Separate ChatGPT optimisation from SEO habits

I sometimes begin this lesson with a composite scene: a Belgian immigration lawyer shows me a search result page with the satisfaction of someone placing a clean file on a judge’s desk. The firm appears on page one for a narrow Dutch-language query about family reunification. The title is sensible. The snippet is not embarrassing. Then we ask ChatGPT a plain client question: “Which small immigration lawyer in Antwerp helps with Belgian residence permits for a non-EU spouse?” The answer names two larger firms, one relocation consultant, and a lawyer who moved offices years earlier. Our careful firm is missing. Not damaged. Just absent, like a stamped annex left outside the bundle.

The shape is familiar. A boutique practice can be visible to a human who searches patiently and still be weakly present inside an AI answer. The lawyer’s instinct is to ask, “What keyword did we lose?” That instinct is understandable. It is also the first habit we need to loosen. ChatGPT is not handing the user a ranked shelf of links. It is making a sentence-shaped judgement about which names are safe to mention, how to describe them, and what uncertainty to smooth over.

The answer is not a search result wearing a nicer coat

A search result page makes the user do the last part of the work. It shows titles, snippets, maps, directory entries, and sometimes ads. The person still has to click, compare, distrust, and assemble meaning. A ChatGPT recommendation answer compresses that comparison into prose. It may say, “You could consider firms that handle Belgian immigration and family reunification matters,” then list names or categories. The compression is the point. It is also where small firms disappear.

For this course, ChatGPT optimisation is work that helps ChatGPT recognise, place, quote and describe a firm accurately from public evidence. I am using the term narrowly. It is not “make the website louder.” It is not a promise that a model will mention you. It is the practical discipline of asking whether the firm can be understood from the material available outside the consultation room.

The legal domain makes this sharper than, say, a bakery or a yoga studio. An immigration firm is not just a place plus opening hours. It has jurisdiction, regulated status, service limits, languages, client types, and procedures that change meaning if the wording slips. “Visa help” is too wide. “Belgian family reunification for non-EU spouses” is narrower, but still not enough if the page does not say whether the firm advises, files, represents, or only explains options. ChatGPT tends to prefer wording it can safely reuse. When the firm’s own words are soft felt, the model may reach for a harder surface elsewhere.

There is a small trap here. If a firm already ranks in Google for a service query, the team may assume ChatGPT will inherit that visibility. Sometimes it does. Often it does not. Ranking is a position in a list. An answer is a constructed paragraph. The model has to decide whether the firm belongs in that paragraph at all.

Memory and browsing are different answer conditions

When I say training memory in this course, I mean offline model knowledge shaped before the user asks, not a live check of current pages. That definition matters because many disappointing firm answers come from a mixture of old associations, broad category knowledge, and whatever the model has internalised about a place or practice area. The outside observer cannot open the model and inspect those associations. We can only test the answer it produces and look carefully at the kind of mistake.

Live browsing is ChatGPT behaviour that consults current web sources before forming an answer. That changes the work, but it does not make the work simple. A browsed answer may inspect current material and still choose a directory snippet over a firm page if the directory is clearer. It may find a source and summarise it badly. It may also search with a phrase that does not match how the firm describes itself.

A recurrent pattern in small professional services is the split answer. In one run, without visible browsing, ChatGPT gives a generic answer and mentions no firms. In another run, with browsing or search-like behaviour, it finds names, but the descriptions come from thin listings. The first problem feels like invisibility. The second feels like misdescription. They need different treatment. You do not fix both by rewriting a title tag and hoping the machine becomes grateful.

There is also a separate product use of the word memory, where a user’s own saved context may affect that user’s conversations. That is not the same thing as training memory in this course. I mention this because the word “memory” is sticky. In legal communications work, sticky words cause sloppy diagnosis. If a partner says, “ChatGPT remembers us wrong,” ask first: are we talking about a public answer any user might get, or a personalised account that has its own prior context?

