Ask an assistant to name suppliers for a specific machining process in southern Germany and it will produce three companies. In our experience the international group with a plant in the region is almost never among them — not because it is worse, but because nothing published in German ties it to that process.
This layer sits above ordinary results rather than replacing them. A country site can hold a respectable position for a term and still be absent from every generated answer about the same subject. Both are forms of presence; neither compensates for the other.
For a subsidiary the asymmetry is sharper than for a domestic firm. The group name may be recognised across three continents and mean nothing to a purchasing department in Esslingen — and the question that produces a generated answer is precisely the one asked by somebody who does not already know who to call.
Why brand recognition abroad does not carry over
An ordinary results page lists ten entries and leaves the choice to a person. A generated answer performs the selection and names two or three. That converts a list where seventh position still has value into a sentence where it does not exist.
What gets selected is not the most heavily optimised material but the least ambiguous. A selection needs attributable facts: which process, which tolerance class, which batch range, which certification. A page describing a global footprint and a commitment to engineering excellence supplies none of them, and its international standing does not substitute.
What can be examined in the absence of a counter
Without a counter, what remains is disciplined observation: ask the questions, keep the answers, look for patterns. The generative market analysis in the Semalt workspace does this across six surfaces, describing the field the way a model describes it rather than the way a results page ranks it.
Six views, and where each becomes useful
For a subsidiary the second surface is usually the one that changes the conversation with head office.
- Market Circle placement. Three tiers: leading, middle, niche. It reports distance from the names a model reaches for first, not quality of manufacture — a distinction worth stating clearly before showing it to anybody senior.
- The model's own description of the domain. Positioning, an estimate of traffic, and openings it identifies. Placing this next to the corporate boilerplate is the exercise that most often changes minds, because the gap is visible in one screen.
- The questions themselves, sorted by intent. Not term lists but phrased questions, separated according to whether somebody is orienting themselves, comparing options or preparing to buy.
- Existing pages flagged as leverage. Pages already published where additional content or better internal linking would pay. In group systems these are almost always three levels down the country menu.
- Where competitors are complete and you are not. Subjects covered exhaustively elsewhere with no equivalent in your own material.
- One consolidated figure for AI-assisted search. A single number across the layer — and exactly the item the three caveats in this article govern.
The second surface has a particular use in a subsidiary. When a model describes the domain in terms of the group's global business rather than the local plant's actual capabilities, that is not a modelling error — it is an accurate reading of what has been published. It also happens to be the clearest available evidence that the country site needs locally written pages, and it arrives from outside the organisation, which makes it considerably more persuasive than the same argument made internally.
The limits, stated before the number is used
A consolidated figure remains useful for as long as its origin stays in view. It is produced by running questions, recording answers and counting how often a domain appears in them — a set of trials, not a census.
Reading those three at once might suggest the subject is not worth pursuing. In fact they are the reason it can be pursued at all. A figure whose limits are declared will survive a group review meeting. One presented without them holds until the first analyst asks how it was calculated — which, in a group finance function, is the second question.
What makes a capability page usable
Watching which pages get drawn upon produces a short and unglamorous list. It has more in common with sound technical documentation than with marketing copy.
| Corporate phrasing | Local replacement | Why it works |
|---|---|---|
| "precision engineering" | "IT6 achievable, IT8 held in series" | can be compared against a rival |
| "flexible volumes" | "from 5 pieces, series to 20,000" | filters out enquiries you would decline |
| "short lead times" | "standard parts in 10 working days" | supplies a figure a selection can use |
| "certified to international standards" | designation, scope and certifying body in text | answers an audit question directly |
| "global presence" | "machining in Waiblingen since 1998" | attaches the group to a place and a date |
None of these edits is an optimisation in the usual sense. Each replaces an adjective with information the plant already holds. The effort is an afternoon, and the benefit extends well past the newer layer: the same figures answer the questions sales currently answers by telephone. Where these pages belong in the structure is covered by our content strategy.
Getting locally specific text past a brand guideline
Most subsidiaries know all of this already. What stops them is not insight but approval: corporate communications owns the tone, and tolerance figures do not appear in the brand guideline.
Rewriting the capability page
Touches template text owned by the group. Requires sign-off, a review cycle and usually a translation memory update.
- Months of elapsed time
- Frequently ends in a compromise
Adding a technical detail page
New pages below the capability page, containing figures rather than claims. Adds to the site instead of altering it.
- Rarely contested by brand owners
- Live within weeks
The distinction matters more than it appears. Brand guidelines govern how a company presents itself; they seldom govern whether a tolerance table may be published. Framing the work as adding technical documentation rather than as rewriting marketing text moves it from a governance question to an operational one, and operational questions get answered by the plant.
What belongs to the brand and what belongs to the plant
Drawing this line once removes most of the friction that stops country content from ever appearing.
- Tone, naming and visual identity belong to the group. Nobody sensible disputes this, and conceding it explicitly at the start removes the suspicion that the country office is trying to go its own way.
- Verifiable technical facts belong to the plant. A tolerance class held in series is not a claim about the brand; it is a statement about a machine, and the people who can verify it work in the building.
- New pages are cheaper to approve than edited ones. An addition below an existing page rarely triggers a translation-memory update or a full review cycle. The same content proposed as a rewrite frequently does.
- Published figures need an owner and a review date. The one condition brand owners reasonably insist on. Naming a person and an annual date usually settles the objection in a single meeting.
