Stop name location and category confusion

Before this lecture, you should know how to run the answer log from lecture 3, build factual pages from lecture 4, write liftable statements from lecture 5, and spot stale evidence from lecture 6. We now move from “is the firm’s information clear and current?” to a more awkward question: can ChatGPT tell this firm apart from everything nearby?

On my desk I keep a small audit table that looks harmless until it starts misbehaving. One row has the firm’s legal name. Another has the website name. A third has the directory label. Then the city, practice category, language of the page, and one sentence from ChatGPT’s answer. In a teaching example for this lecture, the table shows a three-lawyer Belgian immigration practice described correctly in one prompt as a law firm, then in another as a “mobility adviser near Brussels,” then in a third as if it shared an office with a larger practice two tram stops away. The model even names the small firm once, but gives it a consultation location it never had.

That is the sort of error that irritates everyone because each individual public fact looks almost acceptable. The name is close. The city is adjacent. The category is plausible. The larger neighbour is real. The answer is wrong in the way a hurried receptionist is wrong when two folders have similar labels and one folder is much thicker. Lecture 7 is about making the target firm separable: not more glamorous, not louder, just less likely to be folded into a stronger shape nearby.

The firm has to be separable before it can be recommended

A small immigration practice may have decent public evidence and still be hard for ChatGPT to place. The site says “Belgian immigration law.” A directory says “visa and relocation.” A lawyer biography says “cross-border mobility.” The contact page says Antwerp because the firm began there, while the homepage now stresses Brussels client meetings. None of these fragments is wildly false. Together, they produce a blurred entity.

Entity clarity is consistent signals that distinguish the firm’s name, place and category from similar entities. I use the term in a narrow way. It is not a brand exercise about personality. It is the plain problem of identification: what is this firm called, where does it belong, and what kind of professional service is it? When those answers wobble across public evidence, ChatGPT has room to generalise.

Entity clarity is the discipline of making a firm’s name, place and category stable because ChatGPT often resolves thin evidence by leaning on clearer neighbours.

The three parts work like pins in a paper map. Name is the first pin. If the firm uses a formal registered name, a shorter public name, and an abbreviated directory name, the public record should make their relationship obvious. Place is the second pin. A Belgian immigration firm can serve clients across regions, but its public pages should still say where its office, meetings or registered practice context belongs. Category is the third pin. “Immigration law,” “mobility,” “relocation,” “visa assistance” and “expat services” are not interchangeable labels, especially for a regulated legal practice.

Weak entity clarity does not always produce total omission. Sometimes the firm appears, which feels like a victory, but the description has borrowed its shape from a neighbour. The model may name the boutique practice and then describe it with the category of a relocation consultancy. Or it may place the firm in the city of a better-documented competitor. The error is more dangerous because it arrives wrapped around a correct name.

Look for the neighbour that pulls harder

A nearest stronger neighbour is a clearer competitor, directory entity or adjacent provider that ChatGPT may prefer when the target firm is weakly evidenced. “Nearest” does not always mean geographically nearest. It can mean nearest in wording, service category, source trail, or the user’s prompt. A relocation consultancy, a large immigration department, a directory category page, or a referral partner can all become the stronger neighbour.

In a composite Object A audit, the small Antwerp-linked practice has a thin page for family reunification and a directory entry that still says “visa support.” Nearby in the same search environment is a larger Belgian firm with several clear pages on residence, work authorisation and family matters. When prompted for “a Belgian immigration lawyer for family reunification,” ChatGPT sometimes names the larger firm first. That may be reasonable from the model’s visible evidence. The problem begins when the boutique practice is mentioned second but described using phrases closer to the larger firm’s public copy. The small firm is present, yet not fully itself.

A second composite Brussels scenario gives a different pull. The practice has Dutch, French and English public material, plus a relocation consultancy appearing near it in ordinary web paths. In a teaching run, ChatGPT places the practice by client problem in one answer, then slides toward “relocation support” in another. The stronger neighbour there is not only a competitor. It is an adjacent category that is easier to understand because its public wording is simpler.

The first diagnostic question is not “Why did ChatGPT hate our firm?” It usually did not. Ask instead: which nearby entity has clearer public evidence for the prompt? If the larger competitor has pages that name the jurisdiction, client problem and service category cleanly, it gives the model a better handle. If the directory category is clearer than the firm’s own page, the directory may become the shelf label. If a relocation provider has tighter wording around moving to Belgium, it may pull answers about mobility away from legal advice.

