Where the search result sat there is now a finished paragraph with three sources along its edge. We walk the Semalt panel's six AI views from the position of a Finnish manufacturer selling abroad, noting how unstandardised the measurement is.
A Swedish maintenance manager needs a replacement hydraulic cylinder for a harbour crane. He asks in a full sentence, gives pressure and temperature range, and gets six lines back. Beneath: two trade catalogues and a Southern Ostrobothnian maker whose page states the seal's temperature limit in numbers.
The manufacturer got no visit. It got a mention. Three weeks later a quotation request arrives saying the company was found through a comparison, and nobody can tie it to a tracked figure. Analytics shows no session, query or referrer.
This is the layer meant by visibility in AI search. It does not replace rankings but settles on top of them with its own logic. Semalt's AI analytics gathers into six views what is observable today.
The buyer who never opened your site
Ten blue links were a tray. The user got ten options and chose, and the measurement apparatus followed: which position, which click share, which page. When the answer comes written, the choice precedes the list.
In industry and export the shift shows in three ways unlike consumer retail.
- The decision happens earlier. Shortlisting happens inside a paragraph nobody saves. By the time the buyer opens a browser, two or three names are set and the rest ruled out.
- Three sources, not ten. A generated answer leans on a handful of picks. The gap between fourth and fifth place was small; between the third source and the fourth it is all or nothing.
- The question comes in the buyer's language. The German buyer who orders from Pirkanmaa asks in German. Your Finnish site is absent then, though the product fits exactly.
That third point matters more to a nationwide manufacturer than to a local service firm, which competes at home where the engine still steers traffic to a map and a list. An exporter competes in three language areas and exists in two only insofar as a model quotes it.
Rank five and the source list are separate contests
Intuition says a strong ranking brings a place among the sources. In practice the link is looser than anyone assumes. Rank follows how well the page answers the query; citation follows whether it holds one checkable statement for that spot.
A conveyor maker in Seinäjoki can sit fifth for years on belt materials in food processing. If a paragraph appears above the results, sourced to two trade magazines and a German rival, the rank holds but enquiries dry up. Nobody overtook it; something placed itself in front.
| Question | Rank report | Answer layer |
|---|---|---|
| What is measured? | Position for a keyword | Whether a statement qualifies |
| How many fit? | Ten rows, queue closed | A few sources, queue open |
| Source of the figure | Search Console and rank tracking | Samples and modelling |
| Update frequency | Daily, two days behind | In jerks, with model versions |
| Can cause be shown? | Usually, given data | Rarely, chance and cause blur |
| Target of the work | Page, technique, links | One paragraph and its numbers |
The fifth row causes agencies most trouble. Movements in the answer layer surface late, at odd intervals, with no traceable cause. Anyone waiting for proof never starts.
The keyword
Work was planned on a word list: pick terms, write a page per term, watch the rank. One winner per row.
- Page is the smallest unit
- Volume sets the order
The checkable statement
Work is planned on questions: what does the buyer want confirmed, and can one paragraph answer it. Several sources may serve one answer.
- Paragraph is the smallest unit
- Precision sets the order
The consequence is back to front. A small domain with few pages and many numbers often beats a large one whose pages read as sales copy. A tolerance limit can be quoted; a promise cannot.
The effect runs the other way too. A site built on keywords for ten years can rank strongly yet be absent from the answer layer, its texts shaped for an engine rather than a buyer. Plan, budget and report the two layers separately.
The AI competitiveness score and how to read it
The first view calculates an AI competitiveness score for a domain. It estimates how likely a model is to use your pages when answering in your sector. No instrument reading, only a summary of that moment.
The competitiveness score
How often your domain qualifies as a source in your own field.
- Valid only inside its own system. Readings from one tool speak to each other. A rival rests on another sample and other assumptions, so matching scales are coincidence.
- Direction before value. A single reading carries almost nothing. Meaning accumulates over six months or a year, and only if the competitor set held.
- Not for contracts. Keep it out of bonus schemes, agency agreements and board targets. It can move in a week with nothing published.
