Signals and trust
When AI gets your company wrong: brand hallucinations, and how to fix them
By Arnav Mukherjee, founder of TofuBofu · July 13, 2026
Last week we published a research pilot: we asked ChatGPT, Claude, Perplexity, and Gemini the same open question, who are the best managed IT providers in British Columbia, and then checked every company they named against Clutch's real directory and Google search. One result stopped me cold. Every single company Claude recommended, nine out of nine, could not be verified in either source.
Some of those nine may be real firms with thin web footprints. But that is the exact pattern hallucination produces: a confident, well-formatted shortlist with nothing underneath it. And it reaches real buyers, who have no way to tell. This post is about why that happens, what it looks like when it happens to your company, and the playbook that actually corrects it.
Why engines invent things about companies
A language model is a plausibility machine. Ask it a question and it produces the most statistically likely continuation, whether or not the underlying facts exist. When it knows a lot about a subject, plausible and true mostly coincide. When it knows little, plausible wins and true loses, and the model does not flag the difference. It answers a question about your company with the same confident tone whether it is drawing on fifty corroborating sources or zero.
This is precisely the problem retrieval was invented to fix. The original RAG research (Lewis et al., arXiv:2005.11401) showed that letting a model fetch real documents before answering produces more factual, specific answers than generating from memory alone. That is why search-grounded engines like Perplexity hallucinate less about companies: they look you up first. In our BC research, Perplexity was the most grounded engine we tested, and Claude, answering from memory, was the least. Same question, different plumbing, wildly different reliability.
The failure modes have a shape. An engine confuses you with a similarly named business. It describes the company you were two years ago, pre-rebrand, pre-pivot. It merges an acquisition wrongly: in our research, Gemini correctly noted that one BC provider had acquired two others, while other engines still listed the acquired firms as independent companies. Or it simply invents: a service you do not offer, an office you do not have, a price you never charged.
Finding out what AI believes about you
You cannot correct what you have not seen. The audit is straightforward, and worth doing by hand at least once. Ask each major engine the questions your buyers ask: best providers in your category and region, alternatives to your biggest competitor, and your company name directly. Then ask the direct fact-checks: what does this company do, where is it located, what does it cost, who are its competitors. Run each question more than once, because answers vary between runs, and one clean answer does not mean the next buyer gets the same one.
You are hunting for four things: facts that are wrong, facts that are stale, facts that belong to someone else, and absences, the questions where competitors get named and you do not. Log all of it per engine, because the fix differs by engine type. I wrote up the full method in how to measure AI visibility.
See what six AI engines actually say about your company
A free scan runs your buyers' questions across ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, and Bing Copilot, and shows you the real response text.
Run your free scanThe corroboration playbook
You cannot argue with a model. You can only change what it reads. Every fix below works the same way: it closes a gap the engine would otherwise fill with a guess.
1. Publish one canonical version of your facts
Your homepage and about page should state, in plain sentences, exactly what you do, for whom, where, and since when. Add Organization schema carrying the same facts: legal name, domain, locations, founding date. Engines resolve entities by matching features; give them unambiguous features to match.
2. Make every listing tell the same story
Directories, review profiles, LinkedIn, partner pages: same name, same description, same locations. After a rebrand or acquisition, stale listings are where engines pick up the old reality. In our BC research, two acquired companies still appeared as independent firms in AI answers because their old footprints were never cleaned up.
3. Build third-party corroboration
An engine trusts what independent sources agree on. Reviews on the platforms your industry uses (Clutch and G2 for services firms), real participation on Reddit, coverage and citations elsewhere: each one shrinks the gap the model can fill with fiction. This is the same off-site work that drives visibility, which is not a coincidence: corroboration is the cure for both invisibility and hallucination.
4. Fix the search-grounded engines first
Perplexity, Copilot, and Google AI Overviews re-read the live web, so source fixes can change their answers within weeks. Training-based knowledge in ChatGPT and Claude updates on model release cycles, so those corrections take months to fully land. Sequence your effort accordingly, and file feedback reports with the engine for anything defamatory in the meantime.
