Uber Eats still appears as a food delivery option in 94% of Claude Haiku 4.5 answers and 86% of GPT-5 nano answers. Gemini 3.1 Flash-Lite did not list it. Across the full sample, at least one brand that left Ukraine, changed its name or whose legal entity is under Ukrainian NSDC sanctions appeared as a list item in 13.4% of Claude answers, 10.2% of GPT answers and 1.4% of Gemini answers. These are results from nine waves between June and October 2026. We tested API models without web search, not the consumer ChatGPT, Claude or Gemini apps. Why does AI recommend closed businesses if all three models were trained after these events? Outdated brand information in AI answers is not explained by the knowledge cutoff alone.
How we tested the models
Every two weeks, VYDAI sent 50 Ukrainian prompts in each of 49 categories to three API models: Claude Haiku 4.5, GPT-5 nano and Gemini 3.1 Flash-Lite. This article covers waves 3 to 11, from June 14 to October 6, 2026. The sample contains 65,544 successful answers: 22,045 from Claude, 21,449 from GPT and 22,050 from Gemini. The system instruction asked each model to name 7 to 12 companies and allowed a brief freshness caveat at the end.
We counted a brand when its name started a bullet or numbered recommendation. We checked the answer text because the automatic brand extractor missed some names and merged others. A passing mention did not count. For example, telling someone to read reviews on Yandex Maps is not a recommendation to use a Yandex service. Company statuses were checked on October 14, 2026 against dated company statements, decisions of Ukraine's National Security and Defense Council (NSDC), National Bank of Ukraine (NBU) records and media reports. For sanctions, the legal entity named in the decision matters, not just a similar brand name.
| Test detail | Scope |
|---|---|
| Models | Claude Haiku 4.5, GPT-5 nano, Gemini 3.1 Flash-Lite |
| Prompt language and categories | Ukrainian prompts in 49 categories |
| Prompts | 50 per category every two weeks |
| Period | Waves 3 to 11, June 14 to October 6, 2026 |
| Successful answers | 65,544 across the three models |
| Web search | Off |
| Main measure | Brand begins a list item in the answer text |
Outdated recommendations in AI: what 65,544 answers showed
Claude had the largest share of answers with at least one marked brand as a list item. Gemini had the smallest. The pattern stayed within a narrow range across all nine waves: Claude ranged from 12.6% to 13.8%, GPT from 9.5% to 10.8%, and Gemini from 1.1% to 1.9%.
At least one such brand appeared in 5,441 of 65,544 answers, or 8.3%. The largest differences appeared by category. Claude and GPT often named Uber Eats and Delivery Club in food delivery lists, while Gemini did not include either name in its lists.
Which ghost brands appear most often
These businesses have different statuses. Some services stopped operating in Ukraine, some were renamed, and some specific legal entities appear in Ukrainian sanctions decisions. The phrase “left Ukraine” does not describe every row.
| Brand in answers | What happened and when | Claude | GPT | Gemini |
|---|---|---|---|---|
| Uber Eats | Stopped operating in Ukraine on June 3, 2020 (Ukrinform English report) | 94% | 86% | 0% |
| Delivery Club | Yandex acquired the service from VK in September 2022 (Interfax); this study does not assess its Ukrainian operations | 59% | 66% | 0% |
| Lamoda | Suspended its Ukrainian operations and order taking on February 25, 2022 (Interfax-Ukraine) | 68% | 52% | 25% |
| 1C | NSDC sanctions on LLC 1C (Moscow) in 2017 (Decree No. 133/2017); the 2023 decision names CJSC “1C Joint Stock Company” (Moscow) (NSDC decision of April 15, 2023) | 88% | 11% | 0% |
| LLC Bitrix24 | The 2023 NSDC decision names LLC Bitrix24; its cloud service stopped working for Ukrainian users on June 1, 2023 (NSDC decision, DOU report) | 33% | 52% | 7% |
| Alfa-Bank Ukraine | Shareholders approved the Sense Bank name in August 2022 (Ekonomichna Pravda) | 24% | 77% | 0% |
| Sportmaster | NSDC sanctions of February 19, 2021 on Sportmaster Operations Pte. Ltd. (Singapore), owner of LLC Sportmaster-Ukraine (Interfax-Ukraine) | 26% | 44% | 26% |
| Skillbox | The 2023 NSDC decision names LLC Skillbox (decision of April 15, 2023) | 51% | 1% | 0% |
| Skyeng | The 2023 NSDC decision names LLC Skyeng and ANO DPO Skyeng (decision of April 15, 2023) | 16% | 22% | 0% |
| Ukrsotsbank | Merged with Alfa-Bank in October 2019 and lost its license (Ukrinform) | 8% | 0% | 0% |
Percentages show the share of answers in each brand's main category, roughly 450 answers per model, where its name appeared as a list item. Yandex is not in this table because its services appeared across many topics; we discuss those results below. Sanctions are not extended to other companies with similar names.
