The question sounds simple and is answered three different ways depending on how you ask it — which is why most brands come away with a flattering, useless answer.
Only the third one reflects what a buyer experiences, because a buyer does not know your name yet. The first tells you whether you are describable, which is a prerequisite. The second is the trap: it is a question about a candidate you introduced, and its answer is flattery with a confidence interval.
Run all three and you get a diagnostic rather than a score. Run only the second and you get a screenshot for a sales deck.
Two distinct stores of knowledge behave differently, and "knows my brand" conflates them.
During training the model absorbs text. Brand names that appear often, consistently, next to category words, become available. This knowledge is stable, slow, and has no sources to inspect. It is also why a model can confidently describe a company using details that are years out of date, or confidently describe a company that no longer exists.
When the assistant searches, it reads pages at answer time. Here the knowledge is current, citable, and entirely dependent on what those pages say. This is the layer you can influence this month.
The practical consequence: "does the AI know me" has two answers. It may recognise your name from training data and have nothing current about you. It may know nothing from training data and still describe you accurately because it just read your site. Both get reported as "the AI knows your brand", and they require completely different work.
The most common failure we see is not absence — it is fragmentation. A company appears as "Acme", "Acme Studio", "acme-studio.de" and "Acme GmbH" across its own site, its directories and its press mentions. To a person these are obviously one company. To a model resolving entities they are four weakly supported candidates, none of which has enough evidence to be recommended.
This is why fixing naming feels cosmetic and produces disproportionate results. You are not changing your brand; you are collapsing four guesses into one fact.
Three places matter most, in this order: your own pages (title tag, about page, footer, author bios), the directories that appear in your category's citations, and any profile with a description field. Use one form, add the legal suffix only where required, and keep the descriptive sentence identical everywhere.
When you ask "what do you know about X" and get a confident paragraph, check it against reality before believing it. Four things can be happening:
The test for confabulation is mechanical: every specific claim must be traceable to something you or a third party published. If a number appears in the answer and nowhere else, it was generated.
A brand probe tells you whether you are describable. A buying question tells you whether you are recommended. Confusing the two is how brands conclude they are visible when they are merely mentionable.
| Diagnosis | What it means | First action |
|---|---|---|
| Describable and recommended | Working. Protect it. | Keep descriptions consistent; re-check monthly. |
| Describable, never recommended | Known but not chosen. | Get into the comparison lists your category cites. |
| Stale description | Learned once, not updated. | Publish current, quotable facts; earn fresh mentions. |
| Entity collision | Merged with another company. | One name form plus a location and category in every description. |
| Unknown entirely | Not in the sources models read. | Listings and one strong explainer — not more articles. |
| Confabulated details | Gaps filled by the model. | Publish the facts it invented, correctly, in a citable place. |
A wrong description is not just embarrassing in a demo. If an assistant tells a buyer you serve a city you left, or a service you discontinued, that answer is doing damage in a conversation you never see and cannot correct afterwards. Unlike a website, there is no page to edit.
Which is why the discipline is worth having: publish the facts you want repeated, in the form you want them repeated, in enough independent places that the model has no gaps to fill.
Test all three questions at once. The free check runs the buying questions on a real assistant and shows which sources it trusted — the layer you can actually change.
Run the free AI visibility check →Ask an open question without supplying context: what do you know about the brand name and what it does. Compare the answer with your own one-line description. Anything the model adds that you never published is either third-party knowledge or invention, and the two look identical in the text.
Usually because it is resolving your name against a different entity, or because the only sentences it has seen came from pages describing you loosely. Consistent naming plus one quotable description fixes most cases.
No. Discussing an entity and volunteering it in a recommendation are separate behaviours. Many brands are describable and never recommended, because recommendation requires comparison against alternatives.
It is normal for a small business and not permanent. It usually means you are absent from the sources assistants read in your category — a corroboration problem, not a content-quantity problem.
Check every specific claim against something published. Numbers, founding years, locations and service lists that appear nowhere on your site or in a third-party source are generated, even when the rest of the description is accurate.
Most brands do not have a sentence a stranger could repeat. They have a tagline, which is a promise, and an about page, which is a history. A model needs a fact: what you do, for whom, and where.
The formula that survives being quoted: [Brand] is a [category] in [place] that [specific thing] for [specific customer].
The third can be lifted into an answer verbatim and still be correct. That is the test: quoting out of context is exactly what an assistant does, so a sentence that only works in context will not help you.
If your name is generic, translated or shared with another company, add two disambiguators to every description: a category and a place. "Acme" is ambiguous; "Acme, a bookkeeping firm in Leeds" is not. This costs nothing and resolves the collision described above.
Where the identical sentence belongs, by impact:
Organization or LocalBusiness schema with name, description, areaServed and sameAs pointing at your real profiles.Twenty minutes, no publishing and no links. It is the cheapest intervention in this discipline and the one most often skipped because it feels too small to matter.
Used as a diagnostic the probe is useful: it shows whether you are describable, accurate and current. Used as a score it flatters you and hides the number that matters - whether a stranger asking for what you sell is ever told your name.
None of these is a statistic; they are the shapes that recur when you run this test across small businesses. Each one points at a different fix, which is why a single score would have hidden it.
The model produces two fluent paragraphs about the company, with a founding year, a founder name and a service list. Nothing in it appears on the company site. What happened: the name is common enough that the model resolved it to a different business entirely, then described that business well. Fix: disambiguate with category and place, and publish the correct founding facts so the wrong ones have competition.
The model knows the company exists and describes it as it was six years ago: the old city, the old service, the old positioning. What happened: the brand was well covered once, then stopped appearing in places models read. Fix: corroboration, not correction - no amount of editing your own site updates a memory that was formed elsewhere.
The model has never heard of the company, and the buying question is answered with directories the company never claimed. What happened: the business has no presence in the four sources that define its category. Fix: listings first, one strong explainer second, links third - and in that order, because links cannot fix an absence from the pages that are already cited.
Three rules for the recall test:
Then, separately and clearly labelled, run the buying questions. Keep the two outputs in different columns of the same document, and never merge them into one number. That discipline is what makes the diagnosis actionable rather than reassuring.
Written by the OptiLink team. Related: How to get ChatGPT to recommend your brand · Do backlinks matter for AI search?