Do assistants answer the same Canadian buying question differently in French and in English, and how to find out for your own category

Treat them as two separate measurements, because that is what they are. The market and the language both live inside the text of the question rather than in any setting, so a French question about a Quebec buyer is a different question with a different source pool, and running it through an English question set measures the wrong thing. We have measured the English Canadian set, 64 blind readings of ChatGPT and Gemini on 22 October 2026, and we have not yet run the French counterpart. Rather than infer a French result from an English one, this page gives what the English reading establishes, the reasons the French answer should be expected to differ, and the test that settles it for your own category in an afternoon.

Why the two should be expected to differ

• The source pool is different. A French language question draws on French language documents, and in Canada that means Quebec publishers, Quebec professional bodies, and the French versions of federal properties rather than the English ones.
• The legal frame is different, and it is published in French. Quebec's Charter of the French Language, strengthened by the 2022 amendments, governs commercial advertising and public communications in the province, and Quebec's Law 25 adds privacy obligations beyond the federal regime. Those documents and the commentary around them exist primarily in French.
• Supplier markets differ. A Quebec buyer's shortlist is often Quebec firms, and those firms publish in French. An English reading will under represent them by construction.
• The engines' live search behaviour differs by engine already. On our English Canadian reading ChatGPT ran a live web search on all 32 of its readings and Gemini on 20 of its 32, so language is a second axis on top of an engine axis that is already uneven.

What the English Canadian reading establishes

A 198 question Canadian buyer bank in 33 batches, put to both engine surfaces through a data provider on 22 October 2026, with Canada named inside the text of every question, because the provider accepts gl, hl and location parameters and ignores all three. Batch 33 names us in the question and measures a ceiling, excluded from every figure.

Reading	ChatGPT	Gemini	Both
Blind readings	32	32	64
Ran a live web search	32	20	52
Named a rival and not us	9	17	26
Named nobody at all	23	15	38

The engines produced 2,992 citation URLs across 974 distinct hosts. The most cited hosts were developers.google.com in 17 readings, support.google.com in 14, torontoseo.com in 13, clutch.co in 12, canadacreate.com in 11, canada.ca in 10, wisemedia.io in 9, semrush.com in 9, competition-bureau.canada.ca in 9, ised-isde.canada.ca in 9, blog.google in 8, www150.statcan.gc.ca in 8 and ottawaseo.com in 7. One entry, ccn.web.tr, is the data provider's own fetch infrastructure rather than a publisher and is excluded from every source analysis.

Notice what is absent from that list: any French language Quebec property. That absence is a property of the question language rather than a finding about Quebec, and it is the clearest available argument for running the French set separately.

The test, run in an afternoon

• Take twelve buying questions in your category and write each one twice. Once in English naming Canada or the province, once in French naming Quebec or the city.
• Have the French written by somebody who writes for a Quebec reader rather than translated, because a translated question often asks something slightly different.
• Put all twenty four to both engines and save the full answer text and source list for each, with the date, the engine and the language recorded.
• Score each language separately. Named you, named a rival, named nobody, with its own denominator.
• Compare the host lists rather than the scores. The overlap between the two lists is the useful number, because a small overlap means two programmes and a large one means one with a second language layer.

What the result changes

If the French and English host lists	Then
Barely overlap	You have two programmes. Budget French as its own publishing and outreach plan, not as translation
Overlap substantially	One programme with a French layer. Prioritise French versions of your highest value pages
Return Quebec firms in French and national ones in English	Your Quebec competitive set is different from your national one, which changes the comparison pages you write
Name nobody in either	The slot is open in both languages, which is the cheapest position to be in and the one to move on first

The compliance half, which is not optional

For any company selling into Quebec, French is a requirement before it is a measurement question. The Charter of the French Language sets requirements for commercial advertising and public communications in the province and the 2022 amendments strengthened them. Separately, Quebec's Law 25 imposes privacy obligations that touch consent for tracking and the handling of personal information, over and above the federal PIPEDA regime. And CASL applies to commercial electronic messages in both languages, requiring consent, sender identification and a working unsubscribe honoured promptly. The practical version is that a French programme needs a reviewer who reads French and knows Quebec, which is the same person who should be clearing the French copy anyway.

What to do first

• Decide in writing whether Quebec is a market. If it is, French stops being optional at every level of this.
• Run the twenty four question test and compare the host lists rather than the scores.
• Fix the language and region declarations, with en-CA and fr-CA used consistently and every French page linked and annotated. The commonest Canadian fault is a French section that exists and is invisible.
• Publish the French version of your single highest value commercial page first, written rather than translated.

If Quebec is a primary market, the French first position and the privacy law that touches your tracking are set out in French first by law, and Law 25 on your tracking. If the language architecture decision itself is still open, it is the first of three that set the scope and the price, in three decisions that set the scope and the price.

Questions about French and English assistant answers in Canada

Do you have the French Canadian figures?

Not yet. We have the English Canadian set, 64 blind readings on 22 October 2026, and we will not infer a French result from it. The test above is the one to run, and it is the same method.

Can I translate my question set?

Have it written rather than translated, because a translated question often asks something subtly different and you are measuring the question as much as the answer.

Will the same suppliers be named in both?

Expect differences, because a Quebec buyer's shortlist is often Quebec firms publishing in French. The host list overlap is the number that tells you how different.

Is French worth it if most revenue is from Ontario?

If you genuinely do not sell into Quebec, write that down internally so the gap is a decision rather than a failure. If you do sell there, French is a requirement.

How do I keep the two readings comparable?

Same question count, same engines, same frequency, separate denominators, with the language recorded per reading. Then the two can be read against each other without being averaged.

Reading two languages as two measurements

The failure mode here is a single Canadian visibility score covering two languages, which would hide the one thing worth knowing. AI Knows Us reads ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews with the market and the language named inside the question, reports each engine separately with its own denominator, records every cited host, and keeps the raw answer text. Outside the United States and the United Kingdom the cheapest plan that reads all five engines is Growth at 199 dollars a month with 100 buyer questions watched, read on aiknowsus.com/pricing and aiknowsus.com/llms.txt on 22 October 2026, and the question count is the line to size against two language sets rather than one.