A Helsinki maker of cross-direction moisture profilers can be the obvious answer to a buyer's question and never learn the question was asked. This article is about the layer above the ranked list, why it weighs more at twenty searches a month than at twenty thousand, and how little of it can honestly be measured.
For two decades search visibility had one shape: ten blue links, a position number, a click-through rate attached to each rank. The whole reporting stack was built to describe that shape, and it described it well enough.
A second shape now sits on top of it. Ask a procurement engineer's question and the page often opens with a written answer: some sentences of synthesis, sometimes a short list of named vendors, drawn from sources the reader never opens. Assistants outside the search box do the same without any list underneath.
The layer above position one
Treat this as a distinct layer, not as a new ranking factor. Classic ranking decides which documents are eligible and in what order. The layer above reads a handful of them and writes something that did not exist before the query was typed.
Two consequences follow for a supplier. Being eligible is no longer being seen: a page can rank fourth and be paraphrased into the answer, or rank first and be passed over for a directory listing. And the answer often ends the session — no click, no row in your analytics, nothing recording that the exchange happened.
For a consumer brand that is erosion: a percentage of clicks lost off the top of a large number. For a specialised industrial exporter it is closer to substitution, because the number underneath was never large.
Twenty searches a month changes what is at stake
Finnish is the working language of about five and a half million people. Take a phrase a paper-machinery buyer would really use, render it in Finnish, and the monthly demand is not modest — it is arithmetically negligible. Twenty searches a month is a generous reading, and many such phrases register nothing at all, month after month.
The English export terms are thicker, but not by as much as people assume. A phrase naming a class of web-tension controller or an ice-class notation may draw a few hundred searches worldwide in a year, across a dozen countries. That is a real market with six-figure contracts behind it, and still too small a sample for any conventional statistic.
Now put a generated answer on top. If forty people worldwide ask a purchasing question in your category this quarter, and the answer names three suppliers, that naming decision has drafted the shortlist for the quarter. No long tail of other impressions compensates. The distribution is not skewed; there is barely a distribution.
What you can see, and what nobody can
Start with the part that is not negotiable, because most confusion here comes from skipping it. There is no official counter for how often a model names your company. No search engine and no assistant vendor publishes a per-domain citation figure, exposes one through an API, or sells one in a paid tier. The metric does not exist as a first-party number.
What exists is a set of narrow, indirect signals. Referral traffic from assistant hostnames shows up in ordinary analytics when a reader does click through, honestly measuring a small fraction of the behaviour. Neither that nor a rise in brand searches is a citation count.
| Signal | What it genuinely shows | What it cannot show |
|---|---|---|
| Referrals from assistant hostnames | Sessions where a reader followed a link out of an answer | Answers that named you and ended there |
| Rise in unprompted brand searches | Awareness moved for some reason | Which surface caused it |
| Manual prompting of a model | One answer, at one moment, from one location | Whether the next reader gets the same answer |
| Vendor visibility scores | An estimate built from sampled prompts | A count of real citations to real users |
| Classic rank for the same query | Eligibility of your document | Whether the synthesis used it |
Manual prompting deserves its own caution, because it is the first thing everyone does. Typing the category question into an assistant and reading the reply feels like verification. It is a spot check of one sample from a deliberately non-deterministic system, and tomorrow's reply may name different vendors.
- Answers vary by country. The vendors named for a German buyer and for a Singaporean one can differ, and neither list is the one you saw from Helsinki.
- Nothing is logged on your side. A reply that ends the session leaves no impression, no click and no row anywhere in your own reporting.
- Rank and citation are separate questions. Appearing in the ten links neither guarantees nor prevents being used in the synthesis printed above them.
- Brand search is the only echo. Readers who never clicked sometimes search the company name later, which is suggestive and never conclusive.
The six views that try to fill the gap
Given that no first-party figure exists, tooling here works by inference: probe the models systematically, classify what comes back, turn the pattern into a score. That is a legitimate method with an unavoidable ceiling. The six views split cleanly into two halves, and the halves deserve very different amounts of trust.
