An answer engine does not hand you a list to be judged on. It writes a paragraph and cites whatever it leaned on. Ask one about a Quebec City company in English, then in French, and you will often meet two different companies.
Classic search reporting assumes a list: a board of results, your page on it, a position photographed weekly. That assumption is now only partly true. A growing share of questions is answered above the board, in prose the engine composes, naming a few sources and ignoring the rest.
If you are weighing up a supplier here from elsewhere, this layer probably shaped your first impression — you asked a model who the serious players were before opening a website. If you are the anglophone manager inside a francophone firm, the discovery arrives from the other side: the English answer reads well, and the French one, the version your customers see, describes a business that is not quite yours.
What follows is what the layer is, what can honestly be measured about it, and what the generative research views in the Semalt panel do with the question — including where the measurement stops.
When the result is written instead of listed
The mechanical change is small; the consequence is not. An engine used to match a query to documents and order them. Now it reads several documents, composes an answer in the language it was asked, and offers citations underneath. The reader gets a conclusion instead of a shelf, and the click becomes optional.
That breaks a chain twenty years old. Your position can be unchanged while the number of people reaching you falls, because whoever would have clicked already has what they wanted. No rank column explains it, and no click column contradicts it.
The second change bites hardest in a bilingual city. A list is a list of addresses. A generated answer is prose about you, in whichever language it was asked, and a complete opinion can form without anyone visiting a page you control.
Citation and ranking are not the same achievement
It is tempting to treat citation as ranking with extra steps. The two come apart constantly: a page can hold third place all quarter and never be named in the answer above it, while a page ranking nowhere near the top gets quoted for the one sentence that settles the question.
Nobody publishes the selection criteria, so this describes observed behaviour rather than a specification. What gets cited tends to be extractable — a definition, a figure with a date, a list of conditions. What ranks can win on reputation and topical fit without a single liftable sentence in it.
| Property | Classic ranking | Citation in a generated answer |
|---|---|---|
| Unit of competition | The page against other pages | The passage against other passages |
| What wins | Relevance, authority, links | A clear, checkable, extractable statement |
| Where the reader ends up | On your site, usually | Often nowhere — the answer sufficed |
| Official measurement | Impressions, clicks, position | None published by any engine |
| Language sensitivity | Moderate — one page can serve both | Severe — the source pool changes entirely |
What an answer actually gets assembled from
Generated answers are rarely built mostly from the company's own website. Your site supplies the factual spine — address, services, year of founding. Tone, comparison and judgement come from elsewhere, and the elsewhere is consistent from one query to the next.
Encyclopaedic and directory entries
Stable, heavily linked pages that state facts without selling anything. Models lean on them for the skeleton of an answer.
- Check what yours says
- Stale entries persist for years
Review platforms and discussion threads
Where the qualitative language originates. A few detailed complaints can shape the adjectives used about you.
- Read the threads, do not argue there
- Answer the recurring question yourself
Registries, orders and chambers
Professional bodies and licensing registries carry weight, and are often what a French question reaches first.
- Confirm the listing is current
- Match your wording to it
Regional press and trade coverage
A company nobody writes about has no material behind it, and the gap fills with sector generalities.
- One real piece beats ten mentions
- French coverage counts separately
Read that list again with language in mind: every category exists twice here, once in an English-language internet built on Canadian and American material, once in a French-language one built on France. The two barely overlap, and the engine draws from whichever the question used.
- The spine is yours. Credentials, service names and dates are taken from your own pages when those pages state them unambiguously.
- The judgement is not. Comparative claims — better for small businesses, expensive, slow to answer — come from third parties almost without exception.
- Absence is filled, not left blank. Where a fact about your firm is missing, the answer generalises from your sector rather than admitting the gap.
Ask it in French and the sources change entirely
Run the experiment first; it takes four minutes and beats any argument. Take a question a real prospect would ask. Ask it in English, note the sources; ask the equivalent in French, note those. For most businesses here the two lists barely overlap.
