Answer engines do not rank you — they decide whether to quote you. For an Ottawa association, research body or professional-services firm whose credibility is the product, that is a different question from position tracking, it behaves differently in English than in French, and it has to be reported carefully: here an unexplained number ends up in front of a committee that will ask where it came from.
What changes when an answer is generated rather than listed
A classic results page is a list of options. The searcher scans a column of links, judges which one answers the question, and clicks. Every step is observable: the query, the position, the click, the page that received it. Search Console reports it with a two-day lag and you can put it on a slide without hedging.
An answer engine does something structurally different. It retrieves a set of documents, writes a single passage of prose from them, and attaches a few source markers. The reader gets the answer without the list. A handful of documents are named; everything else retrieved shaped the wording and received no attribution. There is no position, no impression count, no click to reconcile.
For a retailer that is an inconvenience. For an organisation whose authority is the business model it is sharper. An association publishes practice guidance precisely so that it becomes the reference point; when a model restates that guidance and credits a secondary commentary instead, the association has supplied the substance and lost the attribution. A research group at one of the capital's universities can find its own findings handed back under someone else's name.
Being cited is not the same as being ranked
A ranking is a property of a query and a page on a public results page. Anyone can reproduce it, a rank tracker can record it on a schedule. A citation is a property of a passage inside one generated answer, produced once, for one phrasing of one question, possibly by a model version that will be replaced next month. Ask the same question three ways and you may get three different source lists.
The unit of analysis also changes. Ranking works at the level of the page: a URL competes for a keyword. Citation works at the level of the passage: a paragraph, definition or table selected as evidence for one clause of the answer. That is why a well-optimised page can rank first and never be quoted, while a short definition page ranking eleventh is pulled into answers repeatedly.
| Aspect | Classic ranking | AI citation |
|---|---|---|
| Where you observe it | Public results page, reproducible | One generated answer, not reliably repeatable |
| Official data source | Search Console, SERP tracking | None published by any engine |
| What you can prove afterwards | Position on a given date | That you were quoted once, if someone saw it |
| Suitable for a tender annex | Yes, with the source named | No |
The two questions therefore need separate answers in your reporting. "Are we findable?" is answered by rank tracking and Search Console. "Are we the source the machine reaches for?" is answered, at best, by inference. Merging them into one number is where credibility is lost.
- You can rank first and go unquoted. Depth helps ranking; extractability helps citation. A page can have one without the other.
- A citation is not a session. Being named produces attribution, not necessarily traffic. Treat it as a reputation signal, not a demand channel.
- Repeatability is weak. Two colleagues asking the same question minutes apart can see different sources. Never build a trend line from a handful of manual checks.
Which sources answer engines reach for
No engine publishes its selection criteria, so everything here is a pattern observed from outside rather than a rule you can hold anyone to. The patterns are consistent enough to act on, and they favour the organisations that cluster in the National Capital Region: bodies publishing primary material rather than commentary on someone else's.
The recurring characteristics are unglamorous. Pages that state a thing plainly in the first two sentences get quoted more than pages that build to a conclusion, and pages that define their terms more than pages assuming the reader knows. A visible date, a named issuing body and a stable address help. Material corroborated elsewhere is reached for more readily than material that exists in one place.
Definitional passages
A short, self-contained paragraph that answers one question completely.
- Answer in the first two sentences
- One question per heading
Primary documents
Guidance you issued, data you collected, standards you maintain.
- Named issuing body
- Plain HTML, not only a PDF
Durable URLs
An address that has not moved in three years accumulates the references that make a source look reliable.
- Redirects preserved through rebuilds
- Language versions linked to each other
Second-party references
Being named by another organisation beats a second page of your own saying the same thing.
- Member and partner directories
- Coverage that links, not just mentions
English and French are two different citation markets
This is where Ottawa stops resembling other Canadian markets. A bilingual organisation here is not running a main site with a translated courtesy version. It runs two publications searched by the same people, often on the same day, and answer engines treat them as separate pools of source material.
The volume of English-language material available to these systems is far larger, and English pages tend to be corroborated by more external references. A question asked in French frequently pulls sources from France rather than Canada, because there is more of that material and it is more densely interlinked. A federal-adjacent association can be the reference in English on a topic and effectively invisible in French on the same one, with a European source holding the position its own French guidance should occupy.
