When a prospective client asks an assistant about your firm, the reply is composed on the spot out of whatever the web happens to say — and your carefully worded About page is one ingredient among several. That is the layer above ranking: influenced rather than controlled, estimated rather than measured.
When the answer is generated instead of listed
A ranked list leaves the work with the reader. Ten links, two lines of preview each, and the decision about which source deserves trust stays with the person searching. A generated answer removes that step: the engine reads a handful of documents, treats some of their contents as settled fact, and returns a paragraph in its own words, usually with a short strip of source links beside it.
Three things follow, and only the first gets much discussion. Clicks thin out on any question a paragraph can close. The unit of competition shifts from a page to a statement: your sentence about how long a review takes now competes with somebody else's sentence about the same thing. And the output is unstable — ask twice from two devices and you may get two wordings and two different source strips.
The third point is the one to hold on to: there is no position one in a generated answer, only a paragraph produced for one person at one moment.
| Dimension | Ranked list | Generated answer |
|---|---|---|
| What competes | A URL against other URLs | A claim against other claims |
| Stability | Same query, near-identical result | Same query, different wording and sources |
| Position | An integer you can track daily | No position exists |
| Measurement | Reported by the source | Inferred from sampling; nothing official |
| Failure mode | You slip to page two | You are summarised, not named |
Being cited and being ranked are not the same event
The two travel together loosely — material an engine retrieves usually ranks somewhere — but they come apart often enough that treating them as one metric will mislead you. A page can rank first and contribute nothing, because it opens with positioning language rather than information. A page well down the list can supply the whole answer, because one paragraph on it states a fact plainly.
A third case should concern a professional-services firm most: your material is used, the answer is accurate, and your name never appears. That is not a fault in the system; it is how summarisation works.
Ranked, not used
Strong position, skipped when the answer is assembled. It opens with ambition rather than facts, so nothing on it can be lifted.
Used, not ranked
A modest page supplies the substance because one paragraph answers the question outright — usually a technical or methodology note.
Used, not named
Your material shapes the answer and the source strip lists someone else. No attribution, no visit, no trace in any tool you own.
Named through a third party
The engine cites a directory or a news archive that describes you — their wording, often the stalest version in circulation.
Which material an answer engine actually pulls from
The mechanism is opaque, but the pattern is consistent enough to plan around. Engines favour material that states something plainly, is corroborated elsewhere, and has not obviously gone stale. Marketing prose fails all three tests at once: no checkable statement, no independent repetition, no date.
- Pages that state facts in the open. Legal entity name, founding year, offices, disciplines, certifications, team size — stated once, plainly, without asking the reader to infer anything.
- Corroboration from outside your domain. An association member listing, a public register, a directory entry, a conference programme. Your own claim is one source; the same claim in two independent places is treated very differently.
- Material with a visible date. A page saying when it was last reviewed beats an undated one wherever currency matters, which for a consultancy is most questions.
- Long-lived third-party text. A 2016 article about a restructuring, an old award notice, a review page, a former employee's public post. None of it under your control, all of it retrievable for years.
That last category is why this is a reputation question wearing a technical costume. Your website is the one source you write; everything else in the pool was written by someone whose interests were not yours.
| Source type | Why an engine reaches for it | What you can do | How fast it changes |
|---|---|---|---|
| Your own facts page | Direct, unambiguous, first-party | Write it, keep it current | Immediately |
| Association and register listings | Independent confirmation of status | Request a correction | Weeks |
| News archives | Dated, editorially filtered | Almost nothing | Effectively never |
| Personal posts by ex-staff | Retrievable text naming your firm | Nothing, and attempts backfire | Never |
The answer to "who are they" is generated, and your firm does not write it
Calgary runs on head offices and on committees. A national engineering consultancy run from one address downtown, a mid-sized accounting practice, a B2B software company selling into three provinces from a single Calgary base — in every case the buyer is several people. A technical lead, a procurement contact, a director who signs, and someone senior who attended none of the meetings and will spend four minutes forming a view.
