Cited, Not Ranked: How AI Search Visibility Works and What It Is Worth

Ask an engine how to spend three days in Vancouver and it writes an itinerary, not a list of links. A few sources get named inside that paragraph, and being one of them is a different job from finishing third.

The businesses feeling this first are the ones whose customers never really typed keywords. Someone planning a week on the coast asks whether the Sea-to-Sky drive is worth doing in March; someone weighing a move to the Lower Mainland asks what a family pays in Burnaby versus Coquitlam. These were always questions; the search box forced them into three words.

Now the question survives intact and the engine composes a reply from whatever it finds. For a kayak outfitter in Deep Cove, a boutique hotel off Water Street or a clinic in Kitsilano, the question that matters is no longer the site's position but whether its pages supplied the reply. The AI Analytics module inside Semalt's unified panel is one attempt to survey that floor with the instruments currently available.

Customers stopped typing keywords and started asking questions

The ranked list handed the decision to the visitor. Ten options, a title, a preview line or two, and somebody made a pick — visibly, which is the only reason any of it could be counted. A composed reply takes that decision back: the summary settles the matter, and plenty of readers never open a page at all.

• Interest stays, visits thin out. Once two sentences on the results page dispose of a question, it is disposed of. Demand held; the session vanished.
• Questions get long and spoken. "Whale watching Vancouver" becomes "is a whale watching tour worth booking in early May with two kids".
• The clicks that remain are heavier. Someone who still clicks after reading a summary is usually checking availability. Fewer arrivals, each nearer a decision.

That last shift is the one monthly reports read backwards. A third off the click-through rate may mean nothing worse than idle browsing drying up while booking traffic held firm. The position never moved; a floor was added over it.

Seasonality makes the pattern harder to read here than almost anywhere. July trip-planning questions are asked in February; Whistler snow questions start in October and collapse in April; property questions spike on policy announcements rather than on the calendar.

Named in the answer, or placed in the list

Folding the two together is tempting, since a good position makes a page easier for a model to encounter. The mechanics diverge far enough, though, that one blended figure tells you nothing actionable.

Positions are scarce and ordered — a single top slot, every gain somebody's loss. The answer layer behaves more like a round table with room for four or five, since one reply usually pulls from several documents. A chair goes to any passage that finishes a sub-question tightly enough to be quoted whole. A page down at ninth gets used when one paragraph names the crossing time, the months it runs and the fare. A page at second is passed over for offering "unforgettable West Coast experiences".

Read the last column and the pattern is plain: booking intent still runs through the ranked list, while anything phrased as a question migrates upward. Positions attach to keywords; mentions attach to statements. A single paragraph about reserving a ferry sailing can turn up under twenty different wordings while its page tops no list anywhere.

Canada complicates it further. One business here is searched in two languages, and the French side divides again — Quebec usage on one hand, France on the other, two audiences that country-level filtering quietly merges. On most commercial queries a local operator also competes against Seattle, because the border means nothing to a search engine.

Which sources an answer engine reaches for

Nobody outside the providers sees the selection rules. Read a few hundred generated replies in one sector, though, and it becomes clear which kinds of source keep reappearing. Four do most of the work here.

The practical reading: for these seats you compete less with the operator down the street than with a platform holding a structured copy of your inventory, and with a three-year-old forum thread a model finds more concrete than your homepage. Neither can be outranked. Both can be out-specified.

How a visibility estimate is put together

So what is being calculated? The AI layer runs to six views, and the honest description of the method is that it triangulates. Query sets are assembled and classified by intent; the domain's coverage of them is assessed; rival domains in the same question space are scored; the result is compressed into a competitiveness score with a market picture around it.

That picture is the Market Circle: rival domains sorted into a top tier, a mid tier and a niche band. Treating the three as one roster of rivals throws the view away, since each answers a separate question.

• Top tier is a map. No eleven-room hotel is going to outpublish a national booking platform. The outer band reveals which sub-topics a model regards as the backbone of the subject.
• Mid tier is the benchmark. Operators at your scale, frequently from the same stretch of Metro Vancouver. The questions they left unanswered are your opening.
• Niche is the method. A twelve-page site that explains one trail or one treatment from start to finish keeps turning up in the relevant replies season after season, while a generalist with twenty times as many pages is never named.

