A buyer preparing for a Cologne fair asks an assistant which suppliers to visit. It names three. You are either in that answer or you are not, and there is no page two to be on.
This layer sits above ordinary search rather than replacing it. The same site can hold a respectable position for a term and be entirely absent from a generated answer about the same subject — and the reverse happens too. Both are forms of visibility, both matter, and neither substitutes for the other.
For a foreign supplier the stakes are higher than for a domestic one, and the reason is structural. A German buyer already knows the established domestic names; they do not need to ask. The question gets asked precisely when somebody is looking beyond the familiar set — which is exactly the moment an exporter wants to exist.
Ten results become three names
A classic results page shows ten links and leaves the choice to a person. A generated answer makes the choice and names two or three. That is not a difference of degree. It converts a list in which seventh place still exists into a statement in which it does not.
What ends up in such an answer is rarely what is best optimised. It is what is most unambiguously described. A model making a selection needs attributable statements: which capability, for whom, in which territory, with what constraint. A page that offers everything and specifies nothing is unusable for that purpose no matter how technically sound it is.
What can be observed instead
With no counter available, what remains is systematic observation: ask questions, collect answers, look for patterns. The generative market analysis in the Semalt workspace does exactly that across six views, describing the market as a model describes it rather than as a results page ranks it.
The six surfaces and what each one answers
One domain, six angles. The interesting part is usually where they disagree with the company's own description of itself.
- Competitiveness with the Market Circle. A placement into top, mid or niche tier. It does not say how good a domain is; it says how far it sits from the names a model reaches for first.
- Model-generated market context. Positioning, traffic estimate and identified opportunities, written from the model's point of view. Comparing it against the company's own boilerplate is frequently the most informative ten minutes of the exercise.
- Query research and intent classification. Not keywords but questions, sorted by whether somebody is informing themselves, comparing, or ready to commission.
- Pages flagged as leverage. Existing pages identified as candidates for further content or better internal linking — usually pages buried three levels down in the menu.
- Competitor strengths and content gaps. Subjects covered comprehensively elsewhere with nothing equivalent on your own site.
- Consolidated AI search visibility. A single figure across the AI-assisted layer — and the point at which the three limits in this article apply.
These views replace none of the conventional reporting. Search Console describes what happened on your property. Position tracking describes where you stand relative to others. The generative analysis describes how a model characterises the market you operate in — a statement about perception rather than about traffic.
What the number is and is not
A consolidated visibility figure is useful as long as its origin stays visible. It comes from observation: questions asked, answers collected, frequency of appearance counted. That is a sample, not a census.
Those three sentences sound deflating and are the reason the exercise is worth doing at all. A figure whose limits are stated can be used sensibly. A figure presented without limits produces decisions nobody can defend the moment somebody asks how it was calculated.
What makes a page usable as an answer
Observing which pages appear in answers produces a short list of characteristics. They are unglamorous and have more to do with disciplined writing than with technology.
- One question, one answer. Sections that answer a specific question completely are taken up far more often than sections that circle a subject.
- Numbers with a unit and a condition. "Typically four to eight weeks" is usable. "Fast" is not. Every figure needs a boundary attached to it.
- Stated scope. Which capability, for which type of customer, in which territory. A supplier that does not delimit itself is indistinguishable in a selection of three.
- Named constraints. Pages that also say what something is not suitable for read as more reliable — to people and to models alike.
- Stable addresses. A page that has sat at the same address for two years and been maintained appears markedly more often than one that moves every redesign.
The overlap with plain good writing is almost total, which is the useful part. A page written so that a hurried reader finds the answer in thirty seconds has already taken the shape a model can reuse. That is one of very few places where two objectives align without compromise; how the underlying term and topic work is organised is covered in our content strategy.
The same claim, written two ways
Nothing here is a keyword change. Every line replaces an adjective with a fact.
- "Fast delivery" becomes "shipped within five working days from Rotterdam". A person reads both as a promise. Only the second can be compared against a competitor, and comparability is what a selection runs on.
- "Experienced team" becomes "installing at Koelnmesse since 2014". The claim gains a date and a place, which makes it attributable to a question about this city rather than to a question about the category.
- "Flexible solutions" becomes "minimum order two units, no maximum". Stating the lower boundary loses the enquiries you would have declined anyway and keeps the ones you want.
- "Full service" becomes a list of four named services and one you do not provide. The excluded item does more work than the four included ones, because it demonstrates that the list is real.
Reworking a page this way takes an hour and changes something that no amount of technical work reaches. It also has an effect that rarely gets mentioned in this context: the quality of incoming enquiries rises, because the people who would not have been a fit stop writing. For an exporter with limited capacity for German-language enquiry handling, that is often the more valuable of the two results.
The three facts an exporter must state
For a company selling into Germany from elsewhere, three specific pieces of information decide whether a page can be used in an answer at all. None of them are marketing statements and all three are routinely missing.
| Fact | Why a model needs it | Why a buyer needs it |
|---|---|---|
| Delivery territory, named | Makes the supplier attributable to a geographic question | Answers the first objection before it is raised |
| Invoicing and currency | Distinguishes a domestic-equivalent supplier from a foreign one | Determines whether procurement can proceed at all |
| Language of support | Separates "sells to Germany" from "operates in German" | Decides whether an enquiry is worth making |
| Lead time, with a number | Provides the comparable value a selection needs | The single most asked question at a stand |
Adding these four lines to the pages that matter is an afternoon of work. It changes how the pages perform in ordinary search as well, because the same information removes the reason a cautious buyer had for not clicking. This is the rare case where the same edit improves two unrelated things.
