Keep firm evidence current despite cutoffs

Before this lecture, you should understand the SEO separation from lecture 1, the source-trail habit from lecture 2, and the extraction work from lecture 5. You should be able to look at a public page and ask whether ChatGPT could lift a service fact accurately, but now we add a harder question: could it lift the current fact?

A composite scenario for this lecture begins with a Brussels mobility-law page that has been fixed in English. The old wording said the firm “supported relocation and residence matters for international professionals.” After a careful rewrite, the English page now says the firm advises on Belgian immigration-law aspects of work-related mobility. Good. Then a Dutch prompt still produces an answer calling the practice “a relocation adviser with legal contacts,” and the lawyers stare at the screen as if the website has betrayed them. It has not, exactly. One language version changed; another public trail kept speaking in the old voice.

There is a second wrinkle. In one answer mode, ChatGPT seems to draw on old public associations. In another, it appears to consult current web pages and still finds the outdated directory first. That difference matters. A stale answer is not always caused by the same kind of staleness. Sometimes the model is leaning on training memory. Sometimes live browsing finds the wrong fresh-looking source. Sometimes the firm’s own pages have been corrected, but the public record around them still has last year’s label stuck to it like a courthouse notice nobody removed.

The page changed; the public record did not

When a boutique firm updates its website, the partners often expect the change to behave like switching on a lamp. The old description should disappear. The new one should be visible. In ordinary human reading, that expectation is partly fair. A referred client lands on the new page and sees the current wording. ChatGPT, however, may be working from a wider and less tidy public record.

Training memory is offline model knowledge shaped before the user asks, not a live check of current pages. Live browsing is ChatGPT behaviour that consults current web sources before forming an answer. Lecture 1 separated those answer conditions because they create different kinds of uncertainty. In lecture 6, that distinction becomes practical. If an answer is shaped by training memory, yesterday’s website edit may not affect it immediately. If an answer uses live browsing, the edit can matter sooner, but only if the right current source is found and understood.

This is why I dislike telling firms, “Just update the page and test again tomorrow.” It may work for a browsing-shaped answer. It may do very little for a memory-shaped answer. It may also expose a third problem: the firm’s page is current, but a directory, referral page, lawyer biography, old PDF or translated profile still carries the obsolete description. The model does not need malice to repeat the wrong thing. It only needs one source that is clearer, older and easier to quote than the corrected page.

A recency signal is a public clue that a page or profile reflects current services and locations. A date can be a recency signal, but it is not the only one. Current office wording, updated lawyer profiles, consistent service labels, language-version alignment and removal of old categories can all help. Stale evidence is accessible public information that no longer reflects the firm’s current name, location or service shape. The danger is not merely age. A ten-year-old page may still be accurate. A profile updated three weeks after an office move may already be stale if it kept the wrong city.

Diagnose the type of staleness before rewriting again

A stale ChatGPT answer tempts people into frantic rewriting. The homepage gets changed. Then the service page. Then the profile. Then the title tag, because someone in the room remembers SEO. The result is often more public movement but not more public clarity. Before another edit, diagnose which kind of staleness you are seeing.

Start with the answer itself. Does it use an old office location? Does it use an old service category? Does it name a lawyer who no longer appears on the site? Does it blend legal advice with relocation logistics? Each wrong detail is a thread. Pull one at a time. A wrong city points toward contact pages, register entries, directory profiles and older biographies. A wrong category points toward service pages, third-party listings and soft language that survived in one language version. A wrong personnel detail may point toward lawyer profiles, press mentions or old PDFs still indexed somewhere.

A teaching example makes the method less abstract. Suppose a Belgian immigration firm moved client meetings from Antwerp to Brussels but kept an Antwerp-linked history on its site. A ChatGPT answer says it is “an Antwerp visa firm.” That answer contains two defects. “Antwerp” may come from a real history that now needs clearer framing. “Visa firm” may come from an old category label. If the team edits only the contact page, the category problem remains. If it edits only the service page, the location problem remains. The stale evidence has two roots, not one.