The next view describes your domain back to you in prose: where it sits, roughly what traffic it draws, what openings the model perceives. Its worth lies in the outside view rather than the figures. A tour operator convinced it is a small-group specialist may have most of its pages given over to one bus route, and the description will say exactly that. Lay the Market Circle alongside the rank-tracking competitor list in the same panel; names appearing in only one of the two are usually the instructive part.

Query research when the unit is a question

The third view is query research done with model assistance, sorted by intent. The older method steered by volume: rank candidates on monthly searches, note the difficulty, start at the top. Two things go wrong with that once replies are composed.

Spoken-length questions barely register any volume. Two hundred variants at fifteen searches apiece disappear the moment a list is sorted, though between them they map a real field of demand — and in trip planning nearly all demand looks like that. Nor did volume ever indicate which questions end in a booking.

Mentions in composed replies come from planning and comparison questions; booking intent stays with the ranked list. Different pages serve each, so the candidate list is sorted by intent before anyone writes a brief. The campaign side of the panel fills its pool the same way — Search Console, live results, seed terms — with each candidate approved, rejected or deferred singly.

Writing pages an answer engine can use

That leaves the practical matter of what a model can actually use. No formula for it exists, but the same handful of traits shows up whenever quoted pages are pulled apart, and most are what has always divided serious writing from brochure filler.

Four properties do most of the work across those rows. Facts that can be checked, because "the crossing takes ninety minutes" is quotable and "we get you there quickly" is not. A single question per heading, answered immediately underneath rather than smeared across four paragraphs. Caveats inside the sentence instead of down in a footnote. And visible currency: a displayed review date pulls surprising weight on pages about seasons and access rules.

Two routes to that work sit in the campaign side of the panel; the difference is who approves the wording.

Whichever stage you use, point the on-site content suggestions at facts you are willing to publish. A standing list of approved specifics — seasons, rate bands, durations, access rules, each owned and dated — lets writers work without opening an approval cycle per sentence. The technical foundation underneath it all is described in our services.

Folding it into the monthly report

The sixth view compresses everything into one global visibility figure across the AI search landscape, spanning an entire portfolio of domains. Its only use is comparison against its own earlier values, since compression is precisely what discards causation. It may rise because the rewrites worked, because a rival's section went stale, or because a model changed on a date outsiders never learn.

• Show it as a trend. Two or three quarters of direction tell you something; one isolated reading tells you next to nothing.
• Use it as an alarm, not a verdict. A steep drop justifies opening the views beneath it. It is not a finding, and certainly not a headline.
• Keep it beside the measured data. Clicks, impressions and positions are the record. This figure is commentary printed next to it.

Practically, separate the two inside the report rather than trusting a footnote. The configurable report builder exports CSV and JSON up to 10,000 rows and renders PDFs up to 250 rows server-side, with your own logo and colours, from time series, metric cards, sortable tables, sparklines and heatmaps. Give the measured metrics their own section, put the estimate in a labelled block after it, and build the caveat into the template.

Competing with platforms, and with a city two hours south

For most businesses here the answer layer sharpens an existing problem. A visitor comparing hotels, a family researching neighbourhoods and a patient choosing a clinic all meet a platform's version of your offering first — structured, complete, uniformly formatted, and often exactly what an engine composes from.

The response is not to compete on breadth but to hold facts a platform structurally cannot: why a tour runs differently in shoulder season, what the parking arrangement really is in December, which conditions a clinic does not treat. Those sentences get lifted, and nobody else can write them for you.

The border supplies the other half. Search does not observe it, so a Seattle operator with a well-documented site turns up in replies to questions asked in Vancouver — and the mechanism runs both ways, putting a Vancouver operator in front of a visitor planning from Washington. The country and device splits in the Search Console views repay setting up properly.

Questions clients ask

None of this devalues the work already done. Well-structured, factually solid pages perform on both layers, which is why the habits set out here are the ones that made a service page worth reading long before any of this arrived. A rough measurement of the new layer still beats none, provided the roughness is repeated whenever the figure is shown.

To point the six views at your own domain and your linked Google data, start a Semalt account and take a baseline reading. Treat it as orientation, and give it two quarters before drawing conclusions.