That asymmetry is invisible to a country-level filter: both versions are Canadian and both audiences sit in the same region. The language split carries the analysis, and it belongs in the reporting from the start rather than bolted on when someone in Gatineau notices the gap.
| Check | English side | French side | Why it differs |
|---|---|---|---|
| Query set | Terms as your sector uses them | Canadian French terminology, not a translation | Recast phrasing changes retrieval entirely |
| Competing sources | Mostly domestic and US bodies | Often France-based publishers | Depth of available material |
| Typical failure | Quoted but not visited | Not retrieved at all | Different problems, different fixes |
If your organisation carries an official-languages obligation there is a second reason to look at both sides. A visible gap between the versions is not only a marketing problem; it becomes awkward when someone asks why the French guidance is harder to find.
How a visibility estimate is built, and what it is worth
Since no engine exposes citation counts, every product reporting "AI visibility" is producing an inference, and it is worth understanding its shape before deciding how much weight it can carry. The AI Analytics views in the Semalt panel are a fair example: six views producing a competitiveness score and a global visibility value for a domain in the AI search landscape.
What goes into it is a model's reading of your domain against a query space. The panel generates market context — positioning, a traffic estimate, where the opportunities appear to be — and places competitors on a "market circle" of top-tier, mid-tier and niche. It runs AI-assisted query research, classifies intent, flags pages it considers levers for content expansion or internal linking, and analyses competitor strengths against your gaps. The output is coherent and often useful. It is still an opinion about your domain, not a tally of times you were quoted.
The useful way to hold such an estimate is comparative and directional. Two domains scored by the same method on the same day can be compared; the same domain in April and again in September can be read as having moved. What the number cannot do is stand alone as a quantity, or be set against a figure from another vendor.
Alongside it sits the measured part of the panel: Search Console analytics across eight views, keyword dynamics tracking entries into and exits from the top three, ten and thirty, and SERP tracking with average position, ranking score and competing domains. That material carries a source. Keep it separate, and the estimated layer becomes safe to use internally.
Content suggestions that survive a review cycle
Most advice about optimising for answer engines assumes you can edit a page this afternoon. In the capital that is usually wrong. Text passes through committee review, official-languages review and, in government-adjacent work, legal sign-off. A recommendation that cannot survive six weeks of that is a wish. The work has to be batched, justified in writing, and framed in terms a reviewer will accept.
Fortunately the changes that matter most are the least contentious, and several travel through review on arguments that have nothing to do with search.
Front-load the answer
Put the conclusion in the opening paragraph.
- One claim per heading
- Definitions written out, not implied
- No dependency on earlier sections
Make provenance visible
Date, version and issuing body on the page itself.
- Last-reviewed date, not just published
- Named department or committee
- Reference number where one exists
Liberate the PDFs
Guidance locked in a PDF is retrieved less reliably than the same text in HTML.
- HTML canonical, PDF as download
- Headings as real headings
- Tables as tables, not images
Close the French gap
Recast rather than translate, and build the French side its own references.
- Canadian French terminology
- Reciprocal language links
- French-language directory listings
Automation helps produce the shortlist and the justification, not the publishing. The two tiers differ mainly in how much control you keep over that step.
AutoSEO — the automated baseline
For organisations wanting a continuous shortlist of keyword and on-page candidates without running the process by hand.
- Automatic keyword discovery. The pool is fed from Search Console, live SERP data and your own seed terms; each candidate is approved, rejected or deferred individually.
- On-site AI suggestions. Page-level recommendations you can take into a review cycle as a written proposal rather than an unexplained change.
- Full analytics and live chat. Search Console and SERP reporting in one panel, with the Stream assistant bound to real project data.
FullSEO — with a human review gate
For bodies where nothing reaches the live site without a named person approving it.
- Human-review mode for on-site changes. Every proposed edit waits for approval, which is what makes the tier compatible with committee and official-languages review.
- Manual keyword selection with fallback. You choose the terms that matter to your mandate; automation fills the remainder rather than overriding you.
- A team behind the automation. Specialists, developers and writers, plus manual link placement against a domain-rating target.
