I Interrogated 6 AI Models About Myself Through a Procurement Manager's Eyes: Amazon, Esoterica, 'Telo Hudelo' and 5 Holes We Closed in a Day

In October 2026, AdIndex published a study that made me uneasy: 56% of Russian companies check a prospective contractor through search engines and AI models before the first call. 40% of corporate clients ask AI for a general overview, prices and reviews, 25% look for financial figures, and 20% dig for lawsuits and compromising material. While I polish slides for a pre-sales pitch, a bank's security officer or a procurement director opens ChatGPT, Perplexity or Gemini and types: «Build a dossier on this contractor, check its sole proprietorship, court cases and real case studies».

I sell enterprise AI adoption. If an AI model lies about me, it lies to my client before I get to say hello. I did not want to guess what a procurement manager sees, so I wrote a script and ran a cross-examination of six AI models. Spoiler: it was funny right up to the moment Gemini read out our closed price list.

How I set up a cross-examination of six models

I queried ChatGPT, Google Gemini, Perplexity Pro, Anthropic Claude, Grok and YandexGPT about three subjects: my surname (Zhemal Khamidun), my products (AlpinaGPT, Vibe Academy, hamidun.com) and my legal entity (sole proprietor Khamidun Zhemal Uldovich, INN 695007138610). Each question ran in two modes:

• L1 — blind, from the model's weights alone. This is how a model answers when the user has web search turned off.
• L2 — with live internet search and counterparty databases.

I wrote the prompts in a procurement manager's voice: «check the contractor», «are there any lawsuits», «how much do the services cost», «what are the real case studies». The script logged the answers in a table, then I read them myself and checked every fact against reality.

What did the AI models invent about me without web search?

In L1 mode, not one model said «I don't know this person». Each stitched together a plausible detective story from scraps of tokens, with such confidence that I briefly doubted my own biography. The main L1 conclusion: without search access, an AI model fills the void with fiction, and the fiction sounds more convincing than the truth.

Gemini turned me into an Amazon guru and invented my INN

Google Gemini 3.1 Pro, without search, declared with a straight face that I am «a well-known Russian-language e-commerce expert on trading on the American Amazon using Private Label and FBA models, and on unblocking banned accounts (Suspension Appeal)». The price of personal mentoring: from 1 000 to 5 000 dollars. I have never sold on Amazon in my life.

It gets more interesting. I asked Gemini to name my INN from memory. The model did not blink and generated a Moscow INN, 771891961657, with a valid checksum and OKVED code 47.91, mail-order retail. My real INN starts with 69, which is the Tver region. A procurement manager who does not double-check the numbers gets a nonexistent legal entity with plausible details in the dossier.

GPT-4.1 sent my academy into esoterica

OpenAI gpt-4.1, without search, answered a question about Vibe Academy, my vibe-coding academy. The model broke the name down to its roots: vibe means vibration, atmosphere. It then helpfully warned the client that courses with such a name are often devoted to «esoterica, managing your emotional state, inner energy and women's practices».

We teach people to build products with AI coding. The only inner energy we have is in the coffee machine.

Claude and Grok added crypto channels and Dubai

Without search, Claude confused my master's degree at MIPT with MGIMO and listed me as the author of a Telegram channel about crypto trading and Web3. Grok-3 went furthest and drew a portrait of the owner of a network of anonymous channels, with Dubai real estate and Roskomnadzor fines for failing to label advertising.

Now picture a security officer inside a corporate perimeter where company policy has disabled web search in the chatbot. That is exactly the text they will get.

An AI model without search never says «I don't remember».

It makes things up, and it makes them up with a checksum.

What changed when I turned on live search?

With web search on, Amazon, esoterica and Dubai vanished from the answers of all six models. The search-enabled models assembled my real profile accurately:

• master's degree at MIPT;
• leading the Agile transformation at Sberbank, 1 300+ specialists;
• consulting at Accenture: AB InBev, Mars, Metalloinvest;
• e-commerce growth at Ulybka Radugi: +436% online sales and +1860% app orders;
• Head of AI at Alpina Digital and CPO of the AlpinaGPT corporate platform: 9 000+ employees at 60+ companies in Russia.