- One language at a time. Your Finnish and English material differ in size. One figure hides that you are ordinary at home and absent in export.
Market Circle: three rings around your own name
The other half of that view is the Market Circle diagram, where other domains sit at three distances from the centre. The edges teach more than the middle number: they show which set you were counted into.
For a Finnish manufacturer the result often surprises. The rival named in every sales meeting may be missing, replaced by a Swedish portal, a Polish contract manufacturer and a trade magazine.
| Ring | Who is there | What you learn | What not to do |
|---|---|---|---|
| Top tier | Large portals, trade media, directories | Topic map of the sector | Do not answer at that breadth |
| Middle tier | Peers nationwide and around the Baltic | Realistic yardstick, the real dividing line | Do not dismiss them as small |
| Narrow field | Slim domains where depth replaces breadth | Technique: one subject to the end | Do not read them as traffic |
The innermost ring teaches most. A domain with a dozen pages on one surface treatment returns as a source whenever a question touches it. No marketing budget, no newsletter, but numbers and stated limits.
How a model describes your company when nobody steers it
The third view asks the model to talk about your domain: what it infers you do, how large it thinks traffic is, where it sees growth. The answer comes as paragraphs with a few numbers in them, description rather than measurement, and useful as that.
Model-generated market context
An outside summary of what your site as a whole says about the company.
- Positioning. Set your company description beside it and mark where the two diverge. That difference list is the view's most concrete output, and its most uncomfortable.
- Traffic estimate. Read it as a range. The smaller the domain and narrower the language, the thinner the material behind it.
- Openings. Suggestions, not instructions. Weigh each against what you make, who buys it and at what margin.
The value lies in the differences. When a firm sees itself as a specialist in dairy process pipework, but most of its pages discuss quality and responsibility in general, the description returned is an ordinary metal-sector operator. That reads the site precisely.
Question research: intent decides, not search volume
In the fourth view the model sifts the question field and attaches to each find what the asker wants. Sorting by volume and working down loses its point. In Finland that has been weak for years; in a narrow field it misleads.
A question written as a full sentence typically draws single-digit searches a month. When Finnish case endings and compounds split one subject across ten surface forms, a word list's total lands below real demand. Intent fastens to the goal, not the inflection. The panel's question view groups finds by intent, so the material arrives sorted.
| Intent | Example | Weight in the answer layer | What the text needs |
|---|---|---|---|
| Wants to understand | hot-dip galvanising versus paint on the coast | Most generated paragraphs rest here | Thorough explanation under its own heading |
| Wants to confirm | service life of a C4 coating | One figure is lifted out | Number, unit and validity limit |
| Wants to compare | stainless or acid-proof food line | Paragraph and visit arise together | Limit values and an honest exclusion |
| Wants to act | turning services Ostrobothnia | Still held by the results list | Location details and a contact route |
| Looking for a name | a product code or brand | Unchanged | No growth here |
The top two rows decide answer-layer visibility almost entirely. The fourth stays the home ground of rankings and needs different text, so it belongs in a different plan. Mixed together they serve neither purpose.
In English the same sum lands in another order of magnitude: many times the questions, many times the sites facing you. The domestic version gives an exporter few contacts but ones far into the decision; the English version gives quantity and a tighter field. One figure hides both.
Content gaps, and what a citable paragraph contains
The two remaining views turn into a work list and are best read in sequence. The first, the competitor strength and content gap analysis, sets rivals' strong areas against your quiet spots. The second surfaces published pages worth expanding or using as linking hubs. The shortfall is in the answers, not the word list.
| What is missing | How it shows | First fix | Effort |
|---|---|---|---|
| Dimensions | Rivals give dimensions and weights; you describe in words | Move the dimension table into the body text | small |
| Tolerances | The page says "precision machining" with no figure | Write out the tolerance class and method | small |
| Material data | Trade media answer material questions for you | List alloys, standards and operating limits | medium |
| Delivery terms | Lead times get asked about; the site is silent | State the real ranges, stage by stage | medium |
| Export language | Covered in Finnish, open in English or German | Write it as its own text, not a translation | large |
The bottom row costs most and is deferred year after year. A machine-translated page repeats the content, but the trade's settled terms become near-synonyms on the way, and the term is what recognition rests on. That is dearer than usual for an exporter, since the same firm sells at home in Finnish and switches language abroad.
