The first time you see it you assume it is a one-off. You ask an assistant what your company does, and the answer is fluent, confident, well structured and wrong in a specific way. It has you in the right industry. It has the wrong buyer. It attributes a feature you retired, or a funding round belonging to a company with a similar name, or a founder who has never worked for you.
Then you ask again with slightly different wording and get a different wrong answer.
The shapes the confusion takes
Once you start collecting these, they sort themselves into a handful of recognisable patterns.
- Category misplacement. You get described as the thing you were three years ago, or as an adjacent category that shares your vocabulary.
- Feature invention. Capabilities you do not have, usually ones typical of your category, get attached to your name because the description was assembled from what companies like yours generally do.
- Merged identity. Two organisations with similar names become one, and the resulting profile is a blend of both.
- Stale fact. Old pricing models, former offices, departed executives, discontinued products. All of it still in circulation.
- Attribution slip. Something you published gets credited elsewhere, or something you did not publish gets credited to you.
The patterns are not equally damaging. A stale office location costs you nothing. A confident description of you as a small-business tool when your entire book of business is enterprise costs you the consideration set, silently, for as long as it persists.
The accounts of what causes LLM brand hallucination that I find useful all come back to the same underlying condition: the text available about you was thin, inconsistent, or old, and the gap got filled with a plausible average.
Where the wrong description comes from
It helps to stop thinking about this as a model problem and start thinking about it as an evidence problem.
Whatever gets said about your company in an answer was assembled from text. Some of it yours. A larger portion of it not. Directory entries somebody filled in years ago and nobody has touched since. A funding announcement with a one-line descriptor a PR agency wrote at speed. A conference programme bio. Old press coverage that framed you as a competitor to a company you have nothing to do with now. Review site categories chosen by whoever set up the listing.
None of those sources were written to be authoritative. They just happen to be numerous and stable, and stability tends to beat accuracy when nothing more recent contradicts them.
Thin evidence is the other half. Companies that are small, young, or operating in a niche without much third-party writing about it get described by inference from their category, because there is not enough specific material to describe them any other way. You can feel the difference in the answers: the ones about well-documented companies are specific and checkable, the ones about thinly documented companies are smooth and generic.
Shared names and the merge problem
Name collisions produce the most stubborn version of this, and almost nobody checks for them before they name the company.
Two companies with the same name in different industries usually stay separate, because the surrounding context differs enough to keep them apart. The trouble starts when the two are adjacent: same name, both in software, one in logistics and one in fintech. Descriptions bleed across. A leadership team from one gets attributed to the other. Answers hedge, or worse, they do not hedge and pick wrong.
There is a similar problem with parent companies, acquisitions and rebrands. A company acquired and folded into a larger group keeps generating answers under the old identity for a long while, and answers about the new identity keep carrying attributes of the old one. Anyone who has been through a rebrand will tell you the corpus does not forget on your announcement schedule.
Practical response, where a collision exists: make everything you control carry a disambiguating phrase alongside the name. Industry, location, function, something. Not as keyword stuffing, just as the kind of clarifying clause a human writer would naturally include when two things share a name.
Facts that never got retired
The drift that catches established companies is different from the confusion that catches new ones.
You launched, you got written about, you grew, you changed. Every stage left behind text describing you accurately at the time and inaccurately now. Nobody goes back to correct a three-year-old blog post on somebody else’s site. The old descriptions sit there, they were never wrong exactly, and they keep informing how you get described today, shaping your brand’s AI reputation. Longevity works against you here in a way it never did in search, where a stale page simply lost its ranking and stopped mattering.
What makes this insidious is that it looks like nothing is happening. There is no error message. Your website is current, your positioning is sharp, your team is aligned, and the answers people get about you are describing a company you stopped being at some point you cannot identify. You find out when a prospect says something odd on a call.
Teams that stay on top of it treat it as maintenance rather than a project. A recurring check, a list of the sources that keep getting cited, a slow programme of getting the worst offenders updated or outweighed. Boring work. It compounds in the right direction, which is more than can be said for a lot of marketing activity.
The fixes that actually hold
There is no single lever, and anyone who tells you there is has something to sell. What there is instead is a sequence that tends to work when followed in order.
- Write the canonical description. One or two sentences covering what the company is, who it serves, what category it sits in. No adjectives, nothing aspirational, nothing that stops being true when the roadmap ships.
- Make it consistent everywhere you control. Site, boilerplate, social profiles, directory listings, executive bios, partner pages, app store entries, support documentation. Identical wording, not variations on a theme.
- Fix the sources you can reach. Directories with edit access, listings with a contact form, review profiles where you have a claimed page. Unglamorous and effective.
- Publish the facts plainly on your own domain. A page that states what you do, who it is for, what it costs to consider, who runs the company, in sentences rather than in marketing copy. Assistants retrieve prose.
- Earn accurate third-party description. New writing about you, in the current framing, from sources that get cited. This is the slow part and the part that eventually outweighs the old material.
- Then maintain it, because step five keeps producing new text that has to stay consistent with step one.
Step two is where teams quietly fail. Not because it is hard, but because it is distributed across five departments and nobody owns it, so the sales deck says one thing, the careers page says another, and the CEO’s conference bio says a third.
What structured data does and does not do
Schema markup gets treated as the fix for this, and it does less and more than that.
Less, because marking up an organisation does not make an assistant describe you correctly. What surfaces in answers is prose that got retrieved and read, and structured data is not prose. You cannot mark your way out of a corpus that contradicts you.
More, because consistent structured data across your properties, with stable identifiers pointing at the same profiles you reference everywhere else, is the cheapest way to make the machine-readable version of your identity agree with the human-readable one. When those two disagree, you have created ambiguity for no benefit at all.
So mark it up, keep it accurate, keep it matching your visible content, and stop expecting it to carry the whole job.
Checking whether any of it worked
You will want a number and you are not going to get one. What you can get is a log.
Write down the questions that matter: what does the company do, who is it for, how does it compare to the obvious alternatives, is it suitable for a specific use case. Ask them across the assistants your buyers use, on a fixed cadence, in consistent wording to monitor your AI search visibility. Record the description that comes back and the sources cited underneath it. Then make a change and keep asking.
The work on fixing entity drift when AI gets your brand wrong that has been most useful to me makes the same point about timelines: the corrections take longer to appear than the errors took to establish, and the interval is long enough that teams give up before the fix lands.
What eventually happens, if you keep at it, is undramatic. One day you ask the question you have been asking for months and the answer is just correct. No feature you never built, no buyer you never served, no company you were confused with. Correct, and slightly dull, and exactly what you wrote down in the first place.