For browsing-shaped answers, test the likely source trail by looking at what current sources are easy to find and quote. The firm’s corrected page may be accurate but buried under a clearer directory snippet. For memory-shaped answers, the test is slower and less certain. Repeated prompts may show that an old association still appears even when browsing is not in play. We cannot see the internal store. We can only observe the output and improve the public evidence that future systems may absorb.

The uncomfortable lesson is simple: do not rewrite the newest page until you know why the old fact is still alive.

Dates help only when the substance is current

I see teams add “updated” labels to pages as if a date alone can disinfect weak evidence. It cannot. A recency signal has to point to current substance. If the page says “Updated May 2026” but still describes the firm as a general mobility adviser, the date only makes the wrong description look freshly wrong.

A good recency signal sits near the fact it supports. On a service page, a short note can say that the page reflects the firm’s current Belgian immigration-law services. On a contact page, current meeting arrangements should be unambiguous. On a lawyer profile, the service category should match the practice pages. In multilingual evidence, the Dutch, French and English versions should not carry different eras of the firm. One language page should not behave like a time capsule.

A composite Brussels case shows the trap. The English page is rewritten to say “Belgian immigration-law advice for employers and cross-border professionals.” The French directory profile still says “relocation support.” The Dutch page still uses a broader mobility phrase and lists an old office area. In a Dutch-language prompt, ChatGPT describes the firm through the Dutch page and directory. In an English prompt, it sometimes gets closer to the corrected version. The firm has not one public record, but several records walking at different speeds.

There is no need to date every paragraph. That can look fussy and sometimes absurd. The better discipline is to make current facts easy to identify. Service pages should show current service shape. Profiles should reinforce the same category. Directory profiles should be corrected where possible. Old pages that remain useful should explain their age or context. A page about a past seminar, for example, should not accidentally become the best public description of the firm’s current work.

Recency also belongs inside liftable statements. “The firm advises on Belgian work-related immigration matters” is useful. “The firm currently advises on Belgian work-related immigration matters and does not provide relocation logistics” is stronger if the old public trail confused the two. Use “currently” carefully, not as decoration. It should repair a known stale association.

Match the update to the answer mode

The course does not teach magic handles for changing ChatGPT. It teaches source work under uncertainty. That means the right update depends on whether the observed problem is more likely tied to browsing, memory or a mixed public trail. The distinction is imperfect, but it prevents bad work.

For live browsing, prioritise current retrieval surfaces. The firm’s own pages should be accurate, clear and easy to quote. Public profiles that appear prominently should be corrected. Official registers should be checked for name and location consistency where relevant. Directory text should be cleaned where the firm has control. If a current source is wrong and easy to find, it can keep feeding wrong answers even after the firm website improves.

For training memory, the work is slower. You cannot force an immediate internal update from outside. You can build a stable public record that says the same correct thing across crawlable sources over time. That sounds unsatisfying because it is unsatisfying. I would rather say that plainly than sell a fantasy of instant correction. Memory-shaped visibility is more like drying a damp wall than wiping a table. You remove the source of damp, ventilate the room, and accept that the stain does not vanish on command.

Mixed cases are common. A prompt may produce an answer that sounds old, then browsing adds one current source, then the final answer blends both. A firm may be described with its current service page but an old city. Or with the right city and wrong category. In those cases, do not argue about whether the answer is “memory” or “browsing” in a pure sense. Track the wrong facts and repair the sources that could be supporting them.

This is also why testing once is thin evidence. Ask the same question in a few reasonable forms: by service, by location, by client problem, and by language. Record whether the old fact repeats. If it appears only in Dutch prompts, the language trail deserves inspection. If it appears across all prompts, the association may be broader. If it appears only when browsing is used, a current source is probably doing damage.

Build a maintenance rhythm, not a panic ritual

Recency work should not happen only after ChatGPT gives an embarrassing answer. Boutique immigration firms already have natural maintenance moments: a service changes, a lawyer joins or leaves, office arrangements change, a language page is translated, a directory profile is renewed, a referral partner updates a page, or a regulation-facing explanation is rewritten. Each moment should trigger a small public-evidence check.

