A project manager six months away from opening a plant somewhere in Sweden asks a model who supplies the thing she needs, and gets three names in a tidy paragraph. Nothing in that paragraph tells her whether any of the three can be on site in the municipality she has shortlisted, which is the only fact the contract will actually turn on.
She is not being careless. She has no colleague here to ask, no county names in her head, and two shortlisted industrial sites four hundred kilometres apart. English is her working language, a chat window is where she starts, and the answer reads as though the question has been settled.
It has not been settled. It has been summarised, and summarising is the operation that removes the distinction her project depends on. A supplier who can be there on Thursday and a supplier who would need a fortnight and a subcontractor both read, in a generated paragraph, as companies that serve Sweden.
This piece works through what can and cannot be observed about that layer, using the generative research section of the Semalt panel, and it is deliberately blunt about where observation stops and estimation starts.
Three names, and no way to tell them apart
Consider what she can verify from the paragraph in front of her. Three company names, perhaps two or three sources beside them. What she cannot see is which of the three has people within driving distance of the site she is leaning towards, because nothing in the summary is organised that way and nothing in her own knowledge fills the gap.
Her domestic counterpart never lands in this position. A buyer who has worked here for fifteen years asks a former colleague, or types a town name into a search box, and the geography does the filtering before any tool gets involved. The regional question is answered by the phrasing of the query itself.
The inbound enquirer has no such phrasing available. She asks at country level because it is the only level she knows, and gets an answer built to match the question.
- The query is national because the asker is new. Nobody names a county they have not yet chosen, so the question arrives at the widest possible resolution and the answer follows it there.
- The stakes sit at the opposite resolution. Everything expensive about the decision — response time, spare parts, an engineer who arrives before the line stops — is regional.
- There is no second query. A list invites comparison, so people scroll. A paragraph invites acceptance, so the three names become the shortlist and the fourth company is not in the room.
What summarising takes out on the way through
A ranked list of ten results preserves difference almost accidentally. Each entry keeps its own domain, its own title and often a place name sitting in that title where the reader can see it. The discriminating is left to the person reading, who is usually better at it than any summariser.
A generated answer does the discriminating first and then reports the conclusion. It discriminates on whatever the underlying documents made easy to extract, and what national websites make easiest to extract is national scope, since that is the claim those pages exist to assert. The regional texture — three depots, a delivery run that leaves at six, a service engineer who lives in the county — is either absent from the page or written in a form nothing can lift out of it.
The consequence runs both ways. A company with thin regional capacity can be named purely because its pages talk at the right altitude, while a company with real depots and real engineers is left out because everything specific it has to say sits in a case study nobody indexed.
| Aspect | Ranked list | Generated answer | Effect on a regional supplier |
|---|---|---|---|
| Who filters | The reader | The model | Place names in titles stop doing their job |
| How many win | Ten, unevenly | Two to four | Being fourth-best is closer to being invisible |
| What survives | Title, domain, snippet | Whatever could be stated in a clause | Delivery detail needs to be one sentence long |
| How scope reads | One claim among ten | The organising principle | National phrasing is quietly rewarded |
| What the asker does next | Opens three tabs | Writes to the names given | The enquiry arrives already shortlisted |
None of this makes the layer hostile. It is a surface with different physics, and a supplier whose advantage is geographic is the one it carries worst.
The list underneath did not disappear
It is tempting to treat this as a replacement. It is not one. The paragraph is assembled from documents that had to be findable first, so most of the classic work still applies unchanged: a page nothing links to and nothing has fetched is no more available to a summariser than to a search result.
So the layer sits on top rather than instead, and what changes is the shape of the reward. On a list, position five still earns a share of the clicks, and a patient company can live there for years. In a paragraph there is no position five; there are the names mentioned and everything else. Against that, several names fit in one answer, so the surface is less winner-take-all than the first position on a list has become.
Gradual, forgiving, place-aware
Progress shows up as movement from thirty to twelve to seven, and each step is visible and worth something on its own.
- Town names work in titles
- Partial visibility still converts
Binary, but not exclusive
You are in the sentence or you are not, though the sentence has room for three or four companies rather than one.
