A technical buyer in Stuttgart asks a model which European companies make the thing you make, and the answer arrives as a paragraph naming four firms. There is no position eleven inside a paragraph. You are in the sentence or you are not, and nothing on your dashboard records which.
For a Copenhagen company selling abroad, this is the part of search that has changed fastest and is discussed least honestly. It sits above the ranking work you already do, it behaves differently in Danish than in English, and almost every number quoted about it is softer than the person quoting it admits.
The paragraph that names four companies
An audiology equipment maker in Herlev builds test booths and calibration units. Its customers are hospital procurement departments, university clinics and distributors in Germany, the Netherlands, the United Kingdom and across the Nordics; Danish orders are a small fraction of turnover. For two decades the path to a new distributor ran through a trade fair, a catalogue and a phone call.
One step in that path has quietly moved. A buyer at a German clinic group, told to bring back two alternatives to the incumbent, increasingly begins by asking a model to name European manufacturers of audiometric equipment. What comes back is prose: four or five companies, a clause of description each, occasionally a source cited alongside. That paragraph functions as a shortlist, and it was assembled before anybody opened a website or requested a quotation.
If your name is in it, you have entered a procurement process you did not know had started. If it is not, nothing happens — and nothing is exactly what you observe. No impression to count, no click that failed to arrive, no downward line to investigate on a Monday morning.
The Danish answer and the English answer are different answers
Ask a sourcing question in Danish and the model draws on a shallow pool: a few domestic trade publications, an association's member list, a handful of company sites, perhaps a business-press profile. The field of candidates is small because the country is small, so a firm well known at home has a fair chance of being named for lack of alternatives.
Ask the same question in English and the pool is continental, then global. Trade publishers, distributor catalogues, procurement platforms, conference proceedings, university tender pages and every competitor from Bavaria to Boston go into the material. The Danish answer flatters you; the English answer is the one your buyer receives.
This is the asymmetry that runs through everything a Copenhagen exporter does online, in its sharpest form. The Danish pages contend with Danish firms; the English pages contend with everyone. In rank tracking that appears as a gap between two average positions, uncomfortable but legible. In generated answers it appears as something blunter: named in one language, unmentioned in the other, with no figure on the screen to mark the difference.
| Asked in | Material the answer draws on | Realistic field of candidates | What being named is worth |
|---|---|---|---|
| Danish | Domestic trade press, associations, a few company sites | Narrow — sometimes fewer firms than the answer has slots | Reinforces a reputation the reader mostly had already |
| English | European and global publishers, catalogues, platforms, rivals | Wide — dozens of plausible names for four positions | Creates a relationship that did not previously exist |
| Danish, absent | Usually a coverage gap rather than a judgement | Fixable by publishing anything substantive | Minor: the buyer probably knows your name anyway |
| English, absent | You were outweighed by better-corroborated candidates | Requires third-party evidence, not more pages | Severe: the enquiry goes to someone else silently |
A model repeats what it can corroborate, not what you claim
The instinct on discovering this layer is to write a page addressed to the machine — dense with adjectives, thick with the phrase it hopes to be matched against. That instinct is wrong in a useful way. Generated answers are built from material that can be cross-checked, and marketing prose is the least checkable text on the internet.
- Facts stated plainly beat claims stated loudly. A sentence naming a standard, a measurement range, a certification or a delivery region can be matched against other sources. "Leading Scandinavian supplier" can be matched against nothing.
- Other people's pages carry unusual weight. A distributor's listing, an association directory or a university's tender record describes you from outside, and outside descriptions are what a model treats as corroboration.
- Named things survive summarisation. Model numbers, protocol names, material grades and regulatory references pass through a paraphrase intact. Slogans do not.
- Contradictions cost you the slot. If your English site, your German distributor's page and your catalogue give three different figures for the same product, the safest thing an answer can do is name a company whose numbers agree with themselves.
Specification pages written for an engineer
Ranges, tolerances, compliance references and interface details set out in ordinary sentences rather than in an image of a data sheet.
- Every number is a fact another source can confirm
- Survives being condensed into one clause
Installation and reference accounts
Where a system went in, what it replaced, which constraints applied and what the measured outcome was.
