An importer in Sao Paulo asks an assistant which European operators can take class 3 flammables into bonded storage. Back come three company names, a sentence of description each, and no sources. He has never set foot in the Netherlands, recognises none of the three, and forwards the list to his purchasing department as if it were research.
That exchange is the layer this article is about. It sits above the ten blue links rather than replacing them, and for a supplier in a port cluster it behaves unlike classic ranking: the output is a shortlist, delivered as prose, and the reader usually has no way to test a claim in it.
Three names arrive and none of them can be checked
A results page is a set of offers with visible authorship. You can see who published each one, notice that two are directories and one is a rival's blog, and form a view before reading a word. A generated answer strips that away and hands over a conclusion: three firms, described confidently, reasoning discarded.
A buyer with local knowledge loses less to that. Someone who has worked a lane for fifteen years reads the three names and thinks that one is really a broker, that another was bought last year, and that the third declines that hazard class. A cargo owner two continents away has none of it. He has a paragraph, and the paragraph reads as settled.
Can anyone do this at all
A capability question, asked before any supplier is known, usually phrased around the cargo and the constraint rather than around a company.
- Answered with a category and two or three examples
- Where a name is either present or permanently absent
What does the rule require
A compliance question about a licence, an authorisation, a segregation rule or a documentary obligation the buyer must satisfy downstream.
- Answered from published regulation and trade explainers
- Suppliers appear only if they explain the rule themselves
Which of these two is better
A comparison, often between named firms or between two ports, where the model synthesises whatever descriptive text it has absorbed about each.
- Thin coverage produces a thin and unflattering paragraph
- Silence is read by the asker as insignificance
How does the procedure work
A process question about a transit movement, a customs procedure, an acceptance protocol or a survey, asked by somebody preparing to buy.
- The clearest published explanation tends to be quoted
- Authorship credit is inconsistent and often dropped
Each of those four is a live buying moment. None produces a click you would recognise in a report, and two may produce no visit at all before your name reaches a procurement inbox.
A remote buyer has nothing to hold the answer against
Every industry is exposed to generated answers. A port cluster is exposed unusually early, and the reason is the geography of the demand rather than the technology. The buyer is somewhere else: a chartering desk in the Gulf, a purchasing office in the American Midwest, a manufacturer in northern Italy arranging distribution from a European seaboard. They will not visit, they have no colleague who used you in 2018, and they cannot ask around a local network because they have none. Whatever the answer says is their entire evidence base until a supplier questionnaire is issued.
A second amplifier follows from how procurement works here. Whoever asks the question rarely awards the work. A planner or an analyst assembles the candidates, pastes the names into an email and hands the file to somebody who never saw the question. By the time the list is discussed, its provenance has evaporated.
What a generated answer is assembled out of
Nobody outside the companies running these systems can state the recipe. What can be observed is which kinds of text tend to be reproduced, and the pattern is unsurprising: plain statements of verifiable fact, written in the vocabulary the asker used, published where a crawler can reach them, and corroborated somewhere other than your own website.
| Input | What it contributes | How much you control it | Failure mode |
|---|---|---|---|
| Your own service pages | The factual spine: scope, capability, authorisations | Entirely | Written as adjectives, so there is nothing to extract |
| Trade directories and port guides | Categorisation and the fact that you exist | Partly, through submissions and corrections | Stale entries listing services you dropped years ago |
| Association and registry listings | Corroboration by a third party the model treats as neutral | Only by maintaining the membership | Listed under a former trading name |
| Trade press and case coverage | Narrative detail and a sense of scale | Indirectly | The only coverage is a decade old |
| Competitor and platform pages | The comparison frame your name is placed inside | Not at all | Your category is defined by a rival's marketing |
| Regulatory and standards text | The definitions the answer leans on | None | You use in-house wording nobody else uses |
The last row deserves a moment. Firms in this trade often keep a private vocabulary: a house term for a handling method, a shorthand a founder invented. It is invisible to a system matching a buyer's regulatory phrasing against published text. Where the legislation says one thing and your site says another, you have opted out of the question.