- Scope language is rewarded
- Improvement is hard to see week by week
For a supplier with uneven regional depth, both surfaces need different material. The Swedish pages go on serving buyers who know the map; the English pages, where the inbound enquirer lands, have to answer a national question without pretending the country is uniform.
What can be counted, and what is only ever inferred
Search Console gives you a log. Clicks, impressions, average position, the query that produced them, split by page, device and country, with the two-day delay already accounted for in the presets. Rank tracking gives you an observation: a position recorded against a keyword at a moment in time, repeated often enough to draw a curve. Both are records of things that happened.
The generative layer offers neither. No assistant publishes a feed of the occasions on which your domain was named. The queries are private, the answers are generated per person, and the same question asked twice can return different companies. There is no ledger, and nobody is withholding one.
What the panel does instead is sample deliberately: take questions a buyer might realistically ask, observe what comes back, weigh it against the competitive picture in ordinary search data, and express the result as a score. A reasonable method, and unavoidably a model reporting on models.
- Recorded. Impressions, clicks, positions, crawl visits and their timestamps. These have a source you can point at and a number that will still be there tomorrow.
- Sampled. What a set of representative questions returns today. Real observation, but of a slice, and the slice was chosen by somebody.
- Inferred. Competitiveness scores, market context, opportunity estimates. These are arguments expressed as figures, and they move when the reasoning behind them moves.
- Unavailable. How often you were named, to whom, on which day. Nobody sells this, and a vendor claiming to have it is describing an estimate.
What the generative research section actually contains
The AI Analytics views run to six, and each is more useful attached to a question you would otherwise be arguing about.
| View | The question it settles | How to treat the output |
|---|---|---|
| Competitiveness score, market circle | Which companies occupy the space we are asking about? | Directional, watched over quarters |
| Model-generated market context | How is our domain described by something that has never met us? | Read as a draft written by an outsider |
| Query research and intent | What is the person behind this phrasing actually trying to do? | The most immediately actionable view |
| Pages flagged as levers | Which existing pages are worth expanding or linking better? | A shortlist to argue with, not a work order |
| Competitor strengths and gaps | What are they answering that we are silent about? | Compare against your own delivery reality |
| Global visibility value | Is the direction of travel right across the portfolio? | Trend only, never an absolute |
Intent classification earns particular attention in a small language market. Swedish volumes are modest outside a few consumer categories, and the English establishment queries are smaller still — a plant opening produces a handful of searches from perhaps three people over four months. Volume is close to useless as a criterion at that scale. Intent is not: a comparison, a specification and an emergency want three different pages, whatever their monthly counts say.
The competitiveness score, and what it is honestly good for
A score with three rings drawn around it
For the quarterly conversation about whether the company is gaining or losing ground in a space nobody can measure directly.
- Top tier. The domains that turn up whatever the phrasing. For a national supplier these are usually the two or three largest names in the sector, and displacing them is a multi-year proposition rather than a campaign.
- Mid tier. The ring where the score can actually be moved. These are companies of roughly your size whose pages happen to be phrased more extractably than yours.
- Niche. Specialists who dominate a narrow question. A regional operator with four counties and a deep site frequently sits here, and outranks you on precisely the questions that convert.
As a quarterly orientation this is genuinely helpful: it shows which ring your rivals occupy and where effort has a chance of paying. As a target it goes wrong quickly, because the number is produced by reasoning rather than by counting.
The same caution governs how the competitiveness score and its market circle appear in reporting. A rising line across four quarters is fair to show a board alongside the recorded figures. A single number presented as a fact is not, and this goes wrong most often in a document written to persuade somebody.
Being described by something that has never visited you
Positioning, traffic estimate and openings, written from outside
For finding out what a summariser would say about your company before a customer finds out for you.
- Positioning as it reads, not as intended. The description is generated from what your pages actually say. Where it is vague, your pages are vague, and no amount of internal clarity compensates.
- An estimate of your traffic and openings. Treat both as prompts for a conversation rather than as figures. The value is in the openings it names that nobody internally had thought to name.