- Ties your name to a place, a sector and a problem
- Gives a third party something to quote later
The English "About us" written by committee
Three paragraphs of values, heritage and ambition, containing no statement that could be checked against another document.
- Nothing in it distinguishes you from a competitor
- Nothing in it can be corroborated or refuted
Pages that exist only in Danish
The technical documentation, the certification notes and the case material all sit on the Danish side and were never carried across.
- The evidence is real but unavailable in the answer's language
- The English tree looks thinner than the company actually is
That last card is the common Danish case, and it is an oversight rather than a decision. The proof of competence accumulated in the home language, because that is where it was produced. The English section came later, as a summary — and a summary is exactly the kind of text that gets read, understood and left out.
What can be observed, and where you would look
None of the above is actionable until you can compare one quarter with the next. That is what the generative research module in the unified Semalt workspace is for: not a view inside anyone's model, but an outside view of how a domain sits in a market as a model would describe it.
Six views over how a market gets described
For a company whose English pages face a field of competitors it has never had to name before.
- A competitiveness score with a market circle. Rivals are separated into top-tier, mid-tier and niche positions, which is a more honest map of an export category than a ranked list of domains.
- Model-generated market context for your domain. How the company is positioned, an estimate of its traffic and where the openings are — a description, not a verdict.
- Query research with intent classification. Candidate phrases sorted by what the searcher is trying to accomplish — which matters more than volume when the useful queries are asked forty times a month across Europe.
- Pages flagged as levers. The module marks pages worth expanding or worth linking to internally, usually surfacing reference material nobody thought of as commercially important.
- Competitor strengths and content gaps. Where a rival is better corroborated than you, and which subjects in your category nobody has documented properly.
- A global visibility figure for the AI search landscape. One number for the whole domain, useful as a trend line and dangerous as a claim — see the next section.
The market circle is the view most worth an hour of a director's time, because it forces a conversation the company has been avoiding. Run it against the Danish side and the top tier holds firms whose managing directors you have met. Run the competitive landscape analysis against the English side and the top tier holds companies with three times your headcount and a decade more trade coverage. Both pictures are true; only one describes the market your growth plan depends on.
Three things to say out loud before anyone quotes a number
This is where the gap between what is sold and what is known is widest. Three statements, none comfortable, all true regardless of whose panel you are looking at.
None of that makes the layer unimportant. The buyer at the German clinic group is genuinely reading a generated paragraph, and the shortlist it produces is genuinely a shortlist. It means only that the honest formulation is narrower than the marketing one.
What you can say internally
That your position in a described market has moved in one direction over three consecutive readings taken the same way.
- Comparative, method-consistent, dated
- Leads to editorial decisions you can act on
What you cannot say to anyone else
That you appear in a stated percentage of answers, or that you outrank a named competitor in a generative result.
- No source exists that could confirm either
- One follow-up question ends the meeting
The work that actually changes the outcome
Once the measurement question is settled, what remains is unglamorous and mostly editorial. It also pays twice: material solid enough to be summarised accurately by a model is usually material a human buyer wanted anyway.
- Move the evidence into English. Test reports, standards work, installation accounts and technical notes that exist only on the Danish side are invisible to the answer your export buyer receives. This is the single highest-yield task on most Danish domains.
- Publish the specifics competitors keep in PDFs. A data sheet locked in a downloadable file is a fact nobody can quote. The same figures in page text can be repeated.
- Make third parties describe you correctly. Distributor listings, association directories and trade coverage are the corroboration layer. A distributor page that describes your product wrongly is worse than no page.
- Say the same thing everywhere. One set of numbers, one product naming convention, one description of what the company does — across your site, your partners' sites and your documentation.
- Write for the buyer described in section one. Not for the domestic reader whose page you translated. Pages built from the wrong questions read as generic.
The English tree needs its own editorial plan rather than a translation queue; our approach is set out under content strategy.
