Ask a machine what a Montreal studio does and you get a paragraph, not a list of ten links. Ask the same thing in French and you often get a different paragraph, assembled from different material, describing what sounds like a different company.
Something has been added on top of ranking rather than replacing it. The results page still exists and the analytics you already read still describe it accurately. Above all that sits a layer where a model composes an answer and names a few sources, and nothing you own reports on what happens there.
That layer is worth understanding before it is worth measuring, and worth measuring long before anybody quotes a figure about it in a meeting.
An answer has no second place
A ranked result is a claim anyone can check. Seventh is seventh: you can open the page, capture it, argue about why. A generated answer is a passage a model assembled from material it judged relevant, and it exists in that shape once. Ask again an hour later and the wording moves.
The consequence lands on whoever is not named. On a results page, tenth still occupies a line somebody can scroll to. In a generated paragraph there is no tenth line — three or four sources are named, and the rest of the field is not so much beaten as absent. Nothing in your reporting separates those outcomes: both look like a query you received no click from.
| Dimension | A ranked list | A generated answer |
|---|---|---|
| Participants visible | Ten or so, in a fixed order | Usually three to five, unordered |
| Stability of the output | Comparable day to day | Reworded on repeat, sources may change |
| What second place looks like | A line below the first one | Nothing at all |
| Who wrote your description | You did, in your own title and snippet | A model did, from whatever it found |
That last row is why this matters commercially. Somebody comparing five suppliers discounts each self-description appropriately. Somebody handed one composed paragraph reads it as neutral background, because it does not sound like marketing — and it is increasingly the only account of your company anyone sees.
Being cited is not the same as being ranked
These get confused, usually in the direction of comfort: a team ranking well assumes it is cited well. The two run on separate inputs, and a company can be strong at one and invisible in the other.
- Ranking rewards the page. A specific URL competes for a specific query, and everything about that page decides the outcome.
- Citation rewards the statement. What gets pulled into an answer is a passage asserting something checkable, and it may sit on a page that has never ranked for anything.
- Citation is often not your page at all. A model summarising your company will happily cite a trade article, a directory entry or a conference programme rather than your own site.
- The unit of competition differs. You compete for a rank against pages targeting your query; you compete for a citation against everything that describes the same subject.
The third point reorders priorities. Most of the effort a company spends on how it is described goes into its own site, the one surface it controls — while the material a model reaches for is disproportionately the material it does not control.
What an answer actually gets built out of
The material that ends up quoted has a recognisable character. It states things plainly, it survives being lifted out of its surroundings, and usually more than one source agrees with it. Prose written to persuade rarely qualifies; prose written to inform routinely does.
Material that gets used
Passages that survive being taken out of context, because they assert something specific and checkable.
- Plain statements of what a firm makes and for whom
- Dates, credits, named people, stated numbers
- Independent write-ups repeating the same facts
Material that gets ignored
Content that reads well on the page and carries nothing extractable away from it.
- Positioning language with no concrete claim
- Facts locked inside images, video or a slider
- Pages repeating a competitor's wording
Repetition across independent surfaces does most of the work. One sentence on your own site is a single source. The same fact in a trade write-up, a conference programme and a portfolio profile is four, and redundancy rather than eloquence is what makes a model confident enough to state it.
Described in English, thinly described in French
Here the problem takes a local shape most published advice does not anticipate. A Montreal company's reputation and its recruiting frequently live in different languages, and not by accident — the buyers are in Los Angeles and London, the people being hired are in Rosemont and Villeray. Both are correct. The difficulty is that a machine building a description does not merge them.
The two pools barely overlap. In English: trade press, portfolio platforms, conference line-ups, award listings, an interview with your technical director. In French: local business coverage, community write-ups, French-language postings, a school partnership page. Different authors, different facts, wildly different depth.
Where the reputation is written
Sector publications and portfolio platforms covering the work itself, in the language the client market reads.