The English answer pulls from Canadian directories, North American review platforms and your own English pages, thin as they may be. The French answer pulls from French-language material — most of which was written in France, for readers in France, about the French market. Not translated Canadian content: native content about another country that shares the language.
The consequences are concrete. A French-sourced answer about vehicle insurance describes a market where bodily-injury coverage is a private product, because in France it is; here that sits with a public plan and the private policy covers something narrower. An answer about changing providers may cite consumer rights that exist under French law and not under Quebec law.
Vocabulary is the quieter half. The everyday professional word differs across the Atlantic often enough that an answer can be fluent, confident and internally consistent while using terms no client in Lévis would recognise. Asked about financial advice in French, a model reaches for the job titles used in France; the regulated title here is a different one.
| What is asked | Where an English answer looks | Where a French answer looks |
|---|---|---|
| Who the local providers are | Canadian directories and listings | French-language aggregators, often France-based |
| How the service works | Provincial and Canadian explainers | French explainers describing French rules |
| What it costs | Canadian ranges, in dollars | French ranges, in euros |
| What the profession is called | The Canadian English term | Frequently the French-from-France term |
| Whether you are reputable | English reviews and coverage | Whatever French material exists — often almost none |
Accurate in English, misdescribed in French
For a restaurant, a French answer written from French sources is an annoyance and a lost enquiry. For a company in insurance or financial services — a large share of the employment base here — it is a different category of problem, and it does not belong to marketing.
Those sectors are supervised provincially: what a firm may call itself, what it may claim about coverage, how a product may be described to a consumer. Now picture an answer engine, asked in French, naming a product category that does not exist here, attaching a cancellation right drawn from French statute, and stating a coverage boundary that is wrong on this side of the Atlantic — while citing your own site among its sources.
Coverage described under the wrong system
The public and private split for road accidents here has no French equivalent, so the answer draws the line in the wrong place.
- State the boundary in French
- Name the supervising authority
Products and titles that do not exist here
Registered savings vehicles and regulated advisory titles differ entirely, and generic French sources supply the French ones.
- Write local product names in full
- Never rely on an acronym alone
The practical response is the only one available. You cannot edit the answer. You can change the material it is composed from: publish, in Canadian French, on pages you own, the statements that keep being got wrong — what the product is called here, what it covers, which rules apply, which authority supervises it. Those pages force nothing; they give the model something local to reach for.
How a visibility estimate is built, and what it rests on
If no counter exists, how does any tool produce a visibility figure? By construction rather than measurement. Assemble questions a customer might realistically ask, put them to models, record which domains were named and how prominently, repeat on a schedule. The score summarises that sample and nothing else.
Generative market research in the panel
For teams wanting an outside read on how a domain is described.
- A competitiveness score with a market circle. Rivals grouped into top-tier, mid-tier and niche bands rather than one ordinal position.
- Model-generated market context. A written read on positioning, traffic and openings for one domain — an outside opinion, not a measurement of your account.
- Query research with intent classification. Candidate questions sorted by what the asker appears to want: the raw material of the sample.
- Pages flagged as levers. Documents the model marks as worth expanding or worth linking internally, plus competitor strengths and content gaps.
- A global visibility value. One figure summarising the domain's standing across the sampled generative landscape.
State the method and its weaknesses declare themselves. Change the question set and the number moves while your company does not. It depends on the language of those questions, which here is the difference between two realities. And it depends on model versions updated to a schedule nobody outside controls, so a drop can be a change in the model rather than in your standing.
What the estimate is good for is direction under fixed conditions — same questions, same language, same interval, read across a quarter. A single value quoted on its own tells you almost nothing.
What to publish so a machine can quote you
The content advice is old advice with a new reason: write the facts down, in the language your market reads, on a page you own, in sentences that survive being lifted out of context. Most business sites here fail that test because the facts live in a PDF, behind a form, or in the heads of the people answering the phone.
