A procurement lead in Stuttgart asks an assistant to name three Czech firms that could take over a subassembly. Back come three names, a sentence of reasoning each, and nothing to click. He has never set foot in the country and cannot verify a word of it. That paragraph is now the long list.
Search results have always been a filter. What is new is that the filtering happens before the reader sees anything to filter. A generated answer arrives already reduced: not two hundred candidates ranked, but three named and the rest silently discarded. For a supplier selling capacity across a border, that reduction is the commercial event, and it happens somewhere you have no account and no log.
Three names and no way to check them
Picture the buyer honestly. He runs a subassembly that has been made in one place for eleven years, and the supplier has just told him the line is being retired. He needs a second source qualified within two quarters, and he is not a specialist in your process. His first move is no longer a directory but a question typed in ordinary language: who can do this, at roughly this volume, to roughly this standard, within reach of his logistics team.
What comes back is fluent, confident and unsourced. It names firms and attaches a reason to each — long-established, works with automotive customers, holds the relevant certification. It may be right. It may equally be describing a company that closed in 2019. The buyer cannot tell, and he does not treat that as a problem: he takes the three names as a starting point and goes looking for each one.
- The question is phrased as a brief, not as a keyword. Volume, standard, region and commercial model arrive in one sentence, the way they would be spoken to a colleague.
- The answer is a shortlist, not a ranking. Three or four names. There is no position eleven to climb to and no second page to appear on.
- The reasoning is reconstructed, not quoted. The sentence explaining why a firm belongs there is assembled from whatever the model absorbed, and may attribute a capability nobody claimed.
- The next search is your company name. Being named produces a branded search hours later — visible in your own reporting even when the generated answer is not.
A machine reading your capability statement
The material a model holds about a mid-sized Czech engineering firm or laboratory is thin and oddly distributed. It is not your brochure. It is whatever text about you exists in a language the model absorbed a great deal of, which for most Czech suppliers means the English half of your own site plus a handful of third-party mentions.
That matters more than it sounds. A firm with a rich Czech website and a four-page English section has told the machine almost nothing. The Czech material exists, and models do read Czech, but there is far less of it in circulation, and the buyer asked his question in German or English. The bridge between the two has to be built on your side.
Stated limits and thresholds
A minimum batch, a maximum envelope, a temperature range, a cleanroom class. Numbers with units are the easiest thing on your site for a machine to extract and reuse correctly.
- Written as prose, not buried in a drawing
- Repeated on the page claiming the capability
Named standards with scope
A certification named in full, with the processes and the facility it covers. Scope is the half nearly always omitted, and it is the half a buyer needs.
- Which site, which processes
- Held since when, renewed when
Capability shown only in pictures
A machine list rendered as a photograph of the shop floor, or an equipment table exported as an image. A human reads it instantly; a model reads nothing.
- Common on manufacturing sites
- Cheap to fix, rarely noticed
The unnamed reference customer
"A leading German automotive group" gives a machine nothing to attach to a sector query, because the sentence contains no term anybody would search for.
- Name the sector where the client cannot be named
- State the component class and volume band
The pattern behind all four is the one that governs ordinary ranking, only harsher: a machine can repeat only a claim it can locate and parse. Your English pages are not marketing here. They are the source document a system will paraphrase to a stranger deciding whether to send you a request for information.
Six views, one score, and what it is made of
The generative research section of the unified Semalt panel runs to six views. It is worth being precise about what each one does, because the difference between an observation and an inference is the whole subject of this article.
| View | What it produces | Nature of the figure | Use to a supplier |
|---|---|---|---|
| Competitiveness score with market circle | Your domain placed against top-tier, mid-tier and niche rivals | Inferred position, not a count | Seeing which tier you are grouped into |
| Market context for a domain | Positioning, a traffic estimate, stated opportunities | Generated description | Reading how a machine would describe you |
| Query research and intent classification | Phrases grouped by what the searcher was trying to do | Classification of real phrases | Separating sourcing from research |
| Content gaps and competitor strengths | Subjects rivals cover that you do not | Comparative, from observed pages | Deciding what to write next |
| Pages flagged as levers | Pages worth expanding or linking internally | Recommendation | Ordering a small content budget |
| Global visibility value | One figure for standing in the generative landscape | Inferred index across the portfolio | Watching direction over quarters |
Read that middle column twice. Four of the six produce something derived. That is not a flaw in the tooling but the honest consequence of instrumenting a channel that emits no logs — and it changes what you are entitled to say about the number afterwards.