- Credits, releases and technical write-ups
- Speaker listings and industry programmes
- Profiles maintained by individual staff
Where the hiring is written
Local coverage and community material deciding what a candidate here learns before applying.
- Local business and culture press
- French-language postings and school links
- Neighbourhood and community mentions
Run one question through both languages and the asymmetry is immediate. In English the answer names three of your projects, the year you were founded and a technique you are associated with. In French it runs two hedged sentences, occasionally attached to a similarly named firm elsewhere, and sometimes it declines to say much at all.
Now put a person in front of each. The producer in Los Angeles gets a specific, confident account and shortlists you. The technical artist in Rosemont, deciding on a Tuesday evening whether to apply, gets a vague one — and vagueness reads as smallness. You have not lost a ranking. You have lost a candidate you never knew was looking, in a market where a dozen studios want that person.
| Which pool is thin | Who feels it first | What it costs | Where the fix sits |
|---|---|---|---|
| French, while English is strong | Candidates and local partners | A hiring pipeline that stays narrow for no visible reason | French coverage and a written French account |
| English, while French is strong | Buyers outside Quebec | Being described as a local firm rather than a specialist one | Trade coverage and portfolio presence in English |
| Both, with a strong site | Everyone, quietly | Your wording is the only source, so answers hedge | Independent mentions in either language |
| Neither — covered twice | Nobody | The rare and expensive case | Keep the facts consistent across both |
Work with no agreed name in either language
The second Montreal problem is subtler and hits the firms doing the most distinctive work. A model attaches a company to a subject through language. If the thing you are known for has no settled name, there is no subject to attach you to, and the citation goes to whoever described something adjacent using words that already exist.
This is common here because the interesting companies sit between categories: a post house with a pipeline that is neither previsualisation nor animation, a research spin-out selling a method with one name in the literature and another among practitioners. In one language the term is unsettled. In two it is unsettled twice, and the French form is often three competing renderings, one of them simply the English word.
- Nameless work cannot be attributed. An engine can only credit you with something it can name. Where no noun exists, the question is never asked in a form that could return you.
- Pick one name and stop varying it. The instinct to describe a technique freshly on every page is fatal here. Choose the phrasing and use it identically on the site, in interviews and in postings.
- Prefer the market's word to your own. Adopting a term you find imprecise costs a little accuracy and buys a subject people actually ask about.
- Fix the French form deliberately. Decide whether the borrowed English term or a French rendering is your standard, then hold every French page to it.
Where nothing has settled, the option left is to name the thing yourself and get somebody else to repeat the name — a talk, a technical post, a trade write-up. It is slow, and it is the only route that produces an attributable subject rather than a phrase living on one website.
How a visibility estimate is built, and what it rests on
Because no counter exists, any figure here is constructed, and knowing how it is constructed separates a useful indicator from a decoration.
Six views over a modelled market
What the module produces, and which part is inference rather than record.
- Competitiveness score and Market Circle. A standing for the domain, rivals grouped as top tier, mid tier and niche.
- Model-generated market context. The system describes your domain back to you: positioning, an estimate of traffic, and where it believes the openings are.
- Query research and intent classification. Candidate questions with intent sorted, which matters more than volume when the unit is a question.
- Leverage pages and content gaps. Pages flagged as worth expanding or linking internally, plus subjects nobody on your side covers.
- One visibility figure for the domain. A global value across the AI search landscape, useful as a direction of travel and nothing more.
Underneath, the method is inference from a sample. Questions are put to models, the sources named in the replies are recorded, your domain is checked against them, and the result is weighted against the competitor set the system believes you belong to. Change the question set, or the day it ran, and the figure moves without anything about your company having changed.
Read as a direction over quarters, next to your own question log, it tells you whether you are becoming easier to describe. Read as a monthly indicator with a target, it produces an argument nobody can win.