A floor above ranking, not a door away from it
It would be convenient if generated answers replaced the result page, because then you could stop maintaining one of them. They do not. The material a model draws on is largely the material that ranks, so classic search work is the foundation this layer stands on, not a legacy activity you are leaving behind.
| What you observe | Most likely cause | Where the fix lives | How long before you can judge it |
|---|---|---|---|
| Named in Danish answers, absent in English ones | The evidence exists only in the home language | Editorial: translate and expand the technical material | One to two quarters |
| Absent in both languages | Nothing about you is documented anywhere checkable | Publishing plus third-party listings | Two quarters and upwards |
| Named, but described wrongly | Contradictory figures across sources | Reconciling your site with partner pages | Weeks, once the contradictions are found |
| English pages rank but are never cited | Pages persuade rather than document | Rewriting towards specifics and evidence | One quarter |
| Nothing changes anywhere for months | The pages may not be indexed | Discovery and indexing, before anything else | Immediately observable |
That last row is the one people skip. A page never fetched cannot rank and cannot be cited; the order in which those problems have to be solved is set out across the English articles here, and it is worth confirming before drawing conclusions from any visibility figure.
Asking questions against your own project data
For the fortnightly half hour when somebody has to say what changed and why.
- One chronological feed per project. Answers, automated reports, newly placed backlinks with the donor's domain rating and traffic, to-dos and campaign news arrive in a single stream rather than in five separate screens.
- Answers bound to real data. A router model decides per question which blocks to load — Search Console, rank tracking, campaign, custom — pulling between none and three by relevance, so replies are grounded in your project.
- Streamed and remembered. Replies arrive token by token and a conversation of up to twenty messages is retained, which is enough to follow one line of enquiry to its end.
- Filtered and searchable. All, links, files or to-do, with to-dos held as active, deferred or dismissed, and full-text search across the history. Keyword and URL lists can be handed over in batches.
A caution belongs with that chapter. An assistant reading your project data is a convenience, not an oracle. It can tell you which keywords entered the top ten last month, because that is recorded. It cannot tell you how often a model named your company, because nobody records that — and asking anyway will produce a confident, useless answer.
Common questions
Can we simply check by asking a model about ourselves?
You can, but treat it as anecdote rather than data. Answers vary between sessions, phrasings and users, and asking about your own company in the language you work in is close to the least representative test available. If you do it, ask the question your foreign buyer would ask, in their language, without naming yourself.
Do we need separate pages written specifically for models?
No, and pages built that way read badly for people without helping anywhere else. What changes outcomes is the character of the material: checkable specifics, evidence rather than assertion, consistency across every source that mentions you. That is ordinary good technical writing, and a human buyer benefits from it first.
Our Danish pages are cited and our English pages are not. Why?
Usually because the domestic field is small enough that you are one of few candidates, while the English field offers dozens. Beyond that, the substantive material often exists only in Danish, so the English tree presents a thinner version of a company that is not thin. Moving that evidence across is the fix.
How often should we look at the visibility figure?
Quarterly is enough, and monthly is already close to reading noise. What you are watching for is direction across several readings taken the same way, not a change between two adjacent points. Anyone reporting this weekly is reporting method variation.
Is any of this worth doing before our English section ranks properly?
Partly. Getting evidence into English and reconciling contradictory figures helps both layers at once, so that work is never wasted. Measuring the answer layer should wait until the pages are indexed and ranking somewhere, because until then you are watching a number with nothing underneath it.
Where to start
Begin with an inventory rather than a tool. List the ten documents that would convince a sceptical foreign engineer that your company can do what it says, and mark which of them exist in English on a page rather than inside a downloadable file. On most Danish export sites that exercise ends with more marks in the wrong column than anyone expected.
Then reconcile your sources. Take your three most important products and compare how they are described on your English pages, on your distributors' pages and in any directory that lists you. Where the figures disagree, decide which is right and correct the rest. Dull work, direct payoff, and it improves how a person reads you as much as how a model does.
Only then is there any point in watching a figure move. The market context and content-gap views sit in the same workspace as Search Console analytics, rank tracking and indexing, so a description of your market and the performance data underneath it can be read in one sitting. To see how your own domain is described rather than take the argument on trust, sign in and run the market circle on your domain, then set the result aside and repeat it in three months.