Use them comparatively and slowly. One reading tells you nothing. Four quarterly readings, with the market circle showing which tier you sit in and which domains sit above you, tell you whether the gap is closing. Moving from the niche group into the mid-tier over a year means something; three points of movement in a fortnight does not.
Nobody publishes how often you were named
Sooner or later somebody in a management meeting asks the obvious question: how many times were we mentioned last quarter? It is a fair question. It is also unanswerable, and saying so plainly does less damage than improvising an answer.
Sample deliberately instead. Write down the fifteen questions a buyer would actually ask before contacting a supplier in your category, phrased as briefs with volumes and standards in them, in the languages your customers use. Ask them. Record which firms are named, in what order, with what claim attached. Repeat the same fifteen next quarter.
- Fix the wording and never improve it. The value lies in comparability across quarters. A better-phrased question in month six destroys the comparison you were building.
- Record the competitors, not only yourself. Which Polish or Portuguese firm keeps appearing is more informative than whether you did, and it is the part your commercial team will act on.
- Log the claim, not just the name. When a model attributes something to you that is wrong, that error came from somewhere — usually a sentence on your own site or an old directory entry.
- Keep it in a spreadsheet, not a slide. Exports run to CSV or JSON at up to 10,000 rows and to PDF at up to 250 rows; this exercise belongs in the first format.
None of that produces a metric. It produces a file of observations, which is a more modest and far more defensible thing to carry into a management meeting than a number nobody can source.
Keep it out of the supplier questionnaire
Here is the failure specific to companies selling into regulated supply chains, and it is always done with good intentions. Somebody sees a strong visibility figure, reads it as market standing, and drops it into the capability statement. Six weeks later it sits in a pre-qualification questionnaire next to turnover, headcount and certification numbers.
The rule is short enough to hand to a marketing team: if a number cannot survive the question "who issued this, and how would I confirm it?", it stays inside the company. Search figures are operational instrumentation. They belong in a marketing review and nowhere near a file a customer's quality department will keep.
| Document | What belongs in it | What does not |
|---|---|---|
| Pre-qualification questionnaire | Certifications with scope, audited figures, references | Any visibility or ranking score at all |
| Capability statement | Processes, tolerances, capacity, equipment, standards | Marketing performance metrics of any kind |
| Internal marketing review | Impressions, band movement, tier, inferred scores | Nothing — this is where they belong |
| Public website | Verifiable claims a buyer will check | Numbers whose source you cannot name in a sentence |
The work that makes a machine able to name you
Strip away the novelty and the levers are unglamorous. A model names firms it has met described in specific terms, in the language of the question, on pages it could reach and parse. Each of those conditions is under your control.
Reachability comes first, because it is cheapest to rule out and most often broken. A page absent from any search index has, in practice, been absorbed by nothing. If your English capability pages were added late, linked from few places and never submitted, discovery is the problem and everything else is theoretical — which is where technical work on the site earns its place before a content budget is spent.
Then specificity. The intent classification in the research views helps directly, separating a phrase typed while sourcing from a near-identical one typed while writing a report. A page built around the first kind — volumes, standards, lead times, commercial model — reads as an answer to a machine and as a qualification document to a human.
Write the qualification page
One page per capability that a buyer could paste into an approved-vendor form: process, scope, standard, limits, batch range, lead time, the facility it happens in.
