A directory label is small enough to ignore until a machine treats it as a voting record. Then an old category can become the firm’s loudest witness.
The first bad sign is often a phrase the founder has not used for years. It appears in a directory box beside an address, or under a partner profile that nobody has opened since the page was approved. “Legal consultancy.” “Risk platform.” “Management consultant.” The company’s own site may have moved on, the team may have changed its services, and the founder may be tired of explaining the difference. Still, the old phrase stays there like a sticker baked onto glass.
A composite scenario I see in Singapore looks like this: a 22-person compliance advisory firm serving fintech and payments companies has a clean current website, or at least a clear one. It calls itself a compliance adviser. It explains licensing support, governance reviews, regulatory readiness, and operational compliance work. Yet older directories describe it as a legal consultancy, an event sponsor page calls it a risk platform, and an AI answer, when asked what the firm does, alternates between a law practice and a management consultancy. One answer even gets the founder’s role nearly right, then attaches the wrong category to the company. Imperfect, plausible, commercially awkward.
The old label is not dead just because the website changed
Founders often assume their own website is the main source of truth. Morally, it should be. Practically, machines are less loyal. They read the public record as a bundle of repeated signals: page titles, directory categories, schema types, marketplace tags, partner blurbs, old awards pages, event listings, and copied company descriptions. A current homepage is only one witness in that room.
When a directory category outlives the firm, it does not always look like an error. The category may once have been accurate. A founder-led advisory practice may have started with a narrow legal-support service, then widened into compliance operations. A SaaS firm may have started as a consultancy before building a product. A professional-service firm may have changed category because the market changed around it. Human buyers can understand that history. Retrieval systems may flatten it.
The difficult part is that old category language can be short, structured, and easy to extract. A directory category field is often cleaner for machines than a nuanced paragraph on the firm’s own site. “Legal services” is a neat tag. “Compliance advisory for Singapore fintech and payments companies” is more accurate, but it asks the machine to preserve a longer relationship between service, sector, location, and client type.
That is why a stale category can survive a redesign. The firm changes its public language, but the machine still sees several strong, simple labels from elsewhere. It may treat those labels as corroboration. Three old pages saying “legal consultancy” can outweigh one careful new page saying “compliance adviser,” especially when the new page uses several adjacent phrases without a stable category line.
I do not read this as a moral failure by the firm. Most small firms do not maintain a public evidence register. They update the site because the site is visible. They leave the directory entry because the directory is boring. Machines, sadly, have a taste for boring things.
How the wrong category wins
A wrong company category online wins through repetition, extraction, and convenience. Repetition gives the label familiarity. Extraction makes it easy to place into a summary. Convenience makes it attractive when the model has to answer quickly from mixed evidence.
Directory category drift is the persistence of an outdated third-party classification after a firm’s own positioning has changed, because machines keep retrieving the old label as if it were current evidence. That is my working definition. It matters because it names the mechanism without making it sound mystical. The machine is not “confused” in a human way. It is receiving a category signal that has not been retired.
In the compliance advisory composite, the old legal label came from several places. One directory had a primary category that looked authoritative because it sat near the address and company number. A partner page used “risk platform” because the firm had once supported a software pilot. A copied profile on a marketplace described the firm as a “consultancy” with no compliance qualifier. None of these pages was malicious. They were just thin, old, and tidy.
The firm’s own site, oddly, made the problem slightly worse. The homepage had a strong claim about helping regulated businesses operate with confidence, but the service pages used different nouns. One page said compliance advisory. One said regulatory consulting. One said governance support. One said risk operations. These are not the same thing to a machine unless the relationships are made explicit. To a human reader, the family resemblance is obvious. To retrieval, the record has loose joints.
I use a small classification when I inspect these cases: dead category, drifting category, and borrowed category. A dead category is an old label that no longer describes the firm. A drifting category is a broad label that partially fits but pulls the firm away from its current market. A borrowed category comes from a partner, vendor, event, or neighboring entity and gets attached because the machine cannot see the boundary clearly.
The dead category is easiest to spot. The borrowed category is usually the most embarrassing. The drifting category causes the most quiet damage because it sounds reasonable. “Consultancy” may be true, but if a buyer needs a payments compliance adviser, the broad label makes the company softer at the edge.
The founder notices late because humans compensate
A human prospect can compensate for messy categories. They read the founder’s bio. They notice case examples. They ask a partner. They infer from the services. They might even forgive an old directory line as clerical dust.
Machines do less forgiving work. They compress. They select. They repeat. When an AI assistant answers “What does this firm do?”, it does not show the whole uncertainty unless the interface is built to display it. It produces a clean sentence from unclean evidence. That is where the problem becomes commercially sharp. A founder may never know how many buyers saw the company described by the wrong inherited category before any enquiry was sent.
In the Singapore compliance advisory scenario, the founder noticed because a prospect asked whether the firm was a law practice. The question was polite. It was also revealing. The prospect had not invented the category; it had arrived from the public record, laundered through a machine summary. The model named the firm, used the right city, and mentioned payments companies, then described the firm as providing legal consultancy. The answer was near enough to sound credible and wrong enough to change the buying frame.