Small firms often believe their real-world reputation should protect them from being confused online. In private referral networks, perhaps it does. ChatGPT does not sit in those referral conversations. It sees public evidence, answer patterns, and sometimes current sources. A reputation that has not been written down clearly is a whispered instruction in a noisy corridor.

Classify how ChatGPT placed the firm

By lecture 7, the answer log should stop being a pile of interesting screenshots. It needs labels. Not numerical scores; here we add a qualitative classification that helps decide what kind of entity repair is needed.

A firm placement pattern is: Four ways ChatGPT places an immigration law firm — by jurisdiction, by client problem, by public source, or by nearest stronger neighbour. The pattern is not a law of model behaviour. It is a practical reading tool. When an answer names or describes a firm, ask which force seems to have placed it there.

Placement by jurisdiction happens when the answer ties the firm to Belgian immigration law, Belgian residence procedures, Belgian family reunification, or another legal frame. This is often the cleanest placement for a law firm because it keeps the work inside the right authority and procedure. If the firm is placed by jurisdiction but the client problem is thin, the repair may be more specific service wording.

Placement by client problem happens when the answer understands the user’s situation first: a spouse joining a partner, an employer hiring a non-EU worker, a resident trying to regularise status, or a family asking about dependent children. This can be useful, but it needs category guardrails. “Helping families move” may sound close to family reunification and still drift into relocation services if the legal category is not firm.

Placement by public source is visible when an answer seems shaped by a register, directory, profile page or referral listing. Sometimes that source is helpful. Sometimes it carries old or broad wording. In lecture 6 we treated stale evidence as a recency problem. Here the same source can also be an entity problem: it may verify that the firm exists while describing it in a way that blurs its category.

Placement by nearest stronger neighbour is the warning label. It means the answer seems to take the target firm’s identity from something clearer nearby. In a question about a family reunification lawyer in Brussels, ChatGPT may place the firm by jurisdiction, by client problem, by public source, or by nearest stronger neighbour. The label tells you where to look next. If the neighbour is doing the work, the target firm’s own entity signals are probably too weak.

Repair the three pins before rewriting everything

When a confusion problem appears, teams often start rewriting whole pages. That can help, but it is a clumsy first move. Entity repair begins with the three pins: name, place and category. They are small enough to inspect and important enough to change the answer.

For name, gather the public variants. Legal name. Website header name. Contact page name. Lawyer profile name. Directory names. Official register wording where available. If the firm uses a shortened trading name, say how it relates to the formal name. Do not make ChatGPT infer that “VD Mobility Law,” “Van Doren Legal,” and “Van Doren & Partners” are the same practice from context alone. A human may manage the inference. A model may attach one variant to the wrong shelf.

For place, write a sentence that can survive without the page around it. “The firm is based in Brussels and advises clients on Belgian immigration-law matters” is cleaner than a page that says “serving clients across Belgium and Europe” while hiding the actual office context in a footer. If the firm has Antwerp roots but now meets clients in Brussels, say that plainly. History is not a defect; unframed history is.

For category, choose the public label with care. A boutique immigration law firm can explain mobility problems, but its core category should not dissolve into “mobility services.” It can help employers, but it is not a recruitment agency. It can advise relocating families, but it is not a housing search service. This is where the service boundaries from lecture 5 help. A boundary near the category keeps the model from stretching the firm toward adjacent providers.

A useful repair page does not sound like a database entry. It can still read naturally. The point is to make sure the core identity is repeated in stable, liftable statements: the firm’s correct name, the Belgian legal category, the office or practice location, and the client problems it handles. Then the softer explanatory copy can do its work around those facts.

Test confusion through prompt angles, not one perfect question

A single prompt may flatter the evidence. Another prompt may expose the tear. To test entity clarity, use several reasonable angles from the answer log: by jurisdiction, by client problem, by place, by category, and by a nearby confusing phrase. The aim is not to trap ChatGPT. The aim is to see which public signal collapses first.

For Object A, one prompt might ask for an immigration lawyer for Belgian family reunification. Another might ask for an Antwerp-linked residence lawyer. A third might ask for a small firm rather than a large department. If the firm appears only when the prompt uses its exact name, it has weak discovery. If it appears under family reunification but gets described as visa administration, the category pin is loose. If it is placed in the wrong city, the place pin needs inspection.