What to publish, and how to report it honestly
The content side is unglamorous and consists mostly of stating what you assume everybody knows. Somewhere there should be a page saying plainly what the company makes, for which industries, since when, from where, at what scale — in both languages, same facts in each. Most firms have this scattered across an interview, a careers page and a slide deck, and nowhere a machine can lift it from.
- Put the checkable facts in one place. Founding year, headcount band, disciplines, sectors served, the technique you want associated with you — stated, not implied.
- Write the French version as a source. A near-copy adds no independent corroboration and reads as thin to the audience that most needs it.
- Get the facts off the site as well. A trade write-up, a technical post under a named author, an accurate portfolio profile.
- Keep names and numbers consistent. Contradictory founding years across surfaces make a model hedge, and hedged answers name somebody else.
- Use the flagged pages as a queue. Leverage and content-gap views point at pages worth expanding; treat them as candidates, not instructions.
On the reporting side, keep this in its own compartment: a short labelled section giving the question set, the date, the languages, who was named, and what changed since last quarter. Beside it sit the measured numbers — clicks, impressions, positions — which carry a different kind of authority and should not be diluted by proximity.
If the area feels like it needs a permanent owner, it does. A quarterly question log, two languages tracked separately and a vocabulary decision held across a dozen surfaces is a standing job. The generative research views carry the observation side; somebody still has to read them and decide.
Handing the standing job to a campaign
For teams that will not sustain a quarterly pass in two languages internally.
- AutoSEO, 149 USD monthly per domain. Terms found and prioritised automatically, links built without your involvement, and the research views described here.
- FullSEO, 500 USD monthly per domain. Manual term selection with automatic fallback, placement held to a domain-rating target, and specialists, developers and writers behind the account.
- Stream, the assistant in My SEO. One feed of answers, reports, new links and open to-dos, tied to the project's own data.
Frequently asked questions
Can we find out how often we are cited in AI answers?
Not exactly, and nobody else can either. You can fix a question set, run it on a schedule, record which sources appear and watch the pattern move. That is a repeatable observation, and worth having. It is not a count, and calling it one will eventually embarrass whoever did.
Our English answers are detailed and our French ones vague. Where do we start?
With the facts rather than the pages. Write one plain French account of what the company does, carrying the same figures as the English one, then get at least one source other than your own site to repeat them. Corroboration in French is the scarce ingredient.
Does ranking well still help if answers are generated?
Yes, twice over. The results page has not disappeared and a large share of searches still ends in a click. Beyond that, material that ranks tends to be material that gets reached for. Ranking work is not wasted, only no longer the whole picture.
A model described our work inaccurately. Can we correct it?
Not directly, and there is no form to submit. What you can change is the material it drew on: find the sources stating the wrong thing, get them corrected where possible, and publish the right version in enough places that it becomes the better-corroborated one. Months, not days.
We do something the industry has no name for. Is that worth solving?
Only if people are looking for it. Where clients arrive by referral, a name matters less. If you want to be found by people who have the problem but not your phone number, they need a phrase to type and a machine needs a subject to attach you to. Take the closest term the market already uses before inventing one.
A modest place to start
Nothing here rewards a big first move. Run the two question sets and keep the replies with a date. Write the plain page of facts in both languages, decide once what your distinctive work is called in each, and spend a quarter getting one or two independent sources to repeat it. Then run the questions again and read the difference rather than the score. The market circle and competitor views will tell you which ring you are compared inside, which is useful context for how ambitious that step should be.
Keep the reporting proportionate to the evidence. A quarterly paragraph on what changed in how machines describe you, beside the measured search data rather than above it, is defensible in front of anyone. A headline number with a target is not, however confident the software sounds. More on this in the work we take on and across the rest of the blog.
When you want the ranking data, the competitor set and the generative market context under one roof rather than in three exports, connect your own domain to the panel and give it a quarter before drawing conclusions. What to watch is not a rising figure. It is the moment the French answer stops being shorter than the English one — because in this city that gap is not a measurement problem. It is a hiring problem wearing a measurement costume.