- One question, one page, the answer near the top. A passage that answers before it elaborates is the only kind that extracts cleanly.
- Local terminology, written the way clients say it. The Quebec professional term, the Quebec product name, the supervising authority — stated, not implied.
- Numbers with dates and scope attached. A figure without a date is unusable to a careful model and unconvincing to a careful reader, for the same reason.
- French pages written, not translated at the last minute. A page composed in French for a Quebec reader outperforms a rendered English one, because the vocabulary matches the query.
The panel contributes at the level of prompting rather than authorship. Pages flagged as expansion levers, gaps where a competitor is consistently the source, and the query research with its intent labels produce a short list worth writing. On a bilingual site, build that list twice — and the French one will be longer.
Keeping the findings in one place
For projects where research, reports and to-dos otherwise scatter across four inboxes.
- One chronological feed per project. Answers, generated reports, new placements with donor authority and traffic, to-dos and campaign notices in a single stream.
- Bound to the project's real data. A routing model decides per question which data blocks to load — none, one, two or three — instead of answering from nothing.
- To-dos in three states. Active, deferred or discarded: enough structure for a content list nobody wants to keep in a spreadsheet.
- Full-text search and batch input. Every message searchable, and keyword or URL lists handed over in bulk rather than one at a time.
Putting it in the monthly report without overclaiming
Reporting is where the damage gets done, because a plausible number in a tidy table acquires authority it never earned. One rule prevents most of it: verified and inferred figures never share a section, an axis, or a headline.
| Line | Where it comes from | How to present it |
|---|---|---|
| Clicks, impressions, CTR | Search Console, counted by Google | As measurement, with the two-day lag noted |
| Positions and competitors | Rank sampling at fixed intervals | As measurement of a sample, method stated |
| Generative visibility score | A question set put to models | As an estimate, labelled, with the question set named |
| Market context write-up | Model-generated prose | As an outside opinion, quoted, never paraphrased as fact |
Frequently asked questions
Can I find out how often an AI actually cited my company?
Not authoritatively. No engine publishes citation data the way Search Console publishes impressions, and no third party sees the full picture. Every figure you are shown comes from someone asking a set of questions and recording the answers — a sample of a moving target, which should always travel with its question set attached.
Why check the French answer separately?
Because it is composed from different sources, and most French-language material on the web was written in France about the French market. The English answer about a Quebec City firm can be accurate while the French one describes products, rules and titles that do not apply here — and your local customers ask in French.
We publish everything in French already. Is that enough?
It is the necessary half. The other half is everything about you that you did not write: directory entries, association listings, review platforms, coverage. Where French-language third-party material is thin, a model fills the gap from generic French sources. Correcting those listings in French has a disproportionate effect here.
Should this number go to our board or our regulator?
No. It is an estimate built on a self-defined sample, with no auditable method and nothing external to reconcile against. In a governance document it will be read as a measurement, and you will not be able to defend it under questioning. Use it internally; send the verifiable click and position data upward.
How long before publishing in French changes anything?
Plan in quarters. The four-to-eight-week window for first measurable movement applies to the ranking layer; the generative layer also depends on models being retrained on a schedule nobody outside can see. Read direction across a quarter and ignore weekly variation.
None of this displaces work that was already worth doing. Clear pages, correct listings, real coverage and honest facts served you when the result was a list, and they are the only levers now that it is a paragraph. What is new is that a machine introduces your company to strangers out of material you did not write — and in this city it does that twice, once per language.
Check both. Decide which French-language facts about your business need to exist somewhere a model can find them, and treat the score attached to that work as an inference: useful for direction, worthless as a headline. Our service pages describe how this fits a French-first publishing process, and the competitor and market-context views show where a rival is already the source.
To see the six research views against your own domain, beside the Search Console figures you can verify, connect a property and open the dashboard. Run the assessment once in each language, keep the two results apart, and within an afternoon you will know whether your French problem is a content problem or a citation problem. The rest of the series sits on our blog, beside the full feature overview.