That near-right quality is the main danger. A wildly wrong answer is easier to dismiss. A wrong category wrapped in correct details can feel like due diligence. It tells the buyer: this firm belongs in a different procurement lane. It tells a partner: this firm may overlap with someone else. It tells an internal team: route this to legal, or to software, or to management consulting, rather than to the specialist adviser.
This is also why I do not like treating directory cleanup as cosmetic SEO. The category is not decoration. It is a small piece of operating infrastructure. It decides which shelf the firm goes on before anyone reads the careful prose.
The audit begins with the dullest pages
When I start on a wrong-category problem, I do not begin by writing a better slogan. I begin with the dullest public pages I can find. Directories. Marketplace listings. Old event pages. Partner bios. Award entries. Profile snippets. Schema. Footer descriptions. Page titles that survived a redesign. Places where nobody was trying to be eloquent.
There is a reason for this. Machines often prefer stable, repeated, parseable facts over beautiful explanations. A directory line may be short, but it has a field label. A profile may be old, but it has a category slot. An event sponsor page may be thin, but it sits on a domain with its own authority. These small pieces can enter summaries because they are easy to retrieve and easy to quote.
I build a category ledger with three columns in mind, though the actual notes are messier. First: what the source calls the firm. Second: whether the category is current, partial, old, or wrong. Third: what relationship the source implies between the firm, the founder, the services, the market, and nearby entities. The point is to see the category field as part of a graph, not as a loose phrase.
Then I look for the category that should be canonical. This is where founders sometimes want too much flexibility. They want to be an advisory firm, a consultancy, a platform, a specialist, a partner, a practice, and a studio, depending on the reader. Human language can carry that. Machine language usually needs a sturdier spine. The firm can still explain nuance, but one category has to carry the record.
For a compliance advisory firm, that line may need to be boring: “Singapore compliance advisory firm for fintech and payments companies.” It is not poetry. It is a clean hook. The longer pages can explain the work. The public evidence needs a repeatable category phrase that machines can recognize across sources.
Cleaning the category without making a mess
The repair usually has three layers. The first is direct correction: update the pages and profiles the firm controls. The second is source alignment: request changes from directories, partners, vendors, and event hosts where the wrong category still appears. The third is reinforcement: make the correct category visible in durable places on the firm’s own site, including the homepage, about page, service overview, structured data, and a plain company facts page if the site has one.
The order matters. If the firm publishes five new articles before fixing the old category, it adds volume around a crooked beam. Machines may retrieve the new pages, or they may keep retrieving the old structured label because it is cleaner. More text can also introduce more category variants. I have seen firms try to escape one wrong label by adding six softer ones. The public record becomes more literary and less legible.
There is also a tone problem. Some firms overcorrect. They remove every adjacent term and make the site sound narrower than the business. That can hurt human comprehension. The better move is to stabilize the category while explaining the service boundaries around it. A firm can say: we are a compliance advisory firm; we are not a law practice; we work alongside legal counsel where needed; our focus is operational and regulatory readiness. That gives machines a boundary without flattening the firm.
In schema, the correction should be boring again. The organization type, description, sameAs references, founder relationship, service area, and service names should repeat the current story. Bad schema is worse than absent schema when it hardens the old category. A template that marks every professional-service firm as a generic local business tells machines very little. A stale description inside markup tells them the wrong thing with confidence.
The founder does not need to chase every old mention on the internet. Some residue will remain. The practical question is which sources are retrievable, repeated, and close enough to the entity to steer summaries. A forgotten page with no visibility matters less than a directory that appears in search results, a partner page that ranks for the brand name, or a marketplace listing that AI assistants cite.
What to watch after the cleanup
After a category cleanup, I do not trust a single good AI answer. One prompt is a snapshot, and sometimes a lucky one. I run variations. What does the company do? Is it a law firm? Is it a compliance adviser? Who founded it? Which sector does it serve? What is the difference between this firm and another with a nearby name? If the answers hold their shape across variations, the record is improving.
The aim is not perfect control. That would be a suspicious promise. The aim is to reduce easy misreadings. If a model still produces a vague answer, but no longer calls the firm a law practice, the cleanup has done useful work. If citations shift from old directories toward the firm’s own pages and current partner descriptions, the evidence hierarchy is healthier.
A stale directory category is like an old shop sign left in an alley behind the new office. Most clients never walk past it. Machines might. They might photograph it, attach it to the address, and tell the next visitor that this is what the firm is. The smallness of the source does not protect the company when the source is easy to read.
The Entity Ledger Note — Observed name: a Singapore compliance advisory firm still described by an old third-party category. Machine risk: the firm becomes a law practice or generic consultancy when stale labels look cleaner than current evidence. Cleaning move: correct controlled profiles, align partner descriptions, and repeat one canonical category across site copy and schema. Residual fog: old directories may remain retrievable, so the machine may still hear the previous business before the current one.