The legal history also came out clean: 0 lawsuits in KAD Arbitr, 0 enforcement proceedings in FSSP, 0 entries in the register of unscrupulous suppliers. At that point I exhaled and was ready to close the laptop. Too soon: live search dug up things we never suspected.

Which five real holes did live web search find?

Search hands the AI model everything in the index, with no «closed» or «outdated» labels. That is how five holes surfaced: leaked closed commercial proposals, a ghost weight-loss domain, a collision between two OGRNIP numbers, distrust of our own landing pages, and weak visibility in blind category search. The first three are engineering and legal hygiene; the last two are about external proof.

Hole 1: closed commercial proposals leaked through sitemap.xml

Our personalized commercial proposal pages sat behind a client-side JavaScript password. It looked closed. But Tilda had a global tag in its settings, <meta name="robots" content="index, follow">, and every proposal URL landed in sitemap.xml automatically.

Googlebot does not run a client-side JS password pop-up. It downloaded the raw HTML with all the tables. As a result, when a procurement manager asked «how much do Khamidun's services cost?», Gemini served up the full price list from a closed proposal for Technonicol: 300 000, 790 000 and 1 280 000 rubles.

A JavaScript password stops a casual visitor, but it does not protect against a search bot at all.

Hole 2: the AlpinaGPT case study on the «Telo Hudelo» domain

We still had an old abandoned domain, telohudelo.online, «Telo Hudelo» («Body Slimmed Down»). Its DNS A record pointed at our server's IP, and the Nginx config had no strict check of the Host header. The server dutifully served a full copy of the news portal news.hamidun.com at that address.

Google indexed the duplicate. Right in the middle of a serious B2B dossier, Gemini cited the source telohudelo.online/en/cases/alpinagpt-corporate-ai six times. Imagine the face of a bank security officer reading a corporate AI case study on the «Telo Hudelo» domain. I imagined it and went to fix it.

Hole 3: two OGRNIP numbers and a wedding code in the register extract

Under my INN 695007138610, counterparty aggregators (T-Bank Business, Rusprofile, Chekko) show two records:

1. The old sole proprietorship from 2015, OGRNIP 315695200002423. I closed it as planned in June 2023.
2. The active sole proprietorship dated 20 March 2024, OGRNIP 324690000014242, main OKVED code 62.01, computer software development.

The old record had built up inbound links and ranked above the new one, and each model stumbled over this in its own way. Claude put the closed OGRNIP in its report. Perplexity Pro noticed the date gap, 2015 versus 2024, and flagged a discrepancy in the registration details. ChatGPT Search did not find the sole proprietorship at all: its Bing engine indexes Russian EGRIP databases poorly, and on our site the details were buried deep in the offer agreement.

Gemini turned out to be the most thorough. It scrolled through the EGRIP extract down to 12 additional codes and pulled out the archived OKVED 93.29.3, «Organization of weddings, anniversaries and musical accompaniment». It then carefully warned the client about a «non-core line of business code». I do not run weddings, and I had completely forgotten that the code existed.

Hole 4: our own landing pages do not count as proof

Perplexity Pro found and unconditionally credited our three-month AI hackathon at LAB Industries: 113 participants, 10+ teams, 6 of which took AI agents all the way to real production. The reason is simple: a detailed article about the hackathon appeared on VC.ru. For all our other projects, Perplexity added a hard caveat:

«The information is posted on the speaker's own resources without independent verification».

We have 45 corporate projects behind us, including the «AI Ambassadors» programs in Drom (Farpost) and «Keiteringburg». But to an AI model, a case study without an article on an external platform or open engineering evidence carries half the weight. Along the way, Perplexity and Gemini also pulled from the archives my student profile as a physics and computer science tutor on profi.ru from 2014. I am not ashamed of it, but it looks amusing in the dossier of a corporate AI contractor.

