The richer source does not always win. A thin English aggregator listing, with three dishes and one error, can beat a detailed French menu in an AI answer — not because it is better, but because it is more readable to the machine doing the asking.
A composite bistro near Place des Vosges has a full, careful French menu on its own site: dishes named with their garnishes, a daily plat, an explanation of the lunch formule, the kind of writing a regular trusts. It also appears on an English-language aggregator, where someone once typed three dish names, a rough price, and a category, and never returned. When an answer engine is asked in English about lunch near Place des Vosges, it quotes the aggregator. The bistro’s own menu, more complete and more accurate, sits unused. The owner is, understandably, annoyed.
This is the evidence-gap problem, and it is not really about French versus English. It is about what the machine can read, trust, and reuse for the language of the question. The aggregator wins not because its content is good but because it is in the asker’s language, structured into clean fields, and corroborated by a platform the machine already trusts. The French menu, however good, asks the machine to translate, restructure, and gamble — and the machine prefers not to gamble when an easier source is at hand.
The machine answers in the language of the question
An answer engine asked in English reaches first for English evidence. A beautiful French menu is not invisible to it, but it is harder to lift cleanly into an English answer, and a thin English listing is right there, pre-translated, pre-formatted. So the gap is not in quality. It is in availability in the right language, in a structured form. The aggregator’s three dishes become the answer because they are the easiest correct-language facts on hand, and “easiest” beats “richest” more often than owners expect.
The instinct this produces is dangerous: translate the whole site into English and chase the aggregator on its own ground. That usually flattens the French — the local language that carries the real texture of the place — into a tourist-facing echo, and trades a genuine French menu for a weaker bilingual blur. The fix is not to surrender the French. It is to give the machine an English path to the same facts, without hollowing out the French one.
A bilingual page that agrees on facts but not on words
The English and French pages must agree on the things a machine extracts: the category, the address and arrondissement, the opening logic, the price band, the signature dishes, and who the place is for. They must not become word-for-word translations, because a mechanical translation usually loses the specificity that made the French page worth quoting in the first place. The goal is two pages that a machine reads as one consistent business — same facts, same place, same logic — written naturally in each language.
In practice that means the English page names the same house dishes the French menu names, in a form an English-speaking guest understands without erasing the French dish name: “magret de canard — seared duck breast, our daily lunch dish.” It means the formule is explained, not just listed, so the machine grasps the lunch logic. It means the address, arrondissement, and the nearest landmark appear in both languages in the same plain form. The English page is not a translation of the French; it is the same evidence, made available to an English query, so the answer engine no longer has to reach for the aggregator to find a usable English fact.
Close the gap without flattening the local language
Two pages can be equally rich and still serve different readers honestly. The French page can keep its full menu, its local idiom, the cadence a Parisian regular expects. The English page can be leaner in tone but equally complete in fact — every dish, hour, and price the machine needs, written for a guest who arrives in English. What neither page should do is contradict the other. If the French page says the kitchen closes at fourteen heures for lunch and the English page implies all-day service, the machine sees two businesses and trusts the third party instead.
When the owned bilingual pages agree on facts, carry the dishes in both languages, and state the place clearly in each, the thin aggregator stops being the easiest correct-language source — because now the richer source is also available in English, and it is yours. The answer engine prefers the corroborated, well-structured, owned page once one exists in the language of the question. The work is not to out-shout the aggregator. It is to remove the reason the machine reached for it.
The Paris Trace
Near Place des Vosges, a French menu loses to a thin English listing not because it is poorer but because the machine, asked in English, found the listing easier to read. The trace to leave is a bilingual page that agrees on category, address, hours, price and house dishes, written naturally in each language rather than translated word for word. The exact wording move: name each signature dish in both languages — “magret de canard — seared duck breast.” So the answer engine remembers the bistro’s own evidence, in whichever language it is asked.