In food delivery, Claude listed Uber Eats in 94% of answers and GPT in 86%. Both also listed Delivery Club in 59% and 66% of answers. Interfax reported that Yandex acquired Delivery Club from VK in September 2022. That source describes an ownership change; we do not assess whether the service ever operated in Ukraine. Gemini listed neither brand. So, is Uber Eats available in Ukraine? No. Uber said the delivery service would stop in Ukraine on June 3, 2020, while Uber rides would continue.
In banking, GPT listed Alfa-Bank in 77% of answers. Yet Alfa-Bank Ukraine was renamed Sense Bank back in August 2022: shareholders approved the new name. The NBU's 2023 annual report says the bank became state-owned in July 2023. Gemini, by contrast, listed Sense Bank in 77% of its banking answers. Across all topics, GPT named Alfa in 379 answers but mentioned Sense in only three of them. Ukrsotsbank merged with Alfa-Bank in October 2019. Claude still listed it in 8% of bank answers and called it a “state bank” in several fragments.
In accounting, Claude listed 1C in 88% of answers about accounting services. Seven of the 30 fragments that mentioned 1C were passing mentions of accounting software: Gemini mentioned BAS/1C in five answers and GPT did so in two, with wording such as “check whether the company works with BAS/1C.” Gemini did not list 1C in this category at all, zero of 450 answers, though it mentioned the software in 126 answers. Claude described 1C:Accounting as “developed in Ukraine,” although both sanctions decisions name Moscow-based 1C legal entities.
The Bitrix24 entity and 1C-Bitrix must be kept separate. The 2023 sanctions decision named LLC Bitrix24 in Kaliningrad; the developer 1C-Bitrix is another company, and this article makes no sanctions claim about 1C-Bitrix. Claude called Bitrix24 a “Ukrainian CRM” and a “domestic CRM.” DOU reported that its cloud service became unavailable to Ukrainian users on June 1, 2023.
In education, Claude listed Skillbox in 51% of answers and Skyeng in 16%. In 81 of 225 fragments where Claude named Skillbox, it itself added “Russian platform” and still left the name in its list. Annex 2 to the NSDC decision of April 15, 2023, put into effect by Decree No. 227/2023, names LLC Skillbox, LLC Skyeng and ANO DPO Skyeng (all Moscow) and LLC Bitrix24 (Kaliningrad), each for ten years. The legal entities, rather than similar brand names, matter in a sanctions check.
In retail, Claude listed Lamoda in 68% of answers, although it suspended Ukrainian order taking in February 2022. Gemini listed Lamoda in 25% of fashion answers. Sportmaster appeared in 26% of Claude lists, 44% of GPT lists and 26% of Gemini lists. These are the only two ghost brands that Gemini listed noticeably. The 2021 sanctions named Singapore-based Sportmaster Operations Pte. Ltd., owner of LLC Sportmaster-Ukraine. The Ukrainian chain later changed its name to Athletics and announced a gradual closure in August 2025, according to Minfin.
The NSDC decision of April 28, 2017, put into effect by Decree No. 133/2017, names LLC Yandex (Moscow) and LLC Yandex.Ukraine; the decision of April 15, 2023 names LLC Yandex again. Claude mentioned Yandex services in 638 answers across 48 of 49 categories. In 336 of those answers, it referred to Yandex Maps, mostly as a suggestion to read reviews there. As a list item, a Yandex service or brand appeared in 246 answers across 28 categories. GPT mentioned Yandex in 66 answers across eight categories; Gemini did not mention it.
Did the model recommend a brand or just mention its past use?
We manually read 262 answer fragments about ten main brands. In 251 cases, or 96%, the brand appeared as a recommendation. None of the fragments about Uber Eats directly said that it had left Ukraine. For Delivery Club, one of 25 fragments said in the list item that it had been used in the past. For Lamoda, one of 25 list items noted that it was unavailable or limited in Ukraine, but kept it in the list.
| Brand | Fragments read | Listed as a recommendation | Caveat in list item | Not a recommendation |
|---|---|---|---|---|
| Uber Eats | 28 | 28 | 0 | 0 |
| Delivery Club | 25 | 25 | 1 | 0 |
| 1C | 30 | 23 | 0 | 7 |
| Bitrix24 | 25 | 25 | 0 | 0 |
| Lamoda | 25 | 25 | 1 | 0 |
| Sportmaster | 25 | 25 | 0 | 0 |
| Skyeng | 25 | 25 | 1 | 0 |
| Skillbox | 30 | 30 | 0 | 0 |
| Ukrsotsbank | 25 | 21 | 0 | 4 |
| Yandex | 26 | 9 | 3 | 14 |
The system instruction allowed a general freshness caveat. Claude added one to 230 of 424 answers that named Uber Eats; GPT did so in 369 of 388. Only one Claude answer specifically said Uber Eats no longer worked in Ukraine. GPT had no such specific correction. A general “check that this is current” does not correct a recommendation unless the answer identifies the problem.