SEO separation starts with the test object

SEO separation is the discipline of not treating search ranking and AI answer visibility as the same output. The discipline sounds dry, almost clerical. It is not. It keeps you from measuring a teacup with a parking meter.

Take a simplified firm scenario. A Dutch-speaking boutique practice in Mechelen has a clear homepage, a page for work permits, and two directory profiles. It appears reasonably well when a human searches its name plus “advocaat vreemdelingenrecht.” Then a potential client asks ChatGPT, “Who can help a Canadian founder move staff to Belgium?” The model may answer around corporate immigration, employer sponsorship, or relocation planning. The firm’s exact keyword ranking is only one possible ingredient. The answer might not use the same phrase, and it might not behave like a user scanning the first ten links.

This is why our first test object is not a keyword. It is an answer situation. Who is asking? What problem do they describe? Which jurisdiction is implied? Do they ask for a named firm, a kind of lawyer, or a comparison? The query “immigration lawyer Belgium” is a blunt instrument. A real client question often arrives with a crooked handle: “My husband is Moroccan and I live in Brussels, do we need a lawyer for family reunification?” That question carries location, relationship, status anxiety, and a request for reassurance. ChatGPT may respond with legal generalities before naming anyone.

SEO habits still matter, of course. Clear pages, crawlable text, readable headings, and consistent service wording help both humans and machines. The mechanism is different enough that we should not pretend one dashboard proves the other. A firm can improve search traffic and remain absent from answer recommendations. A firm can also be named by ChatGPT while its search traffic is modest, because the model has found a crisp association between firm, place and service.

The practical change is small but uncomfortable: stop starting with “Where do we rank?” and start with “What answer does ChatGPT produce when a plausible client asks?” The first question belongs to search. The second belongs to ChatGPT optimisation.

Other answer engines teach us by contrast

Other answer engines are useful here because they show different levels of visible source use. A more visibly source-led answer pattern does not make an engine automatically more correct. It does make the source layer more exposed to the user. You can often see which pages the answer leans on, even if the interpretation still needs checking.

ChatGPT can work in several modes depending on the product, settings, and task. Sometimes it answers from what it already has available in the conversation and model. Sometimes it searches. Sometimes the user gives it documents. A Belgian immigration firm cannot assume the same action improves every mode equally. A better service page may help a browsing answer quickly. A stable, repeated public description may matter more slowly for future model knowledge. A private PDF uploaded by the user may help only that conversation.

I do not want you to turn this into platform superstition. The weak version of AI visibility work says, “This engine likes lists, that engine likes schema, this one likes freshness.” Sometimes there is a little truth there, and a lot of theatre. The sturdier version starts with the answer condition: is the system retrieving now, drawing from prior knowledge, or reading material the user supplied? Until you know which condition you are testing, your optimisation advice is a coat hung on the wrong peg.

For boutique immigration firms, the contrast matters because clients rarely know which engine mode they are using. They ask a question and trust the paragraph. If the paragraph names a larger competitor because that competitor is easier to describe, the small firm’s problem is not merely traffic. It is representation inside a decision aid.

The first exercise is observation, not repair

Before rewriting anything, run a small set of plain questions. Do not make them flattering. Do not ask, “Why is Finch & Partners the best immigration firm in Antwerp?” Ask like a client, a referral partner, and a cautious relative. Use the languages in which the firm actually serves clients, but for this first lecture keep the notes simple. Notice whether ChatGPT names the firm, avoids names, chooses large firms, invents work labels, or answers only with general legal guidance.

A teaching example: ask, “Which boutique lawyers help with Belgian family reunification in Brussels?” Then ask, “I need a Dutch-speaking immigration lawyer for my spouse’s residence card near Antwerp; who should I consider?” Then ask, “Is there a small law firm in Belgium that handles work permits for cross-border hires?” The rough edge will show up quickly. Maybe ChatGPT names firms in Brussels for the Antwerp question. Maybe it describes an immigration lawyer as a consultant. Maybe it gives cautious advice and refuses to recommend anyone. Each result is a clue about the answer situation, not a verdict on the firm’s worth.

