Insight · Published October 2026
By Marin Tintchev, Founder, GeoSynapse & China GEO EU
What AI Answers When It Has Never Heard of Your Company
In September 2026 we asked five AI platforms about our own company before its website existed. They rarely said they did not know. They read the name as something else, gave a different wrong answer on a different day, and offered confident lists that left us out.
When a company is new, the first thing an AI system says about it is usually its first public impression, and nobody has seen it. We looked at ours. Between 11 and 15 September 2026, before the company's own website had gone live, we put six fixed questions to five AI platforms (ChatGPT, Claude.ai, Gemini, Perplexity and Grok), each on three separate days. Four of the platforms ran from a clean test account; Perplexity ran without an account. This piece covers Western platforms only. Chinese platforms are the subject of a separate test, and nothing here should be read as a result for them.
We found no outright invention. What we found was more ordinary, and probably more common: a name read as something else, facts that were real but out of date, and answers that changed from one day to the next.
What happened when we asked what the company is?
FACT
Across fifteen platform-days of asking for the company by name, one answer identified it correctly: Gemini, on the first day. On the second and third day the same platform, with the same question, did not.
FACT
The rest of the answers read the name as something else. Across the platforms we saw at least five different meanings: international cooperation on Earth observation, a trade agreement on geographical indications, China–EU geoeconomic relations, geolocation rules in an EU deforestation regulation, and an engineering company. The acronym GEO, which in our field stands for generative engine optimization, has many older meanings, and the platforms chose among them.
FACT
On at least one run, ChatGPT and Perplexity did the right thing: they listed the possible meanings and asked what context was meant.
Did the wrong answer at least stay the same?
FACT
No. Grok anchored on Earth observation on the first day, on the geographical indications agreement on the second, and on Earth observation again on the third. Gemini's three days produced the correct answer, then two different wrong ones. ChatGPT gave a detailed biography of the founder in one mode and said it found nothing reliable in the other, on the same day.
FACT
In a question about leading providers in this field, ChatGPT and Claude.ai read GEO in the third round as "employer of record" and answered with payroll and staffing outsourcers. In the first two rounds they had answered the same question about the right field. Gemini, Perplexity and Grok read it as AI visibility in all nine of their runs.
INFERENCE
The failure is not a stable wrong belief that one correction would fix. It is a choice among readings that is made again each time, which is why we repeat every question on several days before saying anything about an answer (see [5]).
What about facts that were true?
FACT
When we asked about the founder, three of the five platforms (ChatGPT, Claude.ai, Perplexity) returned a former role as if it were current. It was real, and it appeared on his public professional profile; it was simply no longer where he works. One answer also placed that former employer in the wrong country. Grok returned nothing from the last several years at all, only an older lecture and an old social-media bio. Gemini returned the current picture on all three days, and cited a first-party page.
INFERENCE
These are not inventions. They are outdated but real public signals ranking above the current ones, and the most likely shared source was one public profile whose most recent entry did not yet describe the current work. The evidence for a single shared source is circumstantial.
Did the platforms say they knew no providers?
FACT
On the three questions about providers in this field, the company was named in none of the 45 platform-day answers, and the platforms named more than 80 different organizations between them. On three of the platforms the lists were largely different each time.
INFERENCE
A system asked for providers tends to produce a list rather than report that nothing fits. A list of that kind should not be read as a stable view of the market (see [5]).
Is this already known?
FACT
Parts of it are. A 2025 paper from OpenAI and Georgia Tech argues that facts seen rarely, or once, in training data put a floor under how often a model errs, and that common benchmark scoring gives no credit for saying "I don't know," so guessing is rewarded[1]. A 2024 study found that when a name can mean several things, models often pick the wrong one and apply what they know inconsistently[2]. A benchmark built from Wikipedia disambiguation pages found that when several entities share a name, models often answer for only one or merge facts across them[3]. In a different task, identifying news articles, a 2025 Tow Center study found that search tools answered more than 60 percent of queries wrongly, mostly with confidence, and rarely declined[4].
INFERENCE
Our observations fit this literature. They do not add a new mechanism. What they add is a concrete case: what a new company in a narrow B2B field looks like to five platforms before it has sources of its own. The result is a pre-launch baseline, not a verdict on any later change.
What can any company take from this?
INFERENCE
Four checks travel well. Ask each major platform about your company by name, on several days, and keep the answers. Look at what your name or acronym could mean to someone who has not heard of you. Look at your public profiles for the oldest facts that still outrank the newest ones. And treat a single answer, good or bad, as one sample, not as the state of things.
References
- Kalai, Nachum, Vempala, Zhang, "Why Language Models Hallucinate," arXiv:2509.04664, 4 September 2025 — arxiv.org.
- Sedova, Litschko, Frassinelli, Roth, Plank, "To Know or Not To Know? Analyzing Self-Consistency of Large Language Models under Ambiguity," Findings of EMNLP 2024 — eprints.cs.univie.ac.at.
- Lee, Ye, Choi, "AmbigDocs: Reasoning across Documents on Different Entities under the Same Name," arXiv:2404.12447 — arxiv.org.
- Jaźwińska and Chandrasekar, "AI Search Has A Citation Problem," Tow Center for Digital Journalism, Columbia Journalism Review, 6 March 2025 — cjr.org (task: identifying news articles, not entities).
- "Why AI Answers Change Between Runs," China GEO EU, September 2026 — chinageo.eu.
Frequently asked questions
Does this mean AI platforms invent facts about companies?
Not in what we observed. The answers we examined were misreadings, outdated real facts and inconsistent choices, not made-up claims. Other studies document invention in other settings.
Is this specific to China or to this field?
No. Name and acronym collisions, stale public profiles and run-to-run variation are general. Our case sits in a narrow B2B field with a crowded acronym.
Does China GEO EU offer to fix this for a specific company?
This piece describes a feature of the current AI-answer environment. It is not a service offering.
For why answers vary between runs, see "Why AI Answers Change Between Runs — And What That Means for Measuring Visibility" [5].
Related insights
Insight
Beyond B2B vs B2C: Does Your AI Visibility Strategy Fit Your Business?
B2B or B2C says who the transaction is with, not whose questions an AI system must answer. A four-layer way to look at AI visibility.
Research
From Stuttgart to Shenzhen: Before the First Call, Your Buyers May Already Have Asked AI About You
In Europe and in China, the buying committee turns to AI first. What differs is where it checks the answer. A review of Western and Chinese surveys.
Insight
We Grew a B2C Business on AI Visibility, from Scratch. How Does It Work in B2B?
A yacht business grew from scratch on AI visibility. We then ran the same method on a B2B company with no ads: baseline audit, foundations, re-audit.
About the author
Marin Tintchev is the founder of GeoSynapse and China GEO EU. He has 30+ years in ICT and international business across China and Europe, was educated at Tsinghua University, and is fluent in Mandarin.