Prices are in US dollars. If your budget is set in Canadian dollars for the April-to-March fiscal year, convert once at the rate you plan to use and mark it clearly as approximate — a converted figure that drifts between documents is its own credibility problem. Our services overview sets out how this is scoped for a bilingual site.
Folding this into ordinary reporting without overclaiming
The temptation with a new metric is to give it a tile on the first page, because it is the interesting number. Resist that. An AI visibility estimate belongs as supporting context in a report whose headline figures are measured. The fastest way to damage a reporting relationship here is to show a number to people who ask where it came from and have no answer.
Internally, the fix is provenance labelling: every figure gets a source and a method beside it, in the same typeface, not in a footnote nobody reads. The report builder exports CSV and JSON up to ten thousand rows and PDF up to two hundred and fifty, with your own logo and colours — enough for a layout where measured and estimated sit in visibly different sections.
| Figure | Origin | Where it belongs | How to label it |
|---|---|---|---|
| Clicks, impressions, CTR | Search Console, two-day lag | Anywhere, including tenders | Source and date range |
| Average position, keyword dynamics | SERP tracking | Anywhere, with the tracker named | Tool, location, device |
| AI competitiveness score | Model inference | Internal planning only | Estimate, method named |
| Global AI visibility value | Model inference | Internal planning only | Estimate, directional |
- Report the two languages separately. One combined figure hides the asymmetry that matters most here. Two columns, always.
- Report movement, not level. "Up two bands since April" survives scrutiny; "our AI visibility is 63" invites a question you cannot answer.
- Write the caveat once, in the report. A standing note on what the estimate is and is not saves the same conversation every quarter.
In practice the estimated layer is a content-planning input: it tells you which topics you are thin on and which pages are worth expanding. Those become backlog items with owners and review dates. What reaches the committee is the backlog and the measured results, not the score behind them.
Frequently asked questions
Can I find out how often an answer engine has cited us?
Not reliably. No engine publishes citation logs, and there is no equivalent of Search Console for generated answers. You can sample manually — a fixed set of questions on a fixed schedule — but that is a small sample of a non-repeatable process. Worth doing as a qualitative check, worth nothing as a statistic.
Does classic SEO still matter if answers are generated?
Yes, and more than the framing suggests. The documents an answer engine retrieves are largely the documents that rank, so ranking work remains the foundation. What changes is that ranking alone no longer guarantees the visit, and extractability — plain statements, clear definitions, visible provenance — now earns something it did not before.
We publish in both official languages. Where do we start?
Measure the two sides separately for the same topic set; most organisations find a gap they did not know about. Where the French version is weaker, the usual causes are literal translation that misses the terminology people actually search, and an absence of French-language external references. Both are fixable, the second more slowly.
Can we put an AI visibility score in a tender response?
No. Evaluators verify what they are given, and a figure whose method belongs to a vendor will not survive that. Use measured figures with named sources.
How long before any of this shows up in the numbers?
First measurable movement in the conventional metrics typically appears after four to eight weeks, and that assumes changes actually go live. With committee and official-languages review, the approval cycle is usually the longer part. Plan that path before the content work, not after.
Where to start
The sequence is unremarkable. Establish the measured baseline first, in both languages, so anything you say later has a floor under it. Read the estimated layer as a map rather than a scoreboard, to decide where the content effort goes. Then run the changes through your real review process, batched and justified in writing, so six weeks of approval buys more than one paragraph.
For an association, an institute or a professional-services firm in the National Capital Region the point is narrow: you are already the primary source on your subject, and the work is making that legible to a system that cannot ask you. Clear opening paragraphs, visible provenance, durable addresses, HTML versions of what matters, and a French side that is written rather than translated. None of it is exotic, and all of it is defensible in front of a committee.
If you want the measured and estimated layers side by side without stitching four tools together, the unified panel covers campaign automation, Google data, market research and indexing in one workspace, with per-site sharing so a communications lead and an external consultant see the same figures. You can open the dashboard and connect a property to see what your baseline looks like before committing to a plan. More on bilingual measurement is in our other articles, and the indexing tools are worth a look if you have a large document library that search engines are slow to pick up.