- Use the buyer's noun, not yours. Whatever the hazard class, the customs procedure or the standard is officially called, that is the phrase to put in the heading, with your house term in brackets afterwards if you must keep it.
- State limits as plainly as capabilities. Which classes you decline, which tonnages you cannot take, which countries you do not clear. A precise boundary is easier to quote correctly than an open-ended claim.
- Put numbers where a sentence can pick them up. Clear heights, plug counts, tank capacities, licence numbers, approval dates. A figure inside an image is a figure that does not exist.
- Corroborate off your own domain. A registry entry, a membership listing and a trade publication that agree with your site are worth more than three more pages on your site saying the same thing.
What the generative research module actually does
Because none of the engines publish anything usable about how often a domain is drawn on, the practical tooling in this area works by modelling rather than by counting. That distinction runs through everything below, and it is worth keeping in view while reading the feature list.
Generative market research over one domain
For a supplier who needs to know how a model would describe the market it sells into, and where the description leaves them out.
- A competitiveness score with a market circle. Rivals sorted into top tier, mid tier and niche, which in a port cluster usually separates the pan-European platforms from the operators you actually bid against.
- A model-written market context. Positioning, an estimate of traffic and a list of openings, generated for the domain rather than typed in by you, and therefore useful mainly as a mirror.
- Query research with intent classification. Candidate phrases grouped by what the asker is trying to accomplish, which is how a compliance question gets separated from a price-shopping one.
- Pages flagged as levers. Documents the model marks as worth expanding or worth linking to internally, plus an analysis of where rivals are strong and where the published explanations run out.
The most useful of those is the content-gap analysis, because it produces work rather than opinion. Where the model can describe a procedure but cannot attribute the explanation to anyone in your segment, that gap is an article your operations team could dictate in twenty minutes.
Used that way, the generative research views function as a briefing about the shape of the market as a model perceives it. That is a genuinely new thing to be able to see. It is not, however, a record of anything that happened.
A global visibility figure and what it is honestly worth
The panel does report a global visibility value across the generative landscape, and it is reasonable to look at it. The mistake is treating it as the same species of number as an impression count. An impression is a logged event; the total is a record of things that happened. A generative visibility score is the output of a model that samples prompts, reads what comes back and infers how prominent a domain appears. Change the sample, the phrasing or the day and the figure moves without anything about your company having changed.
| Figure | Where it comes from | What it can support | What it cannot support |
|---|---|---|---|
| Clicks and impressions | Logged events from a verified property | Trend claims, comparisons between pages | Anything about answers that produced no visit |
| Average position | Aggregated over recorded impressions | A rough sense of standing per segment | A statement about one buyer's screen |
| Competitiveness score | Model output over a market description | Tiering rivals, spotting an unnoticed platform | A ranking anyone else would reproduce |
| Generative visibility value | Inference from sampled prompts | A months-long direction of travel | A count of citations, of any kind |
| Content-gap list | Model comparison of coverage | An editorial queue with priorities | Proof that filling a gap changed an answer |
The one place this figure must never appear
Suppliers here answer a great many formal questionnaires: vendor registration forms, tender annexes, insurance schedules, prequalification packs. All share one property that makes them dangerous ground for a modelled number. Everything entered is treated as a representation of fact, checkable and durable.
The rule is a boundary, not a prohibition. Use the figure internally to decide where to publish next, cite it in a marketing review with the caveat attached, then stop. The tender response contains only what a third party could verify.
Evidence a stranger can check
Facts that exist independently of your reporting and survive being audited by somebody who has no reason to trust you.
- Authorisation and licence numbers with issue dates
- Certification scheme, certificate number, expiry
- Volumes handled, with the period stated
Numbers that dissolve under questioning
Modelled or sampled values with no published method, presented in a document where every figure is read as a claim.
- A visibility score with no issuing body
- An estimated citation count in generated answers
- A share of voice nobody outside can reproduce
Building pages a system can safely repeat
If the reporting side is limited, the production side is not. Whatever makes a page usable by a summarising system also makes it usable by the cold second reader in a procurement conversation: a factual claim with a boundary around it, no adjectives, and the figure and its unit in the same sentence. Neither reader can ask you a follow-up question.
