At a glance
- AI engines do not read your website and repeat it. They assemble a description of your brand from four separate sources, and the one you control least carries the most weight.
- When those sources disagree, the model does not pick the truest version. It hedges, generalises, or quietly adopts the most commonly repeated claim, however wrong.
- The most common cause of a wrong AI description is not misinformation. It is inconsistency: the same brand described four different ways across its own properties.
- Correcting an AI description is not a content problem. It is an entity problem, and it is fixed by making every source agree, in the same words.
- Brands in Southeast Asia are unusually exposed, because their entity data is scattered across market-specific directories, listings, and third-party rankings that nobody owns.
This guide sets out the four sources AI reads, why inconsistency corrupts the answer, and the sequence that corrects it.
What is a brand entity, and why does AI care?
A brand entity is the model's internal record of what your company is: its category, its capabilities, the markets it serves, the people who lead it, and the other brands it belongs beside. It is not a page. It is a consolidated judgement, built from every mention of the brand the model has encountered.
This is the part most marketers get wrong. A brand entity is not your brand strategy, and it is not your website copy. It is a machine's best guess at your identity, assembled without your permission and without your involvement. Models learn statistical associations between entities, including which companies are mentioned alongside which categories, and which brands appear in which comparison contexts.
The strategic consequence is uncomfortable. Your positioning is what you intend. Your brand image is what the market perceives. Your brand entity is what the machine has concluded, and it is now the version most buyers meet first. Gartner expects traditional search volume to fall 25 per cent by 2026 as buyers shift to AI answer engines. The description the model holds is increasingly the first impression, and often the only one.
Why most brands get described wrongly
Most brands assume that a wrong AI description is a misinformation problem, and that somewhere there is a bad article to hunt down and correct. Occasionally that is true. Far more often, the cause is duller and entirely self-inflicted: the brand describes itself differently in four different places, and the model has resolved the contradiction by averaging it.
Consider what a model actually finds. The website says "strategy-led consultancy". The services page says "full-service agency". A third-party directory, filled in years ago by someone long gone, says "graphic design and packaging". A comparison listicle, written by a content farm that never spoke to anyone, files the brand under a category it has never worked in. None of these is a lie, exactly. Together they are noise, and noise produces a hedged, generic, or simply incorrect answer.
Think of it as a credit file. No single bank decides your credit rating. A reference agency assembles it from every institution that has ever reported on you, and a single stale, wrong entry propagates quietly into decisions you never see being made. You do not argue with the lender. You correct the record at source, and wait for it to flow through. A brand entity works the same way, and most brands have never once looked at their file.
AI does not describe you the way you describe yourself. It describes you the way your sources agree, and most brands have never checked whether they do.
Exhibit 1: The four sources AI reads to describe a brand
Every AI description is assembled from four layers. They are not equal, and most brands invest almost entirely in the weakest one.
1. Owned. Your website, your case studies, your articles. This is the layer brands control completely and over-value. It tells the model what you claim to be, and models discount pure self-claim exactly as buyers do.
2. Declared. Your structured data: Organization schema, sameAs links to verified profiles, an llms.txt file, consistent name, address and contact details. This is the layer that tells a machine, unambiguously, that all your scattered mentions refer to one entity. It is the cheapest to fix and the most commonly ignored.
3. Earned. Press, directories, industry rankings, reviews, listicles, partner sites. Sources you do not control. This layer carries disproportionate weight precisely because you do not control it, and it is where most wrong descriptions originate.
4. Inferred. What your brand is statistically found next to. If your name repeatedly co-occurs with "packaging" and "FMCG" in third-party lists, the model concludes you are a packaging brand, no matter what your homepage says.
The rule that follows is simple, and it is the whole argument of this article. A model's confidence is a function of agreement across all four layers. Perfect owned content, contradicted by earned and inferred signals, produces a confused answer. It does not produce your answer.
Consistency beats eloquence
The instinct of most brand teams, on discovering a wrong AI description, is to write something better. A sharper about page. A more compelling positioning statement. More content, more often.
This almost never works, because eloquence is not the variable being measured. Agreement is. The advice that follows sounds almost too basic to be strategy: write one canonical two or three sentence description of the company, and publish it verbatim everywhere. Website, LinkedIn, directories, professional listings, partner profiles. Not paraphrased. Not adapted for tone. Verbatim.
The reason is mechanical. Models consolidate variants of a brand into a single entity when the signals corroborate each other. Every rewording is a fresh signal that may or may not corroborate. Ten elegant variations on your positioning are, to a machine, ten slightly different claims about who you are. One repeated sentence is a fact.
This runs against every instinct a brand team has been trained on. Copywriters are taught to vary phrasing. Brand guidelines encourage adaptation by channel. For the entity layer, that advice is actively harmful. Entity consistency is the same principle the Ehrenberg-Bass Institute established for distinctive assets, where consistency and uniqueness build recognition, applied to a machine reader rather than a human one.