Scores, tiers and written context
The model's opinion of where a domain stands in its category, expressed as numbers and prose.
- A competitiveness score with a market circle. Rival domains are sorted into top-tier, mid-tier and niche bands, which on a specialised export term is a shorter and stranger list than the sales team expects.
- Written market context for the domain. Positioning, an estimate of traffic and a set of stated opportunities, produced by the model reading the category rather than reading your logs.
- A global visibility value across the AI search landscape. One aggregate figure per domain, and the number most likely to be misread, which the later sections address directly.
The second half produces lists rather than scores, and a list can be checked by reading it. That difference is why it carries most of the practical value here.
Questions, gaps and pages worth expanding
Output you can verify by hand, which makes it usable even when the scoring is uncertain.
- Query research with intent classification. Candidate questions are generated and sorted by what the asker appears to want, which matters when your own queries are far too few to cluster.
- Pages flagged as levers. Specific URLs marked as worth expanding or worth linking to internally — a shorter and more actionable list than a site-wide audit.
- Competitor strengths and content gaps. Where a rival is covered thoroughly and you are silent, expressed as topics rather than as keywords.
Semalt sets out the reasoning in its overview of generative market research. On an export site the working time goes into the second half, because those views can say something about demand your own property has never recorded.
Reading a market circle three companies wide
The competitor bands were built for categories with hundreds of participants. Run them against a term like ice-class thruster overhaul and the field may be nine firms worldwide, four of them Nordic and two subsidiaries of one group. The tiers sort correctly; they sort a very short list.
That is not a defect, but it changes the reading. In a crowded category the useful question is which tier you occupy. In a nine-firm category it is whether the list is right at all — whether the model's competitive set matches the companies you meet in tenders.
- Check the roster before the ranking. If the circle names three firms you have never bid against and omits the two you always do, the category boundary is wrong and every score derived from it inherits that error.
- Expect adjacent categories to bleed in. Measurement and testing equipment sits next to laboratory supply and next to industrial automation; models frequently merge the three, which inflates the apparent field.
- Watch for directories in the vendor slots. Where a category is thin, answers often name a trade portal or a buyer's guide instead of a manufacturer — a different problem, and a solvable one.
- Read the gap list as a writing brief. A topic where a rival is thorough and you are silent is actionable this month, regardless of whether any score moves.
Fix the description before the score
If the model has the wrong idea of what you make, no amount of link work moves anything. The repair is editorial: say plainly what the product is and which standards it meets.
Work the gap list in order
When the competitive set matches reality, the topic gaps are trustworthy enough to plan a quarter of writing around, one page at a time.
What tends to get a company named
Nobody outside the model vendors can state the mechanism, and any checklist promising citation is overselling. What can be said is which properties make a page usable as a source — ordinary enough that they were worth doing anyway.
Be resolvable as an entity
A model has to be confident that your company name refers to a specific manufacturer of a specific thing, in a specific place, before it will risk naming you.
- State what you make in plain nouns, early
- Keep the company name spelled one way everywhere
Put specifications in text
Ranges, tolerances, certifications and classification-society notations locked inside a PDF datasheet are far less usable than the same figures written into the page.
- Mirror the key numbers in HTML
- Name the standards explicitly, not by allusion
Be described elsewhere too
A claim repeated across independent sources is safer to reproduce than one that appears only on your own site. Trade bodies, distributor pages and conference programmes carry weight here.
- Claim the industry listings you already qualify for
- Placement networks reach further, at a cost
Answer the question as asked
Buyers phrase these things as questions, and Finnish case endings and compounds fracture the same idea across many surface forms. Pages built around questions survive that fracturing better than pages built around a keyword.
- Use the buyer's phrasing as the heading
- Answer in the first two sentences beneath it
The fourth card transfers least well from generic advice, so it is worth restating. When a keyword tool reports zero volume for eleven variants of one Finnish compound, it is not describing absent demand — it is describing a language that spreads one intention across many strings. The question form is the stable version of that intention, and it is also what a generative system is answering.