"Delivery to Germany, Austria and the Netherlands"
Three named countries. Attributable to a question about any of them.
- Usable in a selection
- Removes a buyer objection
"International delivery available"
True, unfalsifiable and useless for a comparison between two suppliers.
- Matches no geographic question
- Leaves the objection in place
How the two layers relate
The obvious question is whether to prioritise one layer over the other. The unhelpful answer is no, because they serve different phases of the same decision. The generated answer comes first, when somebody is sorting a market they do not know. The classic listing comes last, when somebody is checking a name they already have.
Orientation
The buyer does not know the market and wants three names. Attributability decides the outcome.
- Clear scope beats breadth
- Stated numbers beat adjectives
Verification
The name is known; references, terms and reachability are being checked.
- Classic position matters here
- So does a page that answers questions
In practice the same page has to do both jobs. It must be precise enough for a model to reuse and complete enough that a verifying human does not leave. Search Console and position tracking remain the dependable basis; the AI assessment adds a direction rather than a measurement. Both sit next to each other in the Semalt dashboard, which keeps the two layers visible side by side rather than in separate tools.
- Two readers, one page. One skims for thirty seconds, the other reads for ten minutes. Subheadings and short answer paragraphs serve both; unbroken blocks of text serve neither.
- Figures next to the claim they qualify. A lead time belongs in the sentence describing the service, not in an annex. Anything separated is skipped separately.
- No page that depends on another page. A page whose meaning only emerges alongside the home page cannot be used on its own — not by an assistant and not by a visitor who arrived directly from search.
- A visible update date. It works on both audiences. For event-related pages it is the difference between "still current" and "left over from two editions ago".
None of this is new, and that is exactly the point. The newer layer rewards broadly the same qualities good websites have always had; it simply penalises their absence more visibly. Companies that published generic pages for years because generic was sufficient for classic search notice the shortfall much faster now.
A defensible way of working
Because the evidence here is weaker than for conventional metrics, this subject needs a slower rhythm. A quarterly cycle has proven about right — more often produces noise, less often misses movement.
| Step | Activity | Output |
|---|---|---|
| 1 | Write down twenty real customer questions, verbatim | a list without industry jargon |
| 2 | Check which of them have a page of their own | a gap list |
| 3 | Read the market context and competitor strengths | subjects that belong to somebody else |
| 4 | Write or rework two to three pages | attributable answers |
| 5 | Compare tier and visibility figure after a quarter | a direction, not a result |
Step one is the most important and the most frequently skipped. Real customer questions sound nothing like search terms: they are longer, clumsier and full of conditions. "Can you deliver to the hall on the Saturday if the stand is only released on Friday evening?" is not a query — but it is the question that actually gets asked, and a page that answers it is useful in both layers.
Step three has a caveat worth stating. The model-generated market context sometimes describes a domain differently from how its owners would describe it. The instinctive reaction is to call the description wrong. The more useful reaction is to ask what it could plausibly have been derived from — and the answer is almost always the text actually published on the site. Read that way, the discrepancy is not a model error but a precise piece of feedback about your own copy, and acting on it saves several rounds of guesswork. That analysis sits alongside the conventional metrics in My SEO, so a text change and any subsequent metric movement appear on the same timeline.
What this actually produces
Anyone doing this work seriously should know what they end up with. Not a metric to present to a board. A website that describes what it does more precisely, and therefore performs better in both layers.
That is less dramatic than the phrase "AI visibility" suggests and considerably more durable. Attempting instead to raise a derived figure directly means optimising against a sample, which fails as soon as the sample changes. The dependable foundation remains clean structure, precise text and regular review; the technical half of that is covered under technical SEO, and the practical checks for appearing as a supplier who genuinely serves this market under local SEO. The campaign side, where text changes and metric movement can be followed on one timeline, sits in My SEO.
Open the AI market analysis in the dashboard
Frequently asked questions
Can we measure how often our site is named in AI answers?
Not in the sense of a measurement. There is no official counter for a provider to read. What is possible is asking questions systematically, collecting answers and deriving an estimate. That estimate indicates a direction and is useful as such — but it is a sample and should not be treated like a click count.
Should we write content specifically for models?
No, and that is the good news. The characteristics that make a page citable — one question per section, numbers with conditions, stated scope, named limits — are the same ones that satisfy a hurried reader. Writing for models while forgetting the reader loses on both counts.
Will we lose traffic if assistants answer directly?
For purely informational questions, partly yes. Somebody who only wants to know how long something takes clicks through less often. For questions with commissioning intent it is different: a buyer looking for a supplier verifies that supplier. The shift mainly affects pages whose only purpose was a short factual answer — and their value was limited to begin with.
We are a small exporter. Is this worth the effort?
Often more so than for a large domestic competitor. In a crowded field of familiar names the distance is considerable. For a question about a specific capability combined with a specific delivery territory, however, there are very few unambiguously attributable suppliers — and naming your capabilities, territories and constraints clearly makes you one of the few candidates quite quickly.
How does this interact with the fair schedule?
The quarterly rhythm and the fair calendar fit together better than they appear to. Run the question-collection step immediately after an event, while the questions asked at the stand are still fresh and written down. Do the writing in the following quarter. By the time the next edition comes round, the pages have had two to three quarters of history, which is roughly what they need to be reached for at all.