Figure, unit and starting point
A liftable sentence carries a quantity, its unit, and when counting starts. "Fast delivery" carries nowhere; "twelve working days from approved drawing" does.
- No adjectives for figures
- A range where variation exists
The exception in the same paragraph
If the qualifier sits at the foot of the page, the rule gets quoted without it. Put the condition beside the rule itself.
- Rule and limit side by side
- One paragraph, the whole answer
The standard by name and year
The word "certified" does not qualify as a source. The designation, its version and its validity do, and they weigh heavily on customs.
- The designation in full
- The scope written out
Text, not an attachment
Anything inside a brochure, an image or a customer portal is invisible here. So is a page that loads slowly or goes unindexed.
- Numbers in the body text
- Reachable without a login
The list grows faster than anyone can write. Set a ceiling first: ten questions per language, the rest to the next period. Ten finished answers carry further than forty half-done. The technical groundwork is a job of its own, described on our technical SEO services page.
Global visibility value: one line for the board, one risk in the report
The last view compresses everything into a single global visibility value. It covers every domain you control, built to be followed over years. For a management team it is a handy one-line summary; it turns dangerous when that line enters a bonus formula.
Many Finnish groups run three or four domains: the parent site, a brand site, an English export site and perhaps an old web shop nobody removed. The shared value reports their sum, which moves even when one part strengthens as another decays.
- As a comparison figure. Direction is read from the slope. A value with no history says nothing; the first reading is only a baseline.
- As an alarm. A sudden collapse means dive into the detailed views. It gives no licence to draft the conclusion first.
- As a shared language. A common metric shortens meetings, provided its limits are restated aloud.
The rhythm need not be brisk. Open the rings and market description quarterly, take a dozen topics per language from the question view, and renew pages a few at a time. Read results alongside Search Console and rank tracking figures, so every assumption faces a measured number.
Frequently asked questions
How do we even know whether we are mentioned in generated answers?
No direct counter exists, and that is best accepted at the outset. The practical substitute combines three indicators: the panel's estimates, the origin question in sales conversations, and whether short informational queries fade from Search Console while enquiry quality improves. None is proof; together they suffice.
Our sector sees a few dozen searches a month in Finland. Is this worth it?
Often for exactly that reason. When volumes are small, every asker is closer to a purchase and competition for exact technical answers is non-existent. Ten precise questions with answers written to them are the cheapest route in. Measured by volume the work looks pointless; measured by deals it rarely is.
Our product data sits behind a reseller portal. Is that enough?
It is not. Information behind a login does not exist for this layer, nor do PDF catalogues or tables inside images. The fix is not to open the portal but to move onto the public product page what your competitor gives away free: dimensions, materials, operating limits and standards. Prices and customer terms stay behind the login.
Can the competitiveness score be compared with another tool's figure?
It cannot. The figures come from different samples, question sets and assumptions, so one number means different things in different sources. Compare only consecutive readings from one tool, and record where each figure came from and when. The field is still unstandardised.
How often should this be checked if one person handles marketing?
Four times a year is plenty. The rings and market context shift slowly, and closer monitoring mainly produces noise that people then explain. Writing should stay steady instead: a few pages a month beats a twice-yearly push left unfinished both times.
Visibility in AI search is a layer on top of conventional rankings, not a replacement. Its metrics are young, poorly standardised and prone to move for reasons outside your control, but the work is old and familiar: write out what the buyer would otherwise ask on the phone. Dimensions, tolerances, materials and lead times are both your best sales argument and the only material a machine can quote. We publish walkthroughs like this on the blog.
The practical start is narrow: open the AI views for your own domain in the Semalt panel and look first at who lands on your rings and how the model describes your company. If that description differs from what you sell, you have next quarter's order of work.