That fourth person now asks an assistant, and the answer is assembled from the pool above: a directory entry with the address you left in 2019, a news item about a contract that went badly, a review site with eleven entries of which two came from a disgruntled subcontractor, and somewhere in the mix your About page. The engine cannot know which of those you consider representative.
That is what separates this from ordinary ranking work here. Elsewhere search decides whether you are discovered. In this city it mostly decides whether you survive the shortlist — and that can turn on one paragraph nobody at your firm has read.
The prospective client
Checking that a referral holds up before the introductory call. Wants scale, sectors and credentials, and is looking for a reason to say no.
The senior hire
Deciding whether to return a recruiter's call about a firm they do not know. Reads stability signals and weighs old news heavily.
The journalist
Needs a one-line description of your firm on deadline, copies whatever the answer says, and rarely corrects it later.
The prospective partner
A first pass before a longer conversation about working together. Notices contradictions between sources and asks about them directly.
Be honest about what can be done. You cannot control the answer, you cannot instruct an engine to disregard a source, and attempts to suppress third-party material tend to produce more retrievable text than they remove. What you can do is influence the pool: make the verifiable facts easy to find and consistent everywhere they appear, so the plainest and most current version of your story is also the most available one.
How a visibility estimate is built, and what it is worth
Since no counter exists, a visibility figure has to be constructed, always by some variation on one procedure. Take questions a buyer might plausibly ask, put them to a model, record whether your domain appeared and which competitors appeared instead, repeat until one odd answer no longer moves the total, and express the result relative to the field.
That is what the AI Analytics section of the rebuilt Semalt panel does across six views: a competitiveness score with a Market Circle placing top-tier, mid-tier and niche competitors around your domain; a model-generated market context covering positioning, traffic and openings; query research with intent classification; pages flagged as levers for content or internal linking; competitor strengths and content gaps; and a global visibility value across the AI search landscape.
Six views over a layer with no official data
Useful for direction and comparison. Not a meter reading.
- Score and Market Circle. Your domain against top-tier, mid-tier and niche competitors. The mid tier is the band worth arguing with.
- Model-generated market context. Read it as what the web currently says about you, not as an assessment of your firm.
- Query research and intent. Which questions sit around your category, and what the person asking wants.
- Content gaps and page-level levers. Pages flagged for expansion or internal linking, with competitor strengths alongside.
Where it earns its place is direction and contrast. If your score sits below three firms of your size, that is worth a conversation about what they publish and you do not. If it moves the same way over two quarters on an unchanged method, that is a signal. If it drops eleven points in a week, the first hypothesis is that the model changed, not your website.
Making the verifiable facts easy to find
The productive response is not to chase the machine. It is to publish, plainly, what is true and checkable about your firm, and to keep it consistent everywhere your name appears — the only part of this subject fully within your control.
- One page of stated facts. Legal name, year established, offices, leadership, disciplines, certifications, sectors served, headcount. No adjectives, and a line saying when it was last reviewed.
- One clear paragraph per service. Say what the work is and who it is for in the first three sentences. That paragraph is what gets lifted.
- Consistency across profiles. Same legal name, address format and founding year on your site, in directories, on association listings and in staff bios. Contradictions get noticed and repeated.
- Named authors and dated updates. Technical notes signed by a real person, with a revision date, are retrieved more readily than anonymous prose.
- Corrections where the record is stale. Directory entries, old award pages, outdated listings. Slow, dull, more effective than new content.
Two things follow. New and corrected pages must be crawled before anything can retrieve them, which is what the indexing side of the panel is for — the URL and sitemap tools submit through the IndexNow API and log each bot visit, with a daily budget of 1,000 URLs per account and batches of up to 10,000. And on-site suggestions need someone to read them with judgement rather than apply them wholesale.
AutoSEO — suggestions applied on your judgement
For a firm with a marketing lead who can review changes but not write them all.
- Automatic keyword discovery. Candidates from Search Console, live SERP data and your own seed terms, each approved, rejected or deferred.