Record the answer text, the date, the prompt, and whether the answer appears to browse. That is enough for day one. Resist the urge to fix the homepage in the same hour. Legal teams often move from discomfort to editing too fast. They see a wrong answer and start adding paragraphs, badges, testimonials, and procedural detail until the page becomes a cupboard no one can close. First see the shape of the misunderstanding. Then decide what kind of work is actually needed.

ChatGPT optimisation begins when we treat the answer as an observable object. Search rankings remain useful, but they are not the object in front of us here. The object is the paragraph a potential client reads when they ask for help and do not know which names to trust.

What to remember

ChatGPT optimisation starts with answer observation. If you only inspect search position, you may miss the place where the client actually meets the firm: a generated paragraph.

Work that helps ChatGPT recognise, place, quote and describe a firm accurately from public evidence.

Training memory and live browsing create different answer conditions. A stale or generic answer may not have the same cause as a browsed answer built from weak current sources.

SEO separation protects the work from false comfort. A visible search result does not guarantee that ChatGPT can safely name, place or describe a boutique immigration practice.

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

For legal services, vague wording creates special risk. If the firm’s jurisdiction, service and client situation are blurred, the answer may become cautious, generic or pulled toward clearer competitors.

Check yourself

Explain in your own words why a law firm can rank well in search and still be absent from a ChatGPT recommendation answer.

A search result and a ChatGPT answer ask different things of the firm’s material. Search can show a page because it matches a query, leaving the user to inspect and decide. ChatGPT has to build a paragraph and decide whether a firm is safe to mention in that paragraph. A boutique immigration firm may have enough keyword relevance to appear in search, but not enough clear wording for ChatGPT to place it by service, city or client problem. The issue is not only visibility; it is whether the model can describe the firm without guessing.

Give an example from an immigration-law setting where SEO thinking would lead to the wrong first repair.

A firm might see that ChatGPT does not mention it for “Belgian family reunification lawyer” and immediately rewrite its page title around that phrase. That may help search, but it may not solve the answer problem. The model might be avoiding the firm because the page never states whether the lawyers handle spouse residence applications, where they practise, or which languages they support. In that case, the better first repair is not another keyword variation. It is clearer factual wording that lets the answer understand the firm’s role in a realistic client situation.

How would you tell the difference between a training-memory problem and a live-browsing problem in a practical test?

I would first note whether the answer shows signs of checking current web sources, such as cited links or source references. If there is no browsing, the answer may reflect broad stored associations or old knowledge, so I would treat it as a training-memory condition. If browsing is visible, I would look at what the system seems to find and reuse. A browsed answer that mislabels the firm may point to unclear current pages or directory text. The same wrong description can therefore have different likely causes depending on the answer condition.

When should you apply SEO separation, and when would it be unnecessary?

SEO separation is needed whenever someone tries to use search performance as proof of ChatGPT visibility. It is especially useful for boutique firms that already rank for their own name or a narrow service query but do not appear in generated recommendations. It may be less necessary for a task that is purely about ordinary search results, such as improving a title tag for a known keyword. Once the question becomes “what does ChatGPT say about this firm?”, the output has changed. Then search ranking is context, not the main measurement.

How would you explain ChatGPT optimisation to a referral partner who knows legal services but not AI systems?

I would say that ChatGPT optimisation is about making the firm’s public description clear enough that an AI answer can identify and describe it accurately. It is similar to preparing a good referral note, except the reader is a machine that compresses many signals into a short answer. The work is not to make exaggerated claims or chase every keyword. It is to reduce avoidable confusion: where the firm works, what immigration matters it handles, who it helps, and what it should not be mistaken for.