Optimise differently for browsing and memory

Before this lecture, you should know the difference between training memory and live browsing from lecture 1, how public evidence and source trails work from lecture 2, why stale evidence matters from lecture 6, how third-party signals can help or confuse from lecture 9, and how source-level fixes work from lecture 10. We are now separating repair actions by the kind of ChatGPT answer they are meant to influence.

During a small workshop for a Belgian immigration practice, I asked the same question twice: “Which boutique firm can help a Dutch-speaking family with Belgian family reunification in Brussels?” The first answer, with browsing active, found a current service page and described the firm almost correctly, although it still softened the legal category. The second answer, in a setting where no current web check appeared, named a larger competitor and called our target firm “mobility support.” Same prompt. Same firm. Two different rooms in the same house.

That teaching example is simplified, but the pattern is familiar. The lawyer in the room wanted one fix. The communications manager wanted to rewrite the homepage. The directory profile was also wrong, but only in French. At that point the useful question was no longer “How do we optimise for ChatGPT?” It was smaller and sharper: are we trying to improve what ChatGPT can find now, or what future answer behaviour may already have absorbed from the public record?

The same answer may be shaped by different mechanisms

A ChatGPT answer about a Belgian immigration firm can be shaped by stored associations, by current retrieval, or by a mixture that is hard to see from the outside. We should not pretend to know the internal recipe for a particular answer. What we can do is separate the likely repair path.

Browsing optimisation is work that improves what ChatGPT can find and cite during current retrieval, because a browsing answer can only use what its search process can reach, read and trust enough to bring back. The work is practical and near-term: crawlable pages, clear titles, accessible service descriptions, corrected profiles, consistent third-party mentions, and sources that state the current fact plainly.

Memory optimisation is long-term public-record work aimed at future model knowledge. It moves more slowly. It is not a button, not a submission form, and not a promise that next week’s answer will change. If a model has already internalised a loose association between a firm and “relocation support,” the remedy is to make the public record more stable, repeated and correct over time. That may later shape systems trained or refreshed on public material, but the timing is not in the firm’s hands.

The distinction matters because the wrong repair wastes patience. If the problem appears only in browsing answers, then current source surfaces deserve attention first. If the problem appears in answers that do not seem to consult current sources, rewriting one page may still be right, but it should be framed as public-record work, not as an immediate correction. A firm can do useful work without pretending every answer mode obeys the same lever.

Browsing answers need reachable surfaces

In lecture 10, we repaired wrong claims where the weak evidence appeared. Lecture 11 adds another question: can a current retrieval process even reach the corrected evidence?

A retrieval surface is a page, profile, register or source type that a browsing answer may discover. The phrase sounds technical, but the object is ordinary. A Dutch service page is a retrieval surface. A French directory profile is another one. An official register entry, a professional listing, a referral page, a local article, a page title, even a short lawyer biography can become part of the surface ChatGPT searches across when it tries to answer a current question.

For a boutique immigration firm, the best browsing work is often unglamorous. Make the page public. Keep it reachable without a login. Put the core facts in text, not only in a PDF image or a decorative card. State the firm name, location, legal category and client problem near each other. Avoid hiding the useful sentence behind a script that loads after the page has already confused the reader. A human may forgive a beautiful but vague page. A browsing answer may simply leave with the clearest sentence it can carry.

Composite Object B, the Brussels cross-border practice, shows how this plays out. Its Dutch page correctly says family reunification. Its French profile says mobilité internationale, which is not wrong by itself but does not name the legal category early. Its English page speaks to families “moving to Belgium” before it names immigration law. In a browsing answer, the model may retrieve whichever surface looks most directly relevant to the user’s wording. If the user says “move to Belgium,” the English page and a relocation-flavoured directory may come forward before the more precise Dutch page.

So the browsing repair is not “publish more.” It is to make the right surfaces easier to retrieve for the right questions. That may mean adding a compact paragraph to the English family reunification page, correcting the French profile, or making sure the official and professional references use the same firm name and city. The work is less like painting a shopfront and more like placing labels on the correct drawers in a shared archive.

Memory-shaped answers need a stable public record

Memory-shaped answers are more awkward to discuss because they resist quick proof. We can observe answer patterns, but we cannot inspect the training mixture that produced them. Still, the work is not mystical. A future model is more likely to absorb a firm accurately when the public record repeatedly says the same true things across reachable sources.

For memory optimisation, repetition matters, but not the noisy kind. Ten thin profiles saying “visa help” can train the wrong association more efficiently than one careful page can correct it. A stable public record repeats the correct legal category, location and service shape without swelling into claims the firm cannot support. The sentence may be boring. Boring is useful here. “The firm advises on Belgian immigration-law matters for residence, family reunification and work-related permits” does more work than a paragraph of warm fog.

Recency signals still matter, but differently from browsing. In browsing, a current page may be found during the answer. In memory-shaped behaviour, old public material may have left a longer trace. If the firm moved office, changed service focus, or repaired a language mismatch, the public record should show the current version across several places. One corrected page surrounded by old profiles is like a corrected passport tucked into a drawer while everyone at the border still reads the old photocopy.

This is why memory optimisation often feels slow to clients. The firm has done the right work and the answer may still lag. That lag does not make the work foolish. It means the work belongs to a longer public-record strategy. The firm is not arguing with one answer. It is making future wrong answers harder to assemble.

Choose the repair by the failure pattern

When a firm asks what to do next, I start with the failure pattern rather than with a favourite tactic. A browsing failure and a memory-shaped failure can look similar on the surface. Both may call the firm by the wrong category. Both may prefer a larger neighbour. Both may miss a corrected service page. The difference is in the clues.