- On-site AI suggestions. Page-level recommendations you act on selectively — the right posture when pages carry professional claims.
- Analytics and a live assistant. Search Console and SERP reporting alongside Stream, bound to your actual project data.
FullSEO — human review before anything ships
For firms where a wrong sentence is a professional problem, not a marketing one.
- Human-review mode for on-site changes. Nothing reaches a live page without a person approving it. For a regulated practice that is the whole argument.
- Manual keyword selection with fallback. You choose the terms that matter; automation covers the rest.
- Manual placement with a DR target. Directed by hand across a partner network of more than 230,000 sites, with a team of specialists, developers and writers behind it.
Folding this into ordinary reporting without overclaiming
The reporting problem is not technical. An estimate placed beside measured figures inherits their credibility: put a visibility score in the same table as clicks and impressions and nobody remembers which of the three was inferred. The fix is separation — a labelled section, its method stated in one sentence, and no arithmetic between it and anything measured.
Pair it with branded-search data, where the same story appears in numbers you can defend. Impressions on your company name, on named executives, and on your name plus a qualifying word come directly from Search Console. Keyword dynamics — which terms entered or left the top three, top ten and top thirty — carries the same information without inference.
| What you might be tempted to write | What you can actually support |
|---|---|
| "We appear in 34 % of AI answers in our category" | "In a fixed sample of 60 questions, our domain was named in 20" |
| "AI visibility grew 18 % this quarter" | "The estimate rose over two quarters on an unchanged method" |
| "We rank first in AI search" | Nothing. There is no ranking in a generated answer |
| "AI search drove 400 visits" | Referral traffic where the source is identifiable, and no more |
Everything else stays as it is. Exports run to 10,000 rows in CSV or JSON and 250 rows in a server-rendered PDF, the report builder takes your logo and colours, and date ranges default to 28 and 90 days with the two-day Search Console lag accounted for. The AI section sits beside that as commentary rather than a line item — and how it is being read is a fair thing to raise with whoever handles your search work.
Questions that come up in the review
Can we make an assistant stop repeating something inaccurate about us?
Not directly. There is no mechanism for correcting a model's output. What you can change is the balance of the source pool: publish the accurate version plainly and get it corroborated in listings and registers you can influence. Where the inaccuracy sits on a third-party page you can reach, ask for a correction. Where it sits in a news archive or a personal post, it stays.
Is it worth paying separately for AI visibility tracking?
Not as a standalone line item. The estimate is only interpretable next to your ranking and Search Console data, and a tool offering it in isolation is selling a number without its context. Inside a panel that already holds the measured data, it costs nothing extra.
Our score dropped sharply. What happened?
Most likely nothing on your site. The plausible causes in order: the model changed, the question set changed, a competitor published something substantial, or your material genuinely got worse. Check the first two before briefing anyone on the last two.
How long before changes show up?
The same 4 to 8 weeks campaign work usually takes before first measurable movement, and often longer here, because third-party sources must be re-crawled and directories update on their own schedule.
Should we write pages specifically for AI engines?
Write pages that state facts clearly for people and the retrieval benefit follows. Pages written to game a summariser read badly to the senior buyer who opens them, which in a shortlist process is worse than not being cited.
What a defensible position looks like
The honest summary is short. There is a new layer above ranking; it shapes what a committee member reads about your firm before they meet you; you can influence it but not control it; and the numbers describing it are estimates that should be labelled as such wherever they appear.
The practical work is unremarkable: publish the checkable facts, fix the stale record elsewhere, keep the wording consistent, get the pages crawled, and watch the branded queries in data you can defend. Read alongside the rest of our notes for Calgary firms, this is an old discipline with a higher-stakes audience rather than a new one.
To see what this layer says about your own domain rather than read about it in the abstract, the analysis views and the campaign data sit in one workspace: open the panel and run it against your domain, and treat the result as the start of a conversation, not a verdict. The figures you can actually defend live next door, in the Search Console and SERP analytics.