Why Do AI Models Hallucinate Competitors That Do Not Exist?
AI models hallucinate competitors that do not exist because they generate plausible text, and plausible company names are easy to generate. When a model has little real evidence about a market, it can invent a name that sounds right, blend two real companies into one, or keep naming a business that has closed. Researchers at OpenAI argued in 2025 that language models hallucinate largely because training and evaluation reward guessing over admitting uncertainty. A confident list of five competitors scores better than "I don't know". The risk rises where real evidence is thin, as in niche, local and non-English markets, and in answers given from memory rather than from a live search. They mislead buyers, and they distort AI visibility reports that count every name as a real competitor.
Why does this happen?
- Guessing is rewarded. OpenAI researchers describe hallucinations as errors that persist because models are trained and scored in ways that favour a confident answer over an admission of uncertainty.
- Answers expect lists. Questions like "top 10 providers in [town]" invite a full list even when the model knows only three.
- Sparse evidence. In small markets, specialist niches or less common languages, the model has seen few real names and fills gaps with invented ones.
- Name blending. Two similarly named companies, or a company and its product, merge into a single entity that does not exist.
- Outdated knowledge. Companies that closed, merged or rebranded keep appearing when the model answers from training data.
- No search. Answers generated without retrieval cannot be checked against real sources.
Where are phantom competitors most likely?
| Situation | Why the risk is higher |
|---|---|
| Local services in small towns | Few sources name local businesses, so models improvise |
| Niche B2B categories | Few comparison articles or directories list the real players |
| Non-English markets | Less training data and fewer indexed sources in the language |
| Fast-changing markets | New entrants and closures outpace the model's knowledge |
| Questions answered from memory | No search step to ground the names |
When an assistant searches the web and cites sources, invented names are less common. They can still occur when the model misreads or combines what it retrieved.
How do you spot a hallucinated competitor?
For every competitor an assistant names:
- Check it has a website that matches the description.
- Check official records, such as a company registry or licensing body.
- Look for independent traces: reviews, directory listings, press coverage.
- Check the cited source, if there is one, and whether it actually mentions the company.
- Rerun the question with search enabled. Names that disappear once the assistant searches are likely inventions.
- Look for blends, where the name combines parts of two real companies or a company and a product.
Why does this matter for businesses?
- Buyers waste time searching for firms that do not exist, and may lose trust in the answer and in the businesses named alongside them.
- Your brand may be blended with a phantom, picking up invented features or prices.
- Visibility reports get distorted. A tool that counts every name in an answer as a competitor inflates the field and lowers your share of voice. Some "competitors" in a report may be hallucinations.
- Real competitors are displaced. Phantom names take places in a short answer that real businesses, including yours, could have filled.
How assistants mix up real companies is covered in why AI mixes up a brand's features, reviews or case studies with a competitor's.
What can you do about it?
You cannot stop a model inventing names, but you can reduce the gaps it fills:
- Make the real market visible. Publish fair comparison or "alternatives" pages that name the real providers in your category and area.
- Strengthen category directories. Make sure you, and ideally your real competitors, are listed on the directories and review sites assistants cite for your category.
- Be explicit about your own identity and category, so a phantom does not absorb your details.
- Report clear errors through the assistant's feedback option, especially where a phantom is described with your information.
- Clean your tracking data. Keep a verified list of real competitors, flag unknown names for checking, and exclude confirmed phantoms from share-of-voice calculations.
How AI Overviews invent competitors and business details is covered in why Google AI Overviews hallucinate prices, competitors or business details.
What might an illustrative case look like?
Consider an illustrative specialist translation agency for medical devices. When its team asks assistants for "medical device translation agencies in Ireland", two of the five names returned do not exist. One blends the names of two real agencies, and another is a plausible invention with no website. Its visibility tool had been counting both as competitors, lowering the agency's share of voice from about 30% to 20%. The agency cleans its competitor list, and publishes a page comparing the real options for medical device translation in Ireland. It also gets listed on two industry association directories. Over the next quarter, phantom names appear less often in answers that search the web, and the agency is named in more of them.
What do people ask about hallucinated competitors?
Do all AI assistants invent companies?
All can, under some conditions. Answers based on live search and cited sources are less prone to it than answers from memory.
Should we contact the AI company about a fake competitor?
Use the feedback option on the answer. Reports are most useful when a phantom is harming customers or confusing your brand.
Can a hallucinated competitor be a real company elsewhere?
Yes. Sometimes the name belongs to a real business in another country or industry, placed in the wrong market.
How often should we validate our competitor list?
Each time you review tracking results, and whenever an unfamiliar name appears frequently.
Published by The Kovmere Team · updated Sep 29, 2026 · More resources