An inferred score is not a measurement
The visibility figure is genuinely useful and genuinely easy to misuse. Three limits need saying without hedging.
The third limit is specific to a site like yours. An inferred score is built by sampling: generate category prompts, run them, count mentions, aggregate. How much that procedure is worth depends on how well the sampled prompts resemble the real ones. On a high-volume consumer term there is abundant query data to ground the probe set in.
At twenty searches a month that grounding is largely absent. The probe set is mostly synthetic — plausible questions rather than observed ones — because there are not enough observed questions to build from. The inference is weaker here than the same method would be in a large market, and it is worth saying so rather than letting a tidy number imply otherwise.
| Figure | Honest description | Safe use | Unsafe use |
|---|---|---|---|
| AI visibility score | Inferred from sampled prompts | Tracking your own direction over quarters | Any external document or contractual claim |
| Market circle tiers | Model's view of the category | Checking whether your competitive set is understood | Presenting as market share |
| Content gap list | Topic comparison against rivals | Choosing what to write next | Sizing an opportunity in revenue |
| Manual prompt check | One sample, one moment | Sanity-checking an obvious error | Reporting as a result |
| Assistant referrals | Real, counted, and partial | Confirming the channel exists at all | Estimating total generative exposure |
Inside those boundaries the views earn their place. A score moving one way across three quarters, alongside a gap list you have been working through, reasonably indicates the work is landing. That is a modest claim, and modest claims are the only kind this evidence supports. The competitor and content-gap analysis holds up best under scrutiny, because its output is a list you can verify by reading.
Questions that come up
Can I find out how often an assistant recommends my company?
No, not as a counted figure. There is no first-party citation report from any model vendor and none for sale. You can observe referral clicks from assistant hostnames, which is real but partial, and you can use a score inferred from probing. Anyone offering a true count is describing an estimate.
Should this change what we publish, or is it a separate workstream?
It should not become a separate workstream. Everything that makes a page usable as a source — clear entity description, specifications in text rather than only in PDFs, question-shaped headings, corroboration from independent sites — improves ordinary ranking too. Treat it as a further reason to do the same work, not as a parallel budget.
Our Finnish terms show no volume at all. Does this layer even apply to them?
It applies more, not less. A term with no measurable volume is still asked by a handful of real buyers each year, and an answer naming three suppliers is a large share of what they see. The channel is small in absolute terms and disproportionately decisive per query.
How often is it worth reviewing these views?
Quarterly. The models change on their own schedule, and the score is noisy enough that monthly reading invites reaction to sampling variation. Campaign work of any kind typically needs four to eight weeks before first measurable movement, and on a low-volume property that window wants lengthening, not shortening.
Does a paid campaign tier change our position in generated answers?
Not directly, and no honest description claims otherwise. The tiers cover keyword work, link placement and on-site suggestions — 149 USD per month per domain for the automated level, 500 USD for the tier adding manual selection and human review; the breakdown of what each campaign level includes sets out the difference. Better corroboration plausibly makes a company easier to name, but that chain is indirect and unverifiable.
A real layer, an honest instrument panel
Two things are true at once, and holding both is the discipline. The generative layer matters more for a Finnish exporter than for a large consumer brand, because the volumes underneath are so small that one answer can carry a quarter of discovery for a term. And the instrumentation is inferential, sampled, and weakest exactly where your market is thinnest.
The resolution is unglamorous. Do the work that makes a company nameable, since all of it pays off in classic ranking regardless. Read the gap list and the query research, which are checkable by hand. Track the score for direction over quarters and never quote it outside the building.
The rest of our thinking on measuring small markets sits with the other English articles here, and the way we run it alongside ordinary technical work is set out on the service pages. Neither replaces reading your own category with your own eyes, which at this scale is still affordable.
To see what the generative views say about your domain before deciding how much weight they deserve, connect the property through Google sign-in and let the module build its first pass: open the Semalt dashboard and add your site. On an export property the first useful output is rarely the headline figure. It is the moment the circle names a competitor you were not watching, or omits one you thought defined the category.
