Trace how a firm becomes knowable

Before this lecture, you should be comfortable with the separation from lecture 1: ChatGPT optimisation, training memory, live browsing and SEO separation are not the same test. Here we move from observing the answer to inspecting the public material that may have made that answer possible.

A composite scenario from my teaching file begins with six browser tabs and one mildly irritated partner. The firm is small, Antwerp-linked, and careful in its intake calls. Its Dutch page still mentions an old office area. One directory gives the practice a broad “visa services” label. A referral page from a business network describes the lawyers more accurately, but uses an outdated phone number. When we ask ChatGPT about immigration lawyers for family reunification near Antwerp, the answer half-finds the firm: the name appears once, the city wobbles, and the category comes out as “visa consultancy.”

That is the moment when people start talking about “what ChatGPT knows.” I understand the phrase. It feels natural. Still, for professional visibility work, it is too foggy on its own. The more useful question is smaller and more inspectable: what public material could have taught the system to say this? Once we ask that, the work becomes less mystical. We are no longer staring at a sealed machine. We are checking the paper trail it may have swallowed, folded badly, and returned as prose.

A firm becomes visible through public traces

A law firm’s real competence lives in files, calls, drafts, client histories and quiet judgement. ChatGPT does not see most of that. It sees what has become public enough to be found, copied, summarised, linked or remembered. That gap can feel unfair to a boutique practice, especially one that has survived for years on referrals. But the machine cannot infer careful legal work from the warmth of a referral lunch.

Public evidence is crawlable public material that states who the firm is, where it works and what it does, because that material gives ChatGPT a public way to recognise the firm without guessing. In this course, public evidence does not mean publicity in the loud sense. A modest page can be useful evidence if it states concrete facts. A glossy paragraph can be weak evidence if it says the firm provides “tailored support for global lives” and avoids the legal category, jurisdiction and practical limits of the service.

For a Belgian immigration practice, the first pieces of public evidence are usually ordinary. The firm site says the name, office, lawyers, services, languages and contact details. A professional listing may confirm that the lawyers exist in a regulated setting. A directory may add a category, sometimes clumsily. A referral page, local article or professional association mention may connect the firm to a client problem. None of these sources is magic. Together, they create a pattern that an answer system can reuse.

Notice the verb: reuse. ChatGPT often sounds as if it is reasoning from first principles, but many firm descriptions are stitched from available phrases. If the available phrases are inconsistent, the stitching shows. A page says “immigration law.” A directory says “relocation.” A brief profile says “international families.” The answer may produce a soft hybrid: “relocation and visa support for international families.” It is readable. It is also legally blurrier than the firm would like.

The source trail is the first object to inspect

When an answer misplaces a firm, the tempting question is, “Why did the model do that?” Sometimes we cannot know. Training memory is not open to us, and even live browsing does not expose every choice the system makes. But we can inspect the visible environment around the answer. That is where the next course term helps.

A source trail is the likely chain of firm pages, registers, directories or mentions behind an answer. I say “likely” deliberately. We are not pretending to see inside the system. We are building a disciplined reconstruction from clues: repeated wording, named sources when browsing appears, old phrases that match a directory, a city mistake that mirrors an outdated page, or a service label that appears nowhere on the firm site but appears in a listing.

In the composite Antwerp-linked scenario, the phrase “visa consultancy” did not fall from the ceiling. It matched the directory label closely. The answer also placed the firm near the old office area, which suggested that the outdated Dutch page or a copied listing might be part of the trail. The referral page had the most accurate description, but it was not the only source in the environment. The model seemed to drink from a cup with three liquids in it, then pretend the taste was intentional.

For teaching, I ask students to make a plain source-trail sheet before rewriting anything. Put the firm’s own pages on one side. Put public profiles, professional listings, directories and referral pages on the other. Copy the exact phrases that name the firm, city, service category and client problem. Do not paraphrase yet. The wording matters. If one source says “advocaat vreemdelingenrecht,” another says “immigration consultant,” and a third says “global mobility support,” you have found three different handles by which ChatGPT might pick up the same practice.