Where ghost brands take space in recommendations
Food delivery had the highest share of answers with a marked brand: 96.9% for Claude and 92.4% for GPT. Gemini's figure in this category was zero. Clothing, SaaS, banking and accounting also showed notable shares.
| Category | All three models | Claude | GPT | Gemini |
|---|---|---|---|---|
| Food delivery | 63.1% | 96.9% | 92.4% | 0% |
| Clothing, footwear, accessories | 52.1% | 69.6% | 55.8% | 30.9% |
| SaaS and software | 38.1% | 51.1% | 55.7% | 7.6% |
| Banks | 37.2% | 32.2% | 79.1% | 0.2% |
| Accounting services | 33.9% | 88.9% | 12.8% | 0% |
| Sport, leisure and hobbies | 33.3% | 30.2% | 43.7% | 26.2% |
| Education | 27.9% | 60.7% | 22.9% | 0% |
| Marketplaces | 22.6% | 27.6% | 40.1% | 0.2% |
| Restaurants and cafes | 16.4% | 20% | 29.5% | 0% |
| Children's products | 12.1% | 8.2% | 28.2% | 0% |
The other names in the same answers varied by model. In food delivery, Claude named Wolt in 78% of answers and Bolt Food in 24%, alongside Uber Eats in 94%. GPT most often listed Glovo at 87%, Uber Eats at 86% and Wolt at 62%. Gemini's frequent list items were Glovo at 88% and Bolt Food at 86%. In fashion, Claude listed Rozetka at 68%, Lamoda at 65% and Prom.ua at 51%; GPT listed Kasta at 69%, Rozetka at 63%, Answear at 62% and Lamoda at 52%.
These shares show which names appeared in the same lists. They do not measure sales lost by a competitor. We did not measure clicks, leads or orders. A company can use the results to check its own presence in comparable questions and answers, not to calculate a commercial impact.
Why AI recommends businesses that left Ukraine
The LLM knowledge cutoff dates for all three models came after the main events in the table. The answer “the model simply does not know” does not fit these cases: the Ukrsotsbank merger, Uber Eats exit and Alfa-Bank rename all happened before those dates. Claude had the latest training data date among these models and the highest overall share of answers with older brands. Gemini's cutoff was earlier than Claude's, but its share was the lowest.
An LLM knowledge cutoff describes the stated boundary of training data; it does not show which older sources shaped a particular answer.
| Model | Official knowledge date | Answers with at least one old brand |
|---|---|---|
| GPT-5 nano | “May 31, 2024 knowledge cutoff” (OpenAI model page) | 10.2% |
| Gemini 3.1 Flash-Lite | “Gemini 3 models have a knowledge cutoff of January 2025” (Google model documentation) | 1.4% |
| Claude Haiku 4.5 | “trained on data up until July 2025” (Anthropic help page) | 13.4% |
One plausible explanation is that years of older public pages mentioning a brand can outweigh fewer recent pages about its rename or closure. That is a hypothesis about a possible mechanism; this sample did not test it. The system also asked models for 7 to 12 companies, a format that encourages longer lists of familiar names. We did not measure how many errors that instruction caused. The differences in this sample show that a cutoff date alone does not predict the answer. Our study of how often the three models agree on brands across 49 categories shows how differently they build lists in general. Read our comparison of how AI systems choose brands for a broader view of model differences.
Does web search fix outdated AI answers?
OpenAI says its web search tool “allows models to access up-to-date information from the internet and provide answers with sourced citations” in its web search documentation. Anthropic says its Web Search Tool gives Claude current content beyond its knowledge cutoff. Google explains how Grounding with Google Search retrieves, processes and cites results.
This study had search turned off, so it does not compare answers with search against answers without it. In a separate VYDAI study, we checked ChatGPT's smartphone brand recommendations in Ukraine with web search enabled, but that work answered a different question and did not measure brands that had left the market. Search can bring a current page into an answer's context, but readers still need to open the source and check its date and what service it covers.