A simple rhythm works better than a dramatic overhaul. When a service page changes, check the matching lawyer profiles and language versions. When a location changes, check the contact page, public profiles and any official listing the firm relies on. When a category is corrected from relocation support to immigration law, inspect the pages where the old phrase still lingers. Do not make the homepage carry the whole repair. Public evidence is distributed, so maintenance must be distributed too.

The firm also needs a small stale-evidence list. Not a grand dashboard. A practical list: source, old fact, current fact, who can change it, date checked, and next step. Some items will be under the firm’s control. Others will sit with directories or third parties. Some cannot be fixed quickly. The value of the list is that it stops the team from rediscovering the same obsolete profile every quarter with fresh irritation.

There is a craft point here. Current evidence should still be extractable. A page can be up to date and too vague to help. It can also be precise and out of date. The useful page is both current and liftable: clear enough for ChatGPT to reuse, recent enough not to revive old service shapes. That is the line we are trying to hold.

What to remember

A stale ChatGPT answer may come from training memory, live browsing, or a current source that repeats an old fact.

Recency signal is a public clue that a page or profile reflects current services and locations.

Stale evidence is accessible public information that no longer reflects the firm’s current name, location or service shape.

Do not add dates as decoration. A useful recency signal supports a current fact, such as service scope, office context or language-version alignment.

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

Check yourself

Describe in your own words why updating one service page may not correct a stale ChatGPT answer.

Updating one service page may not be enough because ChatGPT can be shaped by more than the newest page on the firm’s site. A memory-shaped answer may still reflect older public associations. A browsing-shaped answer may find a directory, profile or translated page that still uses the old category or location. The corrected page helps, but it has to compete with the rest of the public record. For a Belgian immigration firm, the English page may say “immigration law” while a Dutch page or directory still says “relocation support.” The answer may repeat whichever source is clearer or easier to reuse.

Give an example of stale evidence in an immigration-law firm’s public record and explain the likely risk.

A firm may have moved its client meetings from Antwerp to Brussels while an old directory profile still lists the former office area. The risk is not only that ChatGPT names the wrong city. The wrong location may also pull the firm into the wrong comparison group when a user asks for a lawyer near Brussels. Another example is an old profile describing the practice as “visa assistance” after the firm has clarified that it provides Belgian immigration-law advice. That stale category can make the answer sound administrative rather than legal, which is a serious distortion for a regulated service.

How would you distinguish a stale memory problem from a browsing problem without claiming certainty?

I would avoid pretending that I can see ChatGPT’s internal process. Instead, I would compare answer conditions and repeated prompts. If the wrong fact appears in answers that do not seem to cite or use current sources, it may be tied to older model associations. If the wrong fact appears when current sources are consulted, I would inspect the pages and profiles that are easy to find now. The distinction remains an inference, not proof. The practical response is to repair current public evidence for browsing while building a stable, consistent public record for future memory-shaped answers.

When should a firm use a date or “current” wording on a public page, and when would that be cosmetic?

A date or “current” wording helps when it supports a specific fact that might otherwise be confused with old evidence. For example, if a firm no longer provides relocation logistics, a current service statement can clarify that it now advises on Belgian immigration-law aspects of mobility. That is useful because it repairs a known stale association. The same date becomes cosmetic when the underlying copy remains vague or wrong. A page marked “updated” but still using the old category only makes the bad information look newer. Recency signals need substance behind them.

How would you explain recency maintenance to a partner who wants to “fix ChatGPT” once and move on?

I would explain that ChatGPT visibility is affected by a distributed public record, not by one switch inside the firm’s website. A single fix can help, especially for browsing answers, but old profiles, language versions and directory text may continue to speak for the firm. Recency maintenance is a small rhythm: check the service pages, profiles, registers, directories and translations when something important changes. It is less dramatic than a campaign, but it prevents old facts from hardening. For a boutique immigration practice, accuracy over time is part of professional trust.