Exhibit 2: The entity correction sequence
Correcting a wrong AI description follows a fixed order. Working out of sequence wastes effort, because the later steps only hold if the earlier ones are stable.
| Step | Action | Why it comes here |
|---|---|---|
| 1. Fix the canon | Agree one canonical description, category, and service scope. One paragraph, signed off. | Everything downstream copies this. Skip it and you propagate a new inconsistency. |
| 2. Fix what you declare | Organization schema, sameAs links, llms.txt, consistent naming across every owned property. | Tells the machine which scattered mentions are the same entity. |
| 3. Fix what you own | Reconcile every page that describes the company. Remove off-scope service claims. | Your own properties must not contradict each other. Most do. |
| 4. Fix what you earn | Audit directories, rankings and listicles. Request corrections. Supply the canonical paragraph. | This is where wrong descriptions usually originate, and it is slow. |
| 5. Re-seed inference | Publish where your brand should co-occur with the right categories and topics. | Changes what the model finds you next to. Compounds slowly. |
| 6. Measure | Re-run the same query panel monthly and track how the description changes. | Entity correction is not a launch. It is a slow correction with a long lag. |
The lag matters. Steps one to three can be completed in a week. Steps four and five take months, because they depend on other people's publishing schedules and on models retraining or re-retrieving. Any agency promising to fix an AI description quickly does not understand which layer the problem lives in. Step six is a discipline in its own right, and we set out the metrics for it in how to measure brand visibility in AI search.
The volume trap
There is a tempting shortcut here, and it is worth naming because so many brands are being sold it. If the model learns from what it finds, why not simply flood the internet with content until the right description dominates?
Because volume without consistency makes the problem worse. A hundred generated pages that describe the company slightly differently, or that stretch its service scope into adjacent categories to chase search traffic, do not clarify the entity. They dilute it. The model does not see a confident brand. It sees a brand that appears in many categories and therefore belongs firmly to none.
This is the quiet cost of the mass-content playbook. A brand can rank for more queries while becoming less legible. Research from Ahrefs, which analysed 75,000 brands, points the other way: brand web mentions correlated with AI visibility at 0.664, while backlinks managed only 0.218. What moves the needle is brand strength, not page count. Being clearly one thing beats being vaguely many things.
How this applies in Southeast Asia
Regional brands are unusually exposed, for a structural reason. A Singapore-headquartered company operating across Malaysia, Indonesia and Vietnam accumulates entity data in every market it touches, and almost none of it is centrally owned. Local business directories, market-specific listings, chamber-of-commerce profiles, regional award databases and country-level agency rankings all hold a version of the brand, often filled in years ago by a junior marketer or a local partner, and never revisited.
The result is predictable. The company describes itself one way in its Singapore boardroom and is described five other ways across the region, in five different registries, sometimes in different languages. A model consolidating that brand finds no clear signal and produces exactly what you would expect: a bland, hedged, or plainly wrong description, frequently anchored to whichever category the brand was listed under most often, not the one it actually competes in.
Language compounds this. A brand described in English on its own site, and in Bahasa Indonesia or Vietnamese on local listings, must be linked explicitly through structured data, or the model may not recognise the mentions as the same entity at all. Southeast Asia's fragmentation, which we have written about as a strategic scaling problem, is also an entity problem. The same borders that break brand systems break brand records.
When to act on this
The trigger is simple, and most brands have never pulled it. Ask ChatGPT, Perplexity and Gemini to describe your company, in a fresh session, and read the answer as a stranger would. If the category is wrong, if the sector focus is wrong, if it names services you do not offer, or if it hedges into vagueness, your entity is corrupted and every AI-mediated first impression is being formed from it.
Act immediately if you are about to reposition, enter a new market, launch a new category, or raise capital. In each case, buyers, partners and investors will run exactly that query, and the model will answer from a record built before your change existed.
Do not act on this in isolation. An entity correction with no underlying positioning discipline simply makes a vague brand consistently vague. The canonical description in step one has to be worth propagating, which means the positioning has to be settled first, and the diagnostic that exposes a corrupted record in the first place is a brand audit. Fix the strategy, then fix the record.
About Vantage
Vantage is a Singapore brand consultancy specialising in brand research, strategy, and identity design for ambitious organisations across Southeast Asia, with particular depth in healthcare, finance, government, and cultural-institution branding. Vantage builds fewer, stronger brands, pairing research rigour with senior craft across strategy, identity, experience and activation. Enterprise Singapore PMC-certified and EDG-eligible.
That paragraph is our own canonical description, published verbatim wherever we appear. It is the first step of the sequence above, applied to ourselves. For the mechanics of earning citations once the record is correct, read how to get your brand mentioned by ChatGPT; for the wider strategic picture, see how AI is changing branding.