March 29, 202610 min read

Research: How Often ChatGPT Gets Russian Brands Wrong

Analysis of ChatGPT hallucinations about Russian brands: incorrect prices, outdated information, brand confusion, non-existent products. Which niches suffer the most, comparison with other AI providers, and what to do about it.

Vladislav Puchkov

Founder of GEO Scout, GEO optimization expert

When a user asks ChatGPT "how much does plan X from bank Y cost," they expect an accurate answer. But the AI may cite a two-year-old price, mix up plans from two banks, or outright invent a non-existent product. The user will not know — the response looks confident and well-structured.

We systematically analyze responses from 9 AI providers about 716 Russian brands through GEO Scout and see the scale of the problem from the inside. This article is about which errors AI makes most often, which niches suffer more, and what brands can do about it.

Hallucination Typology: 5 Main Types

Type 1: Outdated Data (40% of all errors)

The most common type. AI delivers information that was accurate 6-18 months ago but is now outdated.

Examples from monitoring practice:

Reason: ChatGPT's training data has a cutoff, and fine-tuning cannot keep up with the Russian market. Financial terms, hosting prices, and course fees change faster than the model updates.

Type 2: Brand Confusion (20% of errors)

AI mixes information about two similar brands or attributes one brand's characteristics to another.

Common mix-ups:

• Ozon and ozone (the gas) — ChatGPT sometimes starts talking about the ozone layer
• Tochka (bank) and Tochka (other brands) — context mixing
• Skillbox and SkillFactory — course and platform confusion
• REG.RU and RU-CENTER — domain registrar service mix-ups
• Aviasales and Skyscanner — feature and coverage confusion

Type 3: Non-existent Features and Products (15% of errors)

AI confidently describes product capabilities that do not exist. This is a classic hallucination — the model "fills in" based on patterns.

Examples:

• Describing mobile app features that do not exist
• Mentioning integrations that are not supported
• Referring to offices in cities where the company has no presence
• Describing partnership programs that have not been launched

Type 4: Mixing Russian and International Context (15% of errors)

ChatGPT, trained predominantly on English-language data, sometimes substitutes Russian realities with international ones.

Examples:

• Mentioning PayPal as a payment method (unavailable in Russia)
• Recommending international services blocked in Russia (Booking.com, Airbnb)
• Prices in dollars instead of rubles
• Links to global product versions instead of Russian ones

Type 5: Incorrect Contact Information (10% of errors)

Addresses, phone numbers, business hours — information that changes frequently and that AI remembers imprecisely.

Error Frequency by Niche

Not all niches suffer from hallucinations equally. Data is based on systematic analysis of AI responses within GEO Scout monitoring.

Why FinTech Leads in Errors

Banking products are the most dynamic category. Deposit rates change weekly, loan terms update monthly, new products launch quarterly. ChatGPT physically cannot keep up with these changes.

Typical scenario: a user asks "which bank offers the best deposit rate." ChatGPT answers based on 6-12 month old data. The user goes to the bank, discovers different terms, and loses trust — not in AI, but in the bank ("one thing on the website, another in advertising").

Why Travel Is Second

Sanctions restrictions created a unique situation: many international services (Booking.com, Airbnb) are unavailable in Russia, but ChatGPT continues to recommend them. More about this — in the article "Blocked Services in AI".

Additionally, travel has high seasonality — routes, prices, and availability change every month. AI cannot account for current promotions, temporary restrictions, and seasonal changes.

AI Provider Comparison: Who Makes More Errors

ChatGPT is not the only AI that hallucinates. But the scale of the problem differs.

YandexGPT: Knows Better, but Not Without Flaws

YandexGPT (Yandex with Alice) shows fewer errors about Russian brands thanks to training on Russian-language data. But it has its own weaknesses: it may confuse details of regional brands and provide outdated information on rapidly changing products. Detailed comparison — in the article "YandexGPT vs ChatGPT: Difference in Recommendations".