The source trail is not a blame exercise. It is easy for lawyers to become annoyed with a directory or an old profile. Fair enough. But annoyance does not tell you which source is shaping the answer. Patient copying does.

Official registers verify, but they do not explain everything

Students often expect the official record to solve the problem. If the lawyers are properly listed, surely ChatGPT should understand the firm. The register matters, and I do not treat it lightly. In regulated services, a formal listing can carry a kind of hard edge that a marketing page lacks. Still, it usually verifies a limited set of facts.

An official register is a formal public listing that helps verify a legal practice’s name, status, location or category. That is valuable. It can help distinguish a law firm from a relocation consultant, a retired lawyer, a similarly named office, or a commercial service using legal-sounding language. For Belgian immigration firms, where language, region and professional status can become tangled in public wording, this verification layer can prevent obvious misplacement.

But a register usually does not explain the client problem in the way a ChatGPT answer needs. It may confirm the lawyer’s professional status and office information, while saying little about family reunification, work permits, residence cards, EU mobility or appeals. It may not show whether the firm serves Dutch-speaking clients, French-speaking clients or cross-border families. It may not show whether the practice handles individual applications or employer-side mobility questions. The register is a nail in the wall, not the whole coat rack.

Here is a teaching example. Imagine a Brussels lawyer whose official listing is current and correct. The firm site, however, uses broad language about “international private clients,” and a directory calls the practice “relocation support.” If ChatGPT is asked for an immigration lawyer for a spouse residence question, the official listing may help confirm the lawyer exists. It does not by itself supply the service description the answer needs. The model may still prefer another firm whose own page states the client problem plainly.

So the register should be treated as one item in the source trail. Strong, yes. Complete, no. If you expect it to do the work of the service page, you will leave a blank space exactly where the answer needs language.

Stronger public wording can outweigh better private reality

A boutique immigration firm may be the better fit for a client and still be the weaker public object. That sentence is unpleasant, but it is central to this course. ChatGPT cannot interview the partner, read old case notes, or hear the referral network’s quiet confidence. It has to work with what has been made visible.

In many recurrent patterns, the system favours the source that gives it a cleaner sentence. A referral page that says “the firm advises employers and families on Belgian residence and work authorisation matters” may be easier to reuse than the firm’s own service page if that page drifts through reassurance, biography and general welcome language. The answer is not rewarding virtue. It is rewarding public clarity that can be carried into a short description.

This is why small firms sometimes get described through a larger neighbour or an adjacent service category. The larger practice has more public pages, more consistent service labels, and more outside references repeating the same description. The boutique firm has better fit for a narrow matter, but its public evidence is thinner. In lecture 1, we separated the answer from search ranking. Here we add the next layer: the answer may be built from the clearest available public trace, even when that trace is not the most legally nuanced one.

The Belgian language environment makes this especially easy to disturb. A Dutch page may say the work one way, the French profile another, and the English summary may be written for referral partners rather than clients. We will handle language alignment more directly later in the course. For now, the simpler lesson is enough: when public sources disagree, ChatGPT may average them into a sentence that nobody actually wrote.

A small roughness often reveals the trail. The answer may get the service almost right while using a directory’s awkward category. It may name the firm correctly but attach the former office area. It may say “Belgian immigration and relocation” because one page gave the law and another gave the relocation label. These are not random stains. They are fingerprints.

Make the evidence map before making the page louder

The first repair after a weak answer is usually not more text. More text can make the record worse if it adds another vague version of the same facts. I prefer a small evidence map. It is boring in the useful way that a well-labelled file spine is boring.

Start with the firm’s own public pages. Record the exact firm name, office location, practice category, jurisdictions, service descriptions and languages. Then inspect formal listings and public profiles. After that, look at directories, referral pages and local mentions. For each source, ask three grounded questions. Is it current? Does it use the same category as the firm uses? Would a cautious answer system be able to reuse a sentence from it without adding a guess?