- Every number stated in prose
- No claim without a scope attached
Fix the record elsewhere
Association listings and trade-body profiles often carry a description of you written years ago. That text is readable, and it is frequently the source of a wrong attribution.
- Correct the outdated entries first
- Consistency beats volume
Writing for the machine
Pages padded with question headings and restated phrases read as filler to both audiences. Nothing here rewards text that says less in more words.
- Thin pages stay thin
- They also fail the buyer who arrives
Chasing the score weekly
An inferred index read every Monday will move, and none of that movement is attributable to anything you did.
- Weekly readings invent false causes
- Tier changes are the signal
Running it as a campaign, not a project
This work is continuous rather than a launch. Both campaign levels carry the generative research views; the difference is how much judgement is delegated.
AutoSEO — delegated judgement
For a firm whose capability vocabulary is standard enough to be found automatically.
- Candidate phrases found and ordered with no list to maintain. The pool draws on Search Console, live result pages and your own seed terms, and each candidate is approved, rejected or deferred one at a time.
- Placements built automatically. Drawn from a partner network of more than 230,000 websites, with no manual selection step.
- The analytics stack alongside. Eight Search Console views and six rank-tracking views in the same account as the generative research, so an inference can be checked against an observation.
FullSEO — judgement kept in the room
For a supplier where a misdescribed capability is a commercial risk, not a typo.
- Phrases chosen by hand, with automatic fallback. You approve the terms describing what you are actually accredited to do; anything you do not reach is picked up automatically.
- Placement against a domain-rating target. A quality threshold set in advance rather than accepted as available.
- Review before anything is published. A team of specialists, developers and writers sits between a suggestion and a live page — which matters when a draft describes a scope you do not hold.
Continuity is handled by the Stream assistant inside My SEO, which keeps answers, automatic reports, newly placed links with donor rating and traffic, to-do items and campaign news in one chronological line. A router model decides per question which data blocks to load — none, or up to three, by relevance — and the feed is searchable, which is how a sampling exercise from two quarters ago gets found again.
Common questions
Should we stop caring about classic rankings?
No, for a mechanical reason. Being named in a generated answer sends the buyer looking for you by name, and what he finds is a result page. The two layers are stacked, not substituted. A firm invisible in ordinary results is also a firm with little material for a model to have absorbed.
Our Czech pages are far stronger than our English ones. Does that help?
Only inside Czech-language questions. A German or English buyer's question is answered from material in that language, and Czech is a small share of what most models absorbed. Domestic strength does not carry across; the English section has to stand on its own material.
A model described our company incorrectly. Can that be corrected?
Not directly, and there is no correction form. What you can do is find and fix the source: usually an outdated page of your own, an association listing or a directory entry nobody has read in years. Consistent, current descriptions in several places are the only lever, and they work slowly.
How often is the sampling exercise worth running?
Quarterly. Monthly is worse rather than better, because the noise between runs gets read as change. Fix the wording once, keep the results in one file, and compare like with like.
Can we see the generative views for several domains at once?
Yes. The account is multi-tenant, site tags act as a global filter, and individual sites can be released to another email address, so an agency or an export manager sees one property and nothing else. Further reading sits in the English articles section.
What to do with the next hour
Write the fifteen questions. Not keywords — questions, phrased the way a buyer with a problem and a deadline would phrase them, with the volume and the standard in the sentence. Ask them, and write down every firm named. That list is the competitive set you are actually in, and for most Czech suppliers it holds at least one company nobody in the building had heard of.
Then take the capability that earns you the most money and read its English page as the machine would: can you extract a scope, a limit, a standard and a lead time from the prose alone? If not, that is the afternoon's work. The content-gap analysis is a useful second pass once the first page is right.
To see how a machine currently describes your domain — its inferred positioning, its tier, and the competitors it groups you with — open the dashboard and run the market context view against your English property. Read it as a description written by a stranger who has only your website to go on, because that is what it is. The gap between that description and what your company actually sells is the brief for the next six months of writing.