How a brand can update its information for AI models
A company cannot change a model version that has already been released. It can make current facts about itself consistent and easy for future model versions and search-based answers to find. Start by checking what a person actually sees in an answer, then update the sources the company controls.
| Situation | What to check or update |
|---|---|
| A service that left still takes a place in the list | Get into current "what to use instead of X" coverage. After Bitrix24 left, Forbes Ukraine (May 2023) compiled lists of Ukrainian and Western CRM replacements. A comparison page on your own site that states when the service left, with a source, adds a text with the correct fact to the web. |
| The brand changed its name | State the current and previous names and the date of the change on the About page. Align official profiles and catalogues. |
| A service is no longer available | State its status and date on the company site. Update product pages, help pages and service profiles the company controls. |
| Facts are confused with another company | Check the legal name, domain, address and official profiles. Use consistent wording that points to verified company pages. |
Wikipedia, Wikidata, catalogues, review sites and media can all contribute to the public record. Correct factual errors there only within each resource's rules. An About page with a rename date cannot guarantee that old references disappear from model answers. Our article on what to do when AI says something wrong about your brand covers a company-level response. For how websites, PR, catalogues and reviews contribute, see which sources AI trusts.
Here is how to update your brand information for AI models so the effect can be measured: first record the same category prompts with the same language, model and search setting. After updates, repeat them and note the date, full answer, brand position and cited sources. This makes it possible to see whether later measurements changed.
How to check if a company is sanctioned in Ukraine and whether it operates
If an AI suggests an unfamiliar or outdated option, check it before signing up or paying. Start with the specific service or legal entity, not only the brand name.
| Check | What to do | What it tells you |
|---|---|---|
| Website and domain | Open the address, compare it with official brand pages and look for a dated name-change notice | Whether the site is available at that address; this alone does not prove the service operates |
| Recent news | Find a dated company statement or reliable report about closure, sale or rebrand | The event and its date |
| Sanctions | Search for the legal entity in the State Register of Sanctions and open the relevant decision | Whether a sanctions record applies to that entity |
| Bank | Check the current name and licence in the NBU list of banks | Whether it appears in the official bank list |
Search the full legal name in the register. A sanctions decision naming LLC Bitrix24 is not a decision about 1C-Bitrix. A bank's historic name also does not mean it still operates under that name. Our DNS check on October 14, 2026 found that bitrix24.ua, alfabank.ua, sportmaster.ua, skillbox.ua, skyeng.ua, delivery-club.ua and ukrsocbank.com.ua did not exist in DNS. A link that does not open is itself a signal to check further. See our study of AI hallucinated company websites for the related domain problem.
Study limitations
This measures three API models without web search. It does not describe every answer in the consumer ChatGPT, Claude or Gemini apps, where the model, system instructions, search tools, personalization and updates may differ. The data covers Ukrainian prompts in 49 categories between June 14 and October 6, 2026.
We counted a brand when it began a list item in the answer text. This measure does not track clicks, orders, lost revenue or market share. Manual reading of 262 fragments helps explain context, but it is not a full review of every answer.
The statuses in the table are different kinds of facts. For sanctions, the named legal entity matters. For a renamed company, the former name may remain in the bank's history. For a service that stopped operating, the country and specific service matter. This study did not assess whether web search reduces these recommendations; that requires a separate test with matching prompts in both modes.
Frequently Asked Questions
Why does ChatGPT give outdated information?
An API model can answer from information formed before a particular date. OpenAI lists a May 31, 2024 knowledge cutoff for GPT-5 nano. In our test without search, GPT listed Alfa-Bank in 77% of banking answers even though shareholders approved the Sense Bank name in August 2022, before that date. So the cutoff date alone does not explain the error.
Is ChatGPT recommending Russian services?
GPT-5 nano listed brands whose legal entities are named in Ukrainian NSDC sanctions decisions: Bitrix24 (LLC Bitrix24) in 52% of SaaS answers, Skyeng (LLC Skyeng) in 22% of education answers and 1C (LLC 1C, CJSC “1C Joint Stock Company”) in 11% of accounting answers. Their statuses differ. Uber Eats stopped operating in Ukraine, Sense Bank is the new name of Alfa-Bank Ukraine, and sanctions apply to specific legal entities. These cases should not be treated as one category.
Is Uber Eats available in Ukraine?
No. Uber said Uber Eats would stop operating in Ukraine on June 3, 2020, while Uber rides would remain available.
Does web search fix outdated AI answers?
It can give a model access to current pages and citations. This study did not compare the frequency of outdated recommendations with search on and off.
How can I update my brand information for AI models?
Date a name or service-status change on your own pages, align public profiles and check the same prompts over time. This helps keep sources accurate, but it does not switch an already released model to new knowledge.
Where does AI get outdated information?
We did not measure the sources behind these answers. Older public references and the prevalence of a former name are possible explanations, but a specific source can only be verified in an answer that provides citations or through a separate source analysis.
How we do it in VYDAI
VYDAI's public ratings show which brands three models list in each Ukrainian category. In the monitoring workspace, teams can review answers by model and date and compare brands that appear alongside their own. VYDAI does not automatically label a brand as a ghost, so company status still needs to be checked against current sources.
You can view a VYDAI demo or register a project. If an old competitor name appears in an answer, save the prompt, date and full response. That gives the next review a concrete record to work from.