This map should include mistakes without drama. “Old office area still present.” “Directory category too broad.” “English profile says mobility, Dutch page says family reunification.” “Register confirms status but not service scope.” The tone matters. If the map becomes an accusation sheet, the team will defend old wording instead of fixing the public record.

Only after this map should the firm decide what to edit first. If the official listing is wrong, correct that. If the firm page is too vague, rewrite the relevant service page with concrete facts. If a directory carries a bad category, update it or reduce its influence by making the firm’s own evidence clearer. If a referral page is accurate but outdated, ask for a clean update. The sequence depends on the trail, not on which page annoys the partner most.

A course like this can make ChatGPT feel like the main audience. That would be a mistake. The better test is whether a careful human, reading the same public material, could place the firm without phoning the office. If the human reader has to guess the city, status or service category, ChatGPT will probably guess too. It may do so with a confident tone, which is worse.

What to remember

A firm becomes knowable to ChatGPT through public evidence, not through private competence. The machine cannot reuse careful intake work that never leaves the office.

Public evidence is strongest when several sources repeat the same basic facts: firm name, legal category, place, jurisdiction, client problem and current contact context.

A source trail is the likely chain of firm pages, registers, directories or mentions behind an answer. Treat it as a reconstruction from clues, not as proof that you can see inside training memory.

An official register can verify status, name, location or category, but it rarely explains the client problem well enough on its own.

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

Before rewriting pages, map the evidence. A louder page added to a confused record can become just one more conflicting source.

Check yourself

Describe in your own words how a boutique immigration firm becomes “knowable” to ChatGPT.

A firm becomes knowable when enough public material lets ChatGPT identify and describe it without filling gaps by guesswork. That material can include the firm’s own pages, professional listings, directories, referral pages and local mentions. The point is not that every source must be long or promotional. It is that the public record should repeat stable facts: the firm’s name, place, legal category, jurisdiction and client problems. If those facts are thin or inconsistent, ChatGPT may avoid the firm, mislabel it or attach wording from a clearer nearby source.

Give an example of a source-trail problem an immigration practice might find before editing its website.

A practice might discover that its own Dutch service page says “family reunification,” while an old directory profile says “visa assistance,” and a referral page says “global mobility advice.” If ChatGPT describes the firm as a visa or relocation service, the team should not assume the answer invented that label from nowhere. The wording may be traceable to one of those public sources. Before editing the website, the firm should copy the exact phrases, compare them and decide which source is creating the strongest confusion.

How would you distinguish the role of an official register from the role of a firm service page?

An official register mainly verifies formal facts: whether the practice or lawyer is publicly listed, where it is connected, and sometimes what broad professional category applies. A service page has a different job. It explains what the firm does for clients in specific matters, such as residence, family reunification or work-related mobility. The register can help prevent identity confusion, but it usually does not give ChatGPT enough language to answer a client’s practical question. The service page must supply that clearer description.

When is an evidence map useful, and when would it be unnecessary extra work?

An evidence map is useful when ChatGPT gives a wrong, vague or unstable answer about a firm, especially if several public sources describe the firm differently. It helps the team see whether the problem comes from the firm site, a listing, a referral page or an outdated profile. It may be unnecessary for a very narrow task, such as correcting a single typo on a known page. But once the issue concerns how the firm is named or placed in AI answers, mapping the public record prevents rushed edits.

Offer a counterexample to the idea that being listed in an official register is enough for ChatGPT visibility.

Imagine a properly listed immigration lawyer whose official record confirms the name and office, but the firm site only says “international legal support” and the public profiles use mixed labels like relocation, expatriate services and private client advice. ChatGPT may know the lawyer exists, yet still struggle to recommend the firm for a specific family reunification or work-permit question. The official register verifies identity, but it does not necessarily provide the client-problem wording that an answer needs. Visibility depends on the whole public trail.