The quotable capability block
A structure for service pages that has to survive both a paraphrasing model and a purchasing manager reading it cold.
- One sentence that could stand alone. Name the service, the cargo it applies to and the geography it covers in a single line that still makes sense with everything around it deleted.
- The constraint, stated as a number. Capacity, height, temperature range, hazard classes accepted, turnaround. Written out as text, never rendered inside a diagram or a photograph.
- The authority behind it. Which authorisation permits the work, issued by whom, valid until when. This is the element that turns a claim into something a stranger can act on.
- The exclusion. What falls outside the service. Stating it costs you nothing, prevents the enquiries you cannot serve, and makes the rest of the page markedly more credible.
Publishing the result can be done by hand or handed over. Campaign automation at either tier differs mainly in who picks the phrases and who signs off edits: 149 USD a month per domain for the automated level, 500 USD for the managed one, which adds hand-picked keywords, a domain-rating target for placements and human review before changes ship.
Progress here is slow to appear and slower to attribute. Ordinary ranking typically shows first movement after four to eight weeks; this layer lags behind that, since the systems must encounter, absorb and then choose to reuse what you published. A chronological project feed serves better than a dashboard, recording what was published, placed and changed in the order it happened — a record worth keeping beside the notes in the English article index.
Questions this raises
Can we find out how often an assistant mentions our company?
Not as a count, no. The systems producing the answers do not publish that data and no external tool has access to it. What you can do is ask the questions your buyers ask, in the languages they ask them in, keep the answers with dates, and repeat the exercise quarterly. That is sampling, and its limits should be written down next to the results.
Our competitor is named in these answers and we are not. What is the likely reason?
Usually corroboration rather than quality. They are described in the same terms by a directory, an association register and a trade publication, so the same facts appear in several independent places. A firm whose entire public description lives on its own website is a firm the model has one source for.
Should we write pages aimed at these systems rather than at people?
No, and the distinction is less useful than it sounds. What makes a page quotable is factual precision, a stated boundary and a number with its unit attached, which is exactly what a purchasing manager reading three quotes needs. Pages written for machines and stripped of judgement tend to fail with both audiences.
An answer described our capabilities incorrectly. Can it be corrected?
Not directly, and there is no channel to request a correction. What you can influence is the input: correct the stale directory entry, update the register listing, and publish an unambiguous statement of what you do and do not handle. The error usually traces back to an outdated third-party description rather than to your own site.
Is this replacing classic search for our kind of buyer?
It is adding a stage, not removing one. The generated answer produces the shortlist; the buyer then opens the pages of the firms named, and increasingly searches those company names directly. Losing your ordinary ranking to chase the new layer removes the ground the new layer stands on.
Where to begin, and what to write down
Start by acting like the buyer you never see. Take the eight questions your commercial team fields most often, phrase them as a foreign cargo owner would, and put them to two or three assistants. Record the answers verbatim with the date. This measures nothing; it collects evidence about how your market is described to people who cannot check it.
Read the results for three things: whether you appear, whether the description is accurate, and which firms are named instead. The third column tends to be the surprise, because it often holds a platform or a directory rather than an operator — a sign that the question is being answered from aggregated listings rather than from supplier pages.
Then fix inputs rather than chase the score. Correct the directory and register entries, publish the four-element capability block on every service page, write down what your operations team explains on the phone, and put the buyer's official vocabulary in the headings. The technical groundwork sits with our services; the writing discipline is simply that of a good tender annexe, applied to a public page.
If you would rather look at the market description than take it on trust, open the panel and run the generative research over your own domain, then set the model's account of your market beside your own. Where the two disagree, the gap is not necessarily an error on either side. It is a fair map of what a stranger with three quotes on the desk can currently find out about you.
Ready to Improve Your SEO?
Get a free SEO audit and find out how we can grow your Rotterdam business.
Request a Free SEO Audit