The essentials
- Semrush, Google Keyword Planner and Google Trends need volume to be accurate. The Quebec market, smaller and more fragmented, has always been read incompletely by them.
- The models behind ChatGPT, Gemini and Perplexity repeat the same mechanism one level up: they reuse what they’ve seen most, which means English first.
- Well-structured Quebec French content is scarce. Being first isn’t enough: you have to be the most reliable, the most precise and the clearest.
To find out how a market searches, you open Semrush, Google Keyword Planner or Google Trends. These tools give an impression of precision. But that precision rests on one thing: volume. The more searches there are, the sharper the picture. The fewer there are, the blurrier, rounder, or simply absent it becomes.
The Quebec market sits on the wrong side of that rule, and that isn’t new. It’s not the fault of any one tool: it’s a consequence of how they’re built. Generative search doesn’t create this problem. It repeats it one level up, in the models’ training data.
Tools built for volume
Google Trends says so itself: it only shows popular terms, and low-volume terms appear as “0”. A term that Quebecers search every week can show up as zero. Keyword Planner groups close variants and, unless an account is spending, shows wide ranges instead of precise numbers. Semrush queries one database per country: for Canada there is only one, with no split by province, and no separate database for Quebec.
Search Console shows the same limit. The page and query dimensions only cover about half of real clicks, because Google anonymizes part of the queries, and that share tends to grow as volumes fall.
Why Quebec takes the hit
A smaller market produces fewer searches, and that volume fragments further: vocabulary variants, anglicisms, queries that switch between French and English, gaps with the French spoken in France. Each term, taken alone, looks tiny, even when the whole represents real customer demand.
The content planning trap
This is where the error costs the most: the editorial calendar. When you plan organic content from search volume, topics that show zero or close to it drop off the list: not enough volume, not worth publishing in Quebec French. But a topic with no measured volume is not a topic with no demand. A precise article, written in the market’s own words, can answer the intent of someone close to a decision, which a generic high-volume query doesn’t always do.
There’s a second reason, specific to generative search. An article no longer only serves to rank for a query: it also serves as a source the models can cite. In this language there is little competing content, so the spot is open even when the tool shows zero.
Displayed volume should be one signal among others, not an exclusion criterion. Before dropping a topic, it’s worth checking whether customers are asking the question (sales, customer service, the site’s internal search), whether real queries exist in Search Console or in campaign search terms, and whether serious French content on the subject already exists.
What it costs in revenue
When a tool shows zero or “not enough data”, the tempting conclusion is that there’s no market. That’s a risky reading. You underinvest in queries that could generate leads, you drop segments that would have paid off, and you build budgets on estimates that were never calibrated for the target market. That cost shows up in no report: it’s revenue nobody went looking for.
AI repeats the same mechanism
A language model that has seen far more English content than French on a given subject doesn’t cite French-language brands less often out of ill will. It does so because, statistically, it has seen fewer of them, and less often structured so that they can be reused as-is in an answer. It’s the same principle as keyword tools: what’s rare in the data is poorly represented in the output.
Most AI content generation and optimization platforms are designed and calibrated in English first. French often arrives as a translation, rarely adapted to the register or to how a French-language market really searches. Applied mechanically to a Quebec site, a recommendation built for English suggests angles that answer different questions from the ones asked here.
An opportunity, if you’re good
The reflex when facing this is worry. The better reading is the opposite: if the models have little material in Quebec French, nobody there has an established advantage yet. In English, a brand fights thousands of equivalent pages. Here, the space is far less contested.
But arriving first isn’t enough. A model tends to reuse what looks reliable, precise and clear, not the first page it finds. Publishing a lot of average content in Quebec French creates no lead: it adds noise to a space where there’s little signal.
Being good means five things:
- Real, verifiable expertise: facts, sourced figures, a clear position instead of generalities.
- The market’s language: the vocabulary people here actually use, not a translation.
- A structure built for extraction: a clear question, a direct answer, clean sections.
- Clear entity signals: who is speaking, about what, and with what legitimacy.
- Consistency between the site and the sources the models consult, and content kept up to date.
It’s the kind of lead we saw built in SEO fifteen years ago: the least contested markets produced the fastest gains for those who arrived with solid work. In two years, that lead will cost far more to catch than to build now.
What if the gap closes on its own?
It’s the obvious objection: models train on more content with each generation, so the gap should shrink by itself. But the English-language web keeps producing far more content than the French-language space, so the next models will see, proportionally, even more English. The gap only closes if well-structured French-language content grows faster than the gap does, which takes deliberate action.
What to do
Three reflexes, in order. First, never read the Quebec market with a single tool. Cross-check Semrush, Search Console, Google Trends and campaign data before concluding: a zero, or an alarming figure in a single source, is often a coverage artifact and not a market signal. Second, start from the real French-language queries of the target market, not from a translation of what works elsewhere, and check which market an AI tool assumes by default before trusting it. Third, treat content structure, the thing that lets a model cite a page with confidence, as an immediate priority and not as an experiment for next year.
One more nuance: Quebec French isn’t interchangeable with the French of France in training data. A model that has mostly seen European content doesn’t always grasp the turns of phrase, the commercial references or the register used here. A brand that thinks it has solved the problem by publishing in French risks producing answers that sound right in Paris and a little foreign in Quebec: the same gap, reproduced inside the French-language space.
The tools have always lagged behind the Quebec market, and they probably will for a while yet. The models will see more well-structured French-language content, and the gap will eventually narrow. The only question that matters is who, by then, will already have taken their place in the answers, and who will have to win it back after the fact.