Subsidiaries that have this conversation once tend not to need it again, because the boundary turns out to be uncontroversial as soon as it is drawn. Those that never have it spend years assuming that publishing a batch range would require corporate approval, and quietly conclude that nothing can be done locally. That assumption, rather than any brand rule, is what keeps most country sites indistinguishable from one another. The page recommendations in the workspace are useful evidence in that discussion, because they name specific existing pages rather than proposing a redesign.
Where the questions come from
Everything here begins with the questions that actually arrive. They read nothing like search terms: longer, hedged, full of conditions. "Could you machine a small batch in 1.4404 if the drawing is not released for another three weeks?" is not a query — it is what a project manager writes at half past four on a Thursday.
- From the enquiry inbox. Requests state the conditions explicitly: material, quantity, date, inspection scope. Those four items are precisely what most capability pages omit.
- From tender documentation. Written in the language of purchasing rather than of marketing, including the standard designations people later search for.
- From lost bids. The most valuable and least popular source. A stated reason for losing almost always names something that could have been published.
- From follow-up questions after a quotation. If three clarifications are needed after every quote, those three gaps belong on the page — in the customer's wording, not the group's.
Twenty such formulations from one quarter carry more weight than any term list pulled from a tool, and they cost nothing beyond the discipline of writing them down as they arrive. The side effect is worth having on its own: a page that closes those gaps reduces telephone time and improves the fit of incoming enquiries, because unsuitable prospects drop out earlier.
"Turning and milling, batches from 5 to 20,000, in 1.4404 and 1.4571"
Three facts a selection can match against a question. All three come from the shop floor.
- Matches a material query
- Matches a batch-size constraint
"Comprehensive machining solutions for demanding industries"
Accurate, unfalsifiable, and identical to what eighteen sister sites say in their own languages.
- Matches no specific question
- Indistinguishable from competitors
The lower card is not badly written. It is written for a different purpose — to present a group consistently — and it performs that purpose adequately. The mistake is expecting it to do a second job it was never designed for. Both cards can coexist on the same site, provided somebody has decided which page carries which one.
A rhythm that matches the evidence
The evidence here is thinner than for conventional metrics, which rules out a weekly cadence. A quarterly cycle has proved about right: looking more often means watching randomness, looking less often means noticing a shift too late to respond.
| When | What happens | What it produces |
|---|---|---|
| Week 1 | Go through the previous quarter's enquiries; copy twenty phrasings verbatim | wording from the market, not from the brochure |
| Week 2 | Check which of those questions already has a page | a list of gaps |
| Week 3 | Read the model's description and the competitor coverage | subjects effectively owned elsewhere |
| Weeks 4–8 | Write or rework two or three technical pages | figures in place of adjectives |
| Quarter end | Compare placement and the consolidated figure | a direction, explicitly not a result |
Week three carries a caveat worth repeating. Where the model's description of the company diverges from the company's own, the instinct is to call it wrong. The more productive response is to ask what it could have been derived from, and the source is invariably the published text. Read that way, the divergence is not an error but precise feedback. The analysis sits alongside the conventional figures in My SEO, so a text change and any subsequent movement appear on one timeline.
One further note on the collection step. Enquiries arriving in English at a German subsidiary are worth separating out rather than lumping in, because they usually come from a different role — an international purchasing function at a customer that also operates across borders. Those enquiries phrase the same requirement differently and often carry a larger order value. A country site that answers only the German phrasing leaves that group to find its way to the group domain, where nothing local is stated at all. Two versions of the same technical page, one per language, is not duplication in this case; it is two audiences with genuinely different vocabulary. The campaign side keeps both term sets in one pool, which makes the split visible rather than accidental.
What this produces, and what it does not
Before anybody commits effort, it is worth being clear about the output. It is not a metric for a group dashboard. It is a country site that states what the plant can actually do — and therefore performs better in both layers at once.
That is a quieter promise than the phrase "AI visibility" implies, and it lasts longer. Working the other way round — pushing an inferred figure upward directly — means optimising against a set of trials, and the advantage evaporates when the trials change. What holds is what held before: an orderly structure, published figures, a regular review. The technical half of that is covered under technical SEO, and which term groups correspond to which order values is settled during keyword research.
Open the market analysis for your country site
Questions from group and country marketing
Can we measure how often we are cited?
Not as a measurement. There is no counter for a provider to read. What is available is running the same questions on a schedule, keeping the answers and inferring a trend. The inference indicates direction and is intended for that; it stays a set of trials and will not bear the weight of a click statistic.
Our group brand is well known. Does that not carry into German answers?
Less than most groups expect. Recognition transfers where the question names the company; it does not transfer where the question names a process and a region. In that case the selection is made from what has been published about that process in that language — and a global capability statement translated into German supplies nothing attributable. This is the single most common surprise for subsidiaries running this exercise for the first time.
Should we write content specifically for models?
No, and the overlap is the useful part. What makes a page usable — one question per section, figures with conditions attached, a stated scope, named limits — is the same set of properties that satisfies a purchasing engineer with ten minutes. Writing for the machine while forgetting the reader loses on both counts.
Will we lose traffic if assistants answer directly?
Partly, and specifically on pages that only ever supplied a fact. Somebody checking which tolerance a standard prescribes clicks through less often than before. Where a purchase is involved it works differently: a supplier that gets named then gets verified. The pages affected are therefore those whose contribution to the business was already marginal.
How often should the consolidated figure be reviewed?
Once a quarter is sufficient. More frequently and you are following the variance of the trials rather than the movement of the market. More informative than the figure itself is its relationship to the conventional metrics: if clicks and positions hold steady while the placement shifts, that warrants a closer look. If both move together, a single cause usually explains them.