5. Re-check monthly, and after every change to your facts
Rebrands, acquisitions, pricing changes, and new locations are when engines blend old and new information. That transition window is where most damaging brand hallucinations are born, so schedule a check right after any of them.
The uncomfortable upside
Here is the reframe I keep coming back to. If engines guess when corroboration is thin, then every firm in your market with a thin footprint is currently represented by guesses, including the competitors ahead of you in the answers. The BC data showed real, well-reviewed firms invisible while unverifiable names got recommended. That is not a stable ranking, it is a vacuum. Corroborated facts beat plausible fiction the moment they exist, because retrieval prefers them. Being the best-documented company in your category is an available position, and in most services markets, nobody has taken it yet.
Frequently asked questions
Why does AI say false things about my company?
Because language models are built to produce plausible text, not verified text. When an engine has thin or conflicting information about your company, it does not say 'I am not sure.' It fills the gap with something statistically plausible: an old fact, a competitor's detail, or an invention. The weaker your corroboration across the sources the engine reads, the more room it has to guess.
How do I find out what AI engines are saying about my company?
Ask them what your buyers would ask, on every major engine, several times each, because answers vary between runs. Ask direct questions too: what does the engine know about your company, your pricing, your services, your locations. Note anything false, outdated, or borrowed from a competitor. A free TofuBofu scan automates this across six engines with buying-intent queries and shows you the actual response text.
Can I get an AI engine to correct false information about my brand?
Not by asking it to. Corrections come from changing what the engine reads. For search-grounded engines like Perplexity, Copilot, and Google AI Overviews, fixing the sources they retrieve, your site, directories, review profiles, third-party pages, changes answers within weeks. For training-based knowledge the timeline is model release cycles. OpenAI, Google, and the others also accept feedback reports on harmful inaccuracies, worth filing but not counting on.
What makes a company prone to AI hallucinations?
Thin, inconsistent, or conflicting public information. A generic name shared with other businesses, different descriptions of what you do across your site and directories, stale listings from a rebrand or acquisition, and little third-party coverage all leave gaps. Engines fill gaps. Companies with consistent, corroborated facts across many independent sources give the model nothing to invent.
What is the fastest fix when AI confuses my company with another one?
Strengthen the signals that disambiguate you: a clear, specific description of what you do repeated consistently on your homepage and about page, Organization schema with your exact name, domain, and locations, and matching profiles on the directories your industry uses. If the confusion is with a similarly named business, add the differentiating facts (geography, industry, founding) prominently, because those are the features an engine uses to tell entities apart.
Do AI engines hallucinate recommendations, not just facts?
Yes. In our own published research we asked four engines for the best managed IT providers in British Columbia. Every company Claude named, nine of nine, could not be verified on the region's Clutch directory or in Google search results. That is not proof of fabrication, some may be real firms with thin footprints, but it is exactly the pattern hallucination produces, and it means unverifiable shortlists reach real buyers.
How often should I check what AI says about my brand?
Monthly is the right cadence for most services firms, matching how fast the engines' retrieval and answers actually drift. Check sooner after events that change your facts: a rebrand, an acquisition, a pricing change, new locations. Those transitions are precisely when engines mix old and new information, which is where damaging hallucinations usually start.
Sources and further reading
- TofuBofu Research: Top MSPs in British Columbia: the 4-engine pilot behind the numbers in this post, with full methodology and per-company data
- Lewis et al., Retrieval-Augmented Generation (arXiv:2005.11401): the research showing retrieval grounds generation in real documents
- G2 2026 AI Search Insight Report: 51% of B2B buyers now start research with an AI chatbot, which is why wrong answers reach real pipelines
- Search Engine Land: fixing your brand's AI hallucinations: an independent practitioner guide to the same problem