have a black-and-white answer instead of guesswork.

Step 7. Check the consistency of your facts across three places. Open side by side your website, your profile in a professional directory, and any press mention. Cross-check: name, description, founding year, city, service offered. LLMs are "extremely sensitive to conflicting data — high variability leads to exclusion" (ALLMO). Three different spellings of your name = three different aircraft on the radar, none of them yours.

A specific trap for non-English speakers: transliteration. If you are "Petro Koval" in one place, "Petro Kovalenko" in another, and only in Cyrillic on your website with no Latin equivalent, the model will not consolidate these into one entity. To it, they are three different people with similar occupations. Write down one canonical version of your name and company — and make sure it appears everywhere, down to the last detail. Tedious work of about an hour, but often that is the exact "code" that was missing.

Human mirror: you've seen this mistake before

Think back to the last time you asked ChatGPT about someone you knew — a mid-level colleague, not a star — and got a response where something was off. A job title from a previous life. A city they'd never lived in. A project they had nothing to do with. You probably smiled and closed the tab.

Now imagine the person on the other side of that screen was deciding whether to give you a €40,000 contract — and was reading your biography in the same style. They didn't smile. They quietly crossed you off the list because "something doesn't add up." You never found out: invisibility sends no notifications.

That is the worst property of this whole story — its silence. A bad Google review you'll see and respond to. A rejection letter you'll read and learn from. But when the model assembles your biography alone with a potential client at eleven-thirty at night, you receive no review, no letter, not even the knowledge that the conversation happened at all. The contract simply doesn't come. The call doesn't ring. And you chalk it up to the market, competition, seasonality — anything except the four-digit code that was missing. The radar sends no receipt for an aircraft it failed to identify; it simply directs attention past it. That is precisely why diagnostics are worth doing before you feel a symptom — because there will be no symptom, only silence, which is easy to mistake for normal.

Data: three common gaps and what to do about them in week one

Here is the one table worth saving. Match your result from steps 1–7 — and you'll see which of the three gaps you fall into. Almost everyone falls into one, rarely two.

| Symptom (what diagnostics showed) | Gap (mechanism) | What to change in week one |
|---|---|---|
| No Google panel, no QID in Wikidata, direct ChatGPT run returns "I don't know" | Entity not built — only you talk about yourself | Get independent sources to mention you: a profile in a professional directory, a mention in an industry publication, a Knowledge Panel claim (Google) |
| Entity exists (panel/QID present), but you're absent on "who do you recommend for…" | Entity not anchored to the industry/query | Align your description everywhere so that your name appears alongside your service and the client's problem — consistently across all touchpoints |
| You are mentioned but facts are mixed/outdated, or robots.txt blocks GPTBot | Entity polluted or closed | Remove Disallow for AI agents (SEJ), reconcile conflicting facts across three places (step 7), submit corrections in the panel |

Notice one figure that explains why. A Muck Rack study (July 2025, millions of AI-cited links from hundreds of thousands of prompts) found that over 95% of sources AI cites are free-access, and around 85% of those are earned media — mentions in independent publications, not your own website (Muck Rack / National Law Review). In fairness: other measurements disagree — a Yext study of 6.8 million citations says instead that 86% of sources come from brand-controlled channels (website, directory listings) (Yext). The difference is methodology: Yext counts from a specific query with specific intent; Muck Rack counts from the total pool. But both converge on the practical conclusion: a significant share of the work lies outside your own domain, where you influence rather than control.

Counter-pressure: isn't this all just marketing fear?

Here I'll raise the question against myself, because otherwise this is a pitch, not a diagnosis. Aren't the people selling "AI visibility" inflating the industry the way ten years ago they inflated "you're not on the first page of Google"?

Partly — yes. An entire consulting industry has spent the last two years frightening people with "invisibility" in exactly the same way. Healthy skepticism applies, and here are three sober caveats.

First: correlation is not causation. The fact that fewer than one brand in five clears both gates does not prove that working on your entity causes visibility. It's equally possible that large, recognized brands are already present everywhere, and a clean entity is a consequence of their scale, not the cause of their visibility. Most likely it runs both ways.

Second: models change monthly. Today's gap may close on its own with the next update — or a new gap may open that didn't exist yesterday.

Third — and most important: what would falsify this entire article? If you ran all seven steps, built a clean entity, aligned your facts — and three months later a direct run still returned "I don't know" or fiction. If that happens in your case, entity is not the deciding factor in your niche, and there's nothing to pay for. That is an honest test, and it is falsifiable — unlike the magic that gets sold.

Why now, in 2018–2026? Three curves converged. In 2018 voice assistants trained people to ask machines out loud, but the machines couldn't yet answer meaningfully. In 2022–2023 ChatGPT learned to answer — and to lie — with confidence. And in 2025 those 800 million per week (TechCrunch) made it the first contact point with your reputation. The window when "AI doesn't see us" was an exotic curiosity has closed.

Re-plating: when to call a specialist, when to handle it yourself

The seven steps above you do yourself, in one evening, for free. That is diagnostics — and it's yours, don't delegate it, because only you know which queries actually bring in money.

Calling a specialist makes sense for treatment, and even then not always. If the diagnosis is "entity polluted or closed" (third row of the table), that is mostly technical hygiene taking an evening or two: fix robots.txt, reconcile facts, file a panel claim. No contractor needed.

But when the gap is "entity not built," and you're talking about a person or company whose reputation is read before contracts worth tens of thousands — building independent presence by trial and error can take years. This is where systematic work makes sense: I've seen it set up as a pipeline by teams building "reputation nodes" with verifiable sources. In our partner practice at WikiBusiness this is a dedicated analytical track — diagnostics on the entity come first, and only then the decision on whether there is anything to build at all. A straight-talking contractor will start from the same diagnostic evening and tell you honestly if there's nothing to treat. The one who jumps straight to selling you a package is selling fear, not code.

Hard kicker

Return in your mind to that control room screen at two in the morning. A blip drifts in the corner, nameless, and it will get a vector of attention last — not because the pilot is bad, but because he is not squawking his code.

Tomorrow someone will decide whether to do business with you — and will ask not Google but the machine: "who do you recommend." The machine will answer with the confidence of a notary. One question: will that answer be you — or someone else's biography, assembled in the silence you left behind.

The seven steps above are that code, the one that turns a blip into an identified aircraft. They're free, and they take one evening. Less than you just spent reading about a problem that, until now, was someone else's. Open ChatGPT and ask about yourself — right now.

---

Partnership disclosure: WikiBusiness is our ongoing partner; mentions in this article are presented as an example of analytical practice, not as a commercial offer.