If a browsing answer cites or appears to use an old directory profile, the repair begins with that profile or with stronger nearby retrieval surfaces. If a browsing answer cannot find the firm for a Dutch prompt but finds it for an English one, the Dutch public evidence may be too thin, too hidden or too weakly connected to the right client problem. If a browsing answer finds the firm but quotes the wrong phrase, the liftable sentence on the surface may be vague or noisy.

If the answer does not appear to retrieve current sources, the repair is slower. Look for repeated public associations that may already exist: old categories, long-standing directory labels, broad biography language, partner pages that describe the firm loosely, or local mentions that keep using a former city. The task is to make the stable record clearer, not to expect one corrected page to rewrite stored associations on command.

A recurrent pattern in Belgian legal visibility is language-specific drift. The Dutch evidence may be legally precise, the French evidence may be broader, and the English evidence may be written for nervous clients rather than for factual extraction. Browsing optimisation asks which language surface a current answer can find. Memory optimisation asks what repeated association each language has been leaving behind. Those are related questions, but not the same one.

There is one repair that helps both modes: clarity repeated in public. A precise service page helps browsing if it can be found now. The same page also contributes to the longer record if it remains stable, crawlable and consistent with other sources. This is why the course keeps returning to public evidence. The mechanism changes, but the material is still the record.

Test without mixing the rooms

After a source-level fix, it is tempting to run a prompt, see a better answer, and declare the work done. Be careful. If the new test used browsing and the original failure did not, you may have tested a different room. If the first prompt was in French and the retest is in English, you may have changed the surface. If the first question asked for “relocation help” and the retest asks for “Belgian immigration law,” you have also changed the pull of the prompt.

Keep the test notes plain. Record the date, language, prompt, whether the answer seemed to use current sources, whether the firm was named, how it was described, and which public sources appeared to shape the answer. Do not turn this into a ranking ceremony. The goal is to see whether the same failure pattern persists under similar conditions.

For browsing tests, inspect the visible sources or source clues. Did the answer find the corrected page? Did it prefer a directory? Did it miss the official register? Did it retrieve the wrong language version? For memory-shaped tests, be more cautious. You are looking at repeated descriptions, not a live source trail. If the same old category appears again and again without current sourcing, mark it as a longer-term public-record problem.

The distinction also protects the firm from bad promises. A consultant should not say, “We fixed the website, so ChatGPT will stop saying that.” A better statement is more honest: “We corrected the source most likely to shape browsing answers, and we strengthened the public record that future answers may learn from.” Less dramatic, yes. More durable.

What to remember

Browsing optimisation works on what ChatGPT can find during current retrieval. Memory optimisation works on the slower public record that may shape future model knowledge.

Retrieval surface: A page, profile, register or source type that a browsing answer may discover.

A browsing fix should start with reachable, crawlable, current surfaces: service pages, profiles, registers and third-party mentions that state the firm’s facts clearly.

Memory-shaped problems require patience. The firm builds repeated, consistent public evidence so future answer behaviour has fewer wrong associations to borrow.

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

Do not test different rooms and call it progress. A browsing retest and a memory-shaped retest answer different questions, even when the prompt looks similar.

Check yourself

Describe in your own words why browsing optimisation and memory optimisation should be handled separately.

Browsing optimisation and memory optimisation work on different parts of the problem. Browsing optimisation is about what ChatGPT can find during a current search-like answer, so the practical work concerns reachable pages, profiles, registers and other source surfaces. Memory optimisation is slower public-record work aimed at future model knowledge. A firm may correct a page today and still appear wrongly in a memory-shaped answer because the old association has already been absorbed or repeated elsewhere. Separating the two prevents false promises and helps the firm choose the right repair for the actual failure pattern.

Give a Belgian immigration-law example where a browsing fix would be the right first move.

A browsing fix would be the right first move if a current ChatGPT answer appears to use an outdated directory profile that calls the firm “relocation support” instead of Belgian immigration law. The source is public, reachable and likely shaping the current answer. I would first try to correct the directory category or description, then make sure the firm’s own service page states the legal category and client problem clearly. I would retest with the same prompt language to see whether the browsing answer now finds or quotes better evidence.

How would you distinguish a stale browsing surface from a longer-term memory-shaped association?

I would look for source clues and consistency across test conditions. If the answer cites or visibly echoes a current web source, and that source contains an old address or wrong category, I would treat it as a stale browsing surface. If the answer gives the same old description without showing current sources, across several similar prompts, I would be more cautious and treat it as a memory-shaped association. The repair may overlap, but the expectation changes: a browsing surface can sometimes be fixed directly, while memory-shaped behaviour may take longer to shift.

When would a corrected service page fail to improve a browsing answer?

A corrected service page may fail to improve a browsing answer if ChatGPT does not discover that page, if another source is clearer, or if the page hides the useful facts too deeply. For example, the firm may correct its English family reunification page, but a French directory profile still says “international mobility support” in a more extractable way. The answer may choose the directory because it looks more directly relevant to the prompt. The fix is then not simply more writing. The retrieval surface itself needs to become easier to find and quote accurately.

How would you explain this lecture to a lawyer who wants one universal ChatGPT fix?

I would say that there is no single fix because ChatGPT answers can be shaped in more than one way. Some answers search current public sources, so the firm should repair pages, profiles and registers that can be found now. Other answers seem shaped by older or stored associations, so the firm needs a steadier public record across time and languages. The practical work still looks familiar: correct weak evidence, state the firm’s category clearly, and keep sources aligned. The difference is how quickly we expect the answer pattern to respond.