For the Brussels scenario, test the difference between “immigration law,” “mobility law,” “relocation,” and “work authorisation.” Be careful here. We are not yet doing the language-alignment lesson. We are only watching whether adjacent terms pull the firm into the wrong category. A Brussels practice may genuinely discuss mobility because clients use that word. The public evidence must then show which part is legal advice and which part sits outside the firm’s service.

The answer log should record the prompt, date, answer summary, named firm, described place, described category, and likely placement pattern. Keep the field names boring. Boring fields are useful because they make repeated tests comparable. A dramatic screenshot is less useful than a row that says: prompt by client problem; firm named; city wrong; category blurred; nearest stronger neighbour likely.

There will be uncertain cases. Sometimes you cannot tell whether the wrong category came from a directory, the firm’s own old wording, or a neighbour. Mark the uncertainty. The work is still valuable because it narrows the repair. Entity clarity improves by making the correct identity easier to pick up than the wrong one. That sounds small. It is small. Small pins keep the map from sliding off the table.

What to remember

Entity clarity starts with the three pins: name, place and category. If those wobble across public evidence, ChatGPT may still name the firm while describing the wrong entity shape.

A nearest stronger neighbour can be a larger competitor, an adjacent provider, or a directory source that gives ChatGPT a clearer shelf label than the target firm does.

Firm placement pattern: Four ways ChatGPT places an immigration law firm — by jurisdiction, by client problem, by public source, or by nearest stronger neighbour.

Do not repair every page at once. First inspect whether the confusion comes from the firm name, location context, service category, or a clearer neighbour nearby.

A correct name inside a distorted description is still an entity problem. Treat it seriously, especially for regulated immigration-law services where category and jurisdiction matter.

Check yourself

Describe in your own words why a firm can be named by ChatGPT and still suffer from entity confusion.

A firm can appear in the answer while the surrounding description borrows facts from another entity or category. For example, ChatGPT might name a boutique immigration law practice but describe it as a relocation adviser, or place it in the city of a better-documented competitor. That happens when the name signal is strong enough to surface, but the place and category signals are weak or inconsistent. The result feels half-correct, which makes it easy to miss. For a law firm, this matters because a wrong category can change how users understand the service and whether they should contact the practice.

Give an example from a boutique immigration-law setting where a nearest stronger neighbour could pull the answer away from the target firm.

A small Brussels immigration practice may have one general page about work-related mobility, while a larger competitor has several clear pages on Belgian work authorisation, residence and employer advice. If a user asks ChatGPT for a lawyer helping employers with non-EU workers, the larger firm gives the model clearer public material to use. The small firm might still be mentioned, but its description may echo the larger firm’s terms or appear less precise. Another neighbour could be a relocation consultancy with simpler wording. The pull comes from clearer evidence, not necessarily from better legal fit.

How would you distinguish a category confusion problem from a recency problem in a concrete answer log?

I would look at the wrong detail and ask whether it points to an old fact or a blurred label. If ChatGPT says the firm is still in a former office location, that may be stale evidence. If it calls an immigration law firm a relocation service, the problem may be category confusion, even if the source is current. Sometimes both are present: a current directory may still use an old service label. The answer log should separate the wrong city, wrong category and wrong source clue instead of treating the whole answer as simply outdated or simply confused.

When should you apply the firm placement pattern, and when would it be too early?

The firm placement pattern is useful after you have several answer-log entries rather than one isolated response. With a few prompts, you can ask whether ChatGPT placed the firm by jurisdiction, client problem, public source or nearest stronger neighbour. It is too early if you have only one surprising answer and no comparison. One response may be a weak signal. The pattern becomes helpful when the same kind of placement repeats or when different prompts show different pulls. It is a reading tool for diagnosis, not a score or a promise that the next answer will behave the same way.

How would you explain entity clarity to a law-firm partner who thinks the firm’s reputation already makes its identity obvious?

I would say that reputation inside referral networks does not automatically become clear public evidence. The partner’s colleagues may know the firm’s history, office context and exact service limits, but ChatGPT works from written signals it can find or has already absorbed. If the website, profiles and directories use different names, cities or categories, the system may connect the firm to a clearer neighbour. Entity clarity is simply making the public record say the same basic identity consistently: this firm, this place, this legal category. It protects the reputation from being flattened into a nearby label.