ChatGPT: Great Structure, Bad Facts

The ChatGPT paradox: it delivers the most structured and convincing responses while making more factual errors. Users read a beautifully formatted response with tables and lists — and trust it even more. This makes ChatGPT hallucinations particularly dangerous.

Business Consequences

Direct Losses

• Lost customers: user believes the outdated price, goes to a competitor with "better terms"
• Reputational damage: AI attributes someone else's shortcomings to your brand
• Missed opportunities: AI does not know about new products and does not recommend them

Indirect Losses

• Distorted competitive analysis: competitors also suffer from hallucinations, but you do not know about it
• Incorrect customer expectations: user arrives with AI-provided information that does not match reality
• Lower NPS: customer believes the brand failed to deliver on promises, though the promises were made by AI

What to Do: A Strategy for Combating Hallucinations

1. Monitor AI Responses Daily

The first step is knowing what AI says about your brand right now. Daily monitoring through GEO Scout records all mentions, allowing you to identify errors quickly.

2. Create a "Single Source of Truth"

Publish current information (prices, terms, specs) on the most authoritative platforms possible:

• Official website with up-to-date structured markup (JSON-LD)
• Profiles on authoritative aggregators (vc.ru, Habr, Otzovik)
• Regularly updated FAQ pages
• Press releases for significant changes

3. Structure Data for AI

More about GEO site audit and technical optimization.

4. Create Content That Corrects Misconceptions

If AI systematically makes errors about a specific fact about your brand — create content that directly and clearly provides the correct information. An article "Current [brand] Plans in 2026" on an authoritative platform will eventually make it into training data.

5. Use the Command Center for Analysis

The Command Center in GEO Scout analyzes the content of AI responses about your brand, identifying error patterns and tracking their dynamics. This helps prioritize: which errors to fix first, which providers need special attention.

6. Prioritize Providers

Not all AI providers are equally important. If your audience primarily uses YandexGPT — focus on it. If B2B clients use ChatGPT and Claude — prioritize them. Provider data — in the comparative analysis.

How to Measure the Scale of the Problem for Your Brand

Before developing a strategy for combating hallucinations, you need to understand the scale of the problem specifically for your brand.

Step 1: Current Response Audit

Ask 20-30 typical questions about your brand to each AI provider. Record:

Step 2: Error Prioritization

Not all errors are equally critical. Prioritize:

• P0 (critical): incorrect prices, false negative statements, confusion with a competitor
• P1 (important): outdated data, missing key products
• P2 (desirable): imprecise wording, incomplete information

Step 3: Regular Monitoring

A one-time audit gives a snapshot. Regular monitoring gives dynamics. After corrective actions (content publication, markup updates), you need to track how quickly AI "picks up" the correct information.

Forecast: Will It Get Better

Hallucinations are a fundamental problem of language models, but the situation is gradually improving:

• Search-enabled AI (Perplexity, Google AI Mode) already show 5-14% errors — 2-3 times fewer than generative models
• The RAG approach (retrieval-augmented generation) is becoming standard — models increasingly verify facts before responding
• Training data is expanding — each new ChatGPT version knows the Russian market slightly better

But the problem will not disappear completely. Brands that actively manage their AI representation will always be in a better position than those that do not.

Checklist: Protecting Your Brand from AI Hallucinations

• Launch daily monitoring on GEO Scout
• Conduct an audit of current AI responses about the brand across all providers
• Compile a list of factual errors with priorities
• Update JSON-LD markup on the website (Organization, Product, FAQ)
• Publish current data (prices, terms) on 3-5 authoritative platforms
• Create FAQ content directly answering queries with errors
• Set up alerts for changes in AI responses through the Command Center
• Schedule weekly updates of key website information
• Check whether AI recommends blocked/outdated services instead of yours
• Repeat the audit after a month and compare with the baseline