- Gaps against competitors. Questions your rivals answer somewhere on their sites and you do not. Most are not worth answering; two or three usually are, and they are cheap.
The first reading is uncomfortable for most companies, and the discomfort is the useful part. A manufacturer that thinks of itself as the specialist in one process reads a summary calling it a general supplier with a broad catalogue — and the summary is not wrong about the pages, only about the company. That gap is a content problem with a known fix.
The fix, for a business whose real advantage is regional, is not to assert coverage more loudly. It is to publish the delivery facts in a form that survives being compressed into a clause. A model can lift a two-hour response radius from the Sundsvall depot covering four counties out of a page. It cannot lift anything at all out of nationwide service with a local presence, because that sentence contains no information to lift.
Facts with a shape
Named towns, stated response times, a count of engineers, the day the run goes out.
- One clause each, in the body text
- On the page that ranks, not only in a PDF
Scope without substance
Coverage claims, lists of county names, adjectives about reliability every competitor also uses.
- Indistinguishable between suppliers
- Length without anything extractable
Work you have actually done
A named sector, a described problem, a stated outcome and where it happened.
- Put the place in the first paragraph
- English versions for the inbound reader
Evidence locked in the wrong format
References inside a brochure, capabilities in a slide deck, delivery terms known only to the sales team.
- Nothing extracts from an image
- Move it onto a page first
Stream, the project assistant, is where this gets worked through: a chronological feed per project carrying answers, automatic reports, newly placed links with their donor metrics, to-dos and campaign news. The way the assistant loads project data governs how much weight to give its replies — a router decides per question which blocks of real project data to pull in, between zero and three, from Search Console, rank tracking, campaign records or your own sources. A question that pulls in none is being answered from general reasoning. Exports run to CSV or JSON up to 10,000 rows and PDF up to 250 rows. The technical groundwork sits under our services.
Questions this raises once somebody has looked
How would we know whether a model has ever named us at all?
By asking it, repeatedly, in the phrasings a real buyer would use, and writing down what comes back. That is sampling rather than measurement, and it is the honest ceiling of what is available. Anything presented as a count of citations is an estimate in a more confident outfit.
Our sector gets almost no searches in Swedish. Does any of this apply to us?
More than it applies to high-volume sectors. Where domestic volume is small, each arriving enquiry matters individually, and the inbound English one — the firm setting up operations here — is disproportionately likely to have begun in a chat window, having no other starting point.
Should the English pages describe the regional structure, or would that read as a limitation?
Describe it, framed as capacity rather than as boundaries. Saying you operate nationally from four depots, and naming them, reads as substantial. Listing the counties you do not really serve reads as an apology. Claiming uniform coverage you cannot deliver reads as credible right up until the first callout.
Can we set a target for this in next year's budget?
Set targets on recorded figures — enquiries, positions on named terms, entries into the top ten — and carry the generative score alongside as context. A target on an inferred number produces work aimed at the number rather than at the enquiry, and it will move on its own for reasons nobody can audit.
How long before any of this changes what a model says about us?
Assume the general timescale of search work rather than something faster: first measurable movement after roughly four to eight weeks, and longer here, since the summarising layer draws on material that has to be found, kept and then relied upon.
Write the facts that survive being compressed
The layer above the rankings is real, it is where an enquirer who has never worked in this country begins, and what it does worst is preserve the regional distinction that decides whether a supplier can take the job. No tool resolves that combination.
What can be done is narrower and more useful than a visibility programme. Find out how your company reads from outside. Fix the pages whose description is vague because the page is vague. Publish the delivery facts as facts — towns, times, counts, days — so that whatever compresses your site has something specific to carry, and so the enquirer comparing three names has a reason to pick yours. Then measure what is measurable and let the score be context rather than a promise. Automation sits behind this at 149 USD a month per domain, the steered tier at 500, with over 230,000 partner sites for placements. Earlier pieces on regional structure are collected in the English blog, and the first concrete step is to read the market context for your own domain and see whether you recognise the company in it: open the panel and run the first market read.