Hole 5: in blind search, rankings overshadow us

I ran the last test without my surname and asked the AI models as a cold client would: «who should I call to run an AI workshop?». Gemini put me among the top three practitioners in Russia. ChatGPT Search, Perplexity and Claude built their answers around 9 ranking and roundup articles by others on VC.ru, TenChat and DTF.

The conclusion is unpleasant: on a category query, an AI model retells other people's lists. If you are not on the lists, you are not in the answer.

How did we close the technical holes in one day?

In a single working day we killed the mirror on the abandoned domain, cleaned up the sitemap, published reference company details in machine-readable form, and filed an application with the Federal Tax Service (FNS) to remove the extra OKVED code. Below is a step-by-step playbook you can repeat for your own business.

Step 1: killed the telohudelo.online mirror in Nginx

We added a strict Host header check to the Nginx config:

if ($host !~* ^(news\.hamidun\.com)$) { return 410; }

We also added the header X-Robots-Tag: noindex, noarchive. The 410 code tells Google the page is permanently gone, so the phantom domain drops out of the index without lengthy rechecks.

Step 2: cleaned up sitemap.xml and hid the proposals

We removed the global index, follow from the Tilda header. All 15 service pages and draft proposals were switched to server-side noindex. The sitemap.xml file slimmed down from 49 to 28 clean public pages, the only thing in this story that actually slimmed down.

The rule I now repeat to the team: whatever is closed with a JavaScript password is open to a search engine. Closing has to happen on the server.

Step 3: published company details for bots and Bing

We put the reference summary in three places at once: the open footer of hamidun.com, JSON-LD Schema.org markup (Person + Organization + LegalService) and the machine-readable file /llms.txt. The text is identical everywhere:

• Sole proprietor Khamidun Zhemal Uldovich, INN 695007138610;
• status: active individual entrepreneur;
• active OGRNIP 324690000014242 dated 20.03.2024;
• main OKVED 62.01, computer software development;
• previous OGRNIP 315695200002423 (2015–2023) closed as planned at the entrepreneur's own decision, with no debts or legal disputes.

An AI model that sees two OGRNIP numbers looks for an explanation. With no explanation available, it either invents one or writes about a discrepancy. We gave it a ready answer.

At the same time, we opened robots.txt on academy.hamidun.com to AI search agents: OAI-SearchBot, GPTBot, ClaudeBot, PerplexityBot, Google-Extended.

Step 4: cleaned up EGRIP and old profiles

Through the FNS personal account, we submitted an application on form R24001 to remove the archived code 93.29.3. We updated old profiles on aggregators so that the active sole proprietorship's record stops losing to the closed one.

What are we doing about the holes we cannot close in a day?

Holes 4 and 5 cannot be fixed in a day: AI models trust external platforms more than our landing pages, and on category queries they retell other people's rankings. The cure is publications in which someone other than us confirms the case. The LAB Industries hackathon showed the mechanics: one detailed article on VC.ru, and Perplexity credited the project without caveats. So the plan is as follows:

1. Take corporate case studies to external platforms, with numbers and engineering detail.
2. Ask clients to confirm case studies on their own resources, where an NDA allows it.
3. Once a quarter, repeat the interrogation of six models with the same script and compare the answers.

This article is also part of the treatment. AI models now have an external source stating that they made up everything about Amazon, esoterica and Dubai.

What should you check about your own business today?

The five checks from our audit can be done in one evening, and they show exactly what a procurement manager will see through an AI model:

1. Ask ChatGPT, Gemini and Perplexity who you are with search turned off. Write down everything they invent: this is how your business looks to a security officer in a closed corporate perimeter.
2. Turn search on and ask «how much do the services cost» and «check the sole proprietorship». If closed prices surface, look for the leak in sitemap.xml and meta tags.
3. Go through all your old domains: where do the A records point, and does the server return content for a foreign Host.
4. Open the full EGRIP or EGRUL extract, including additional OKVED codes. The AI model will read it to the end, even if you never did.
5. Put your company details in the footer, Schema.org and /llms.txt, and explain the history of closed legal entities yourself before the model does it for you.

The procurement manager asks the AI model before they ask me.

So I start answering in advance: in the site's code, in the EGRIP extract and on other people's platforms.