Someone at your company has already typed a question into a chatbot, watched it recommend three competitors, and pasted the screenshot into Slack. The obvious follow-up — how often does that happen, and is it improving — has no honest answer with a number on it. It has a useful answer without one.
Generative answers now sit on top of the ordinary result page, and for some questions they replace it. A developer evaluating tooling, a finance lead scoping a vendor, a candidate checking what a company builds: all three increasingly get a synthesised paragraph before ten blue links. If that paragraph names three products and yours is not one, the click never happens and nothing in your analytics records the loss.
The screenshot that arrives on a Tuesday
The alarm is proportionate. What follows it usually is not. Within a day someone has run twenty more prompts by hand, someone else has found a vendor promising to track "AI share of voice", and a slide appears carrying a percentage. Nobody asks where it came from, which in a company whose engineers interrogate every latency chart is an odd lapse.
So it is worth being blunt before describing any tooling, including the tooling this article is about.
That is not a reason to ignore the layer. It is a reason to be explicit about which grade of evidence a figure belongs to — a habit this readership drops only when the subject turns to marketing.
Three grades of evidence, and the mistake is mixing them
Every number about your presence in machine-generated answers falls into one of a few categories, and they are not interchangeable. Sorting them once, in writing, prevents most of the arguments that follow.
| Grade | Example | How it is produced | What it can support |
|---|---|---|---|
| Observed | Clicks, impressions, position for a query | Logged by the engine and reported back to you | Thresholds, alerts, commitments |
| Sampled | Your own fixed prompt panel, run weekly | You ask, you record, you state the method | Trend within your own method |
| Inferred | A visibility or competitiveness score | A model reasoning over public signals | Prioritisation, ordering of work |
| Anecdotal | The screenshot in Slack | One prompt, one session, personalised | A reason to look, nothing more |
The line between the second and third rows is the one that gets lost. A sampled figure is weak but transparent: you defined the prompts, you know the sample size, you can hand the procedure to a sceptic. An inferred figure looks stronger and explains itself less, because the derivation runs through a model whose weights you cannot inspect.
An engineer will recognise the shape of that immediately. A synthetic probe firing every minute from one region tells you the probe passed; real user telemetry tells you what people experienced. Reporting the first as though it were the second is how incident reviews go wrong, and reporting an inferred score as an observed count is the same error in different clothes.
The company a model describes is the company your documentation describes
Generative systems assemble a picture of an organisation from whatever text about it exists in volume. In a Stockholm company that runs internally in English, the largest surface area by far is not the marketing site. It is the documentation — every configuration key, every method, every release note — plus a changelog, a set of job adverts and, if the company is old enough, years of incident write-ups.
That inverts the usual assumption behind positioning work. The carefully argued sentence on the front page is one sentence. The documentation says something adjacent to it four thousand times, in the vocabulary engineers chose, and repetition is what a model weighs.
Documentation
The largest body of text about your product, written for people who already bought it.
- Sets the working vocabulary
- Rarely says what you compete with
Careers pages
Job adverts describe the stack, the market and the stage of the company in plain sentences.
- Often your clearest positioning
- Usually on a vendor subdomain
Status page
A permanent archive of every incident, phrased in the language of failure.
- Signals that the product is real
- Also your public downtime record
Swedish pages
A handful of translated pages, maintained by nobody, linked from a switcher.
- Little for a model to draw on
- Swedish answers go elsewhere
The last card is the specific Nordic case. A Swedish-language question about your category is answered from whatever Swedish text exists, which for an English-first company is almost none of yours. You are not described badly in Swedish; you are not described at all.
Six views, and the derivation behind each one
The AI section is six views rather than one score, and the separation matters for the reason set out above: each has its own derivation and therefore its own weight. Reading them as a blended judgement throws away the only thing that makes them usable.
Six views over the answer layer
For teams deciding where to spend effort, not for teams reporting a result upward.
- A competitiveness score with a market circle. Competitors are grouped into top tier, mid tier and niche rather than ranked in one list.
- Model-generated market context for a domain. Positioning, an estimate of traffic and a set of stated opportunities, written as prose about your site.
- Query research with intent classification. Candidate questions grouped by what the asker appears to want, which is the part worth arguing with.
- Pages flagged as levers. Specific URLs marked as worth expanding or worth linking to internally, rather than a site-wide grade.
- Competitor strengths and content gaps. Where another domain covers ground you do not, expressed as topics rather than keywords.
- A global visibility figure for the portfolio. One number across every connected domain — the most inferred of the six, and the least suited to a report.
Four of the six produce lists you can act on; two produce scores. The lists carry the value. A flagged page with a stated reason can be checked against your own knowledge of the site in a minute. A score of sixty-one can be checked against nothing.
Why the market circle contains nobody you pitch against
The competitor set surfaced here rarely matches the one in the strategy deck, and the mismatch is informative rather than a defect. A Stockholm studio expecting two other studios finds an app-store aggregator, a general reference article about the genre and a jobs board. None of those take its revenue. All of them occupy ground a model walks across on its way to an answer.
Coverage, not size
A domain lands here by covering the category comprehensively.
- Reference and comparison sites
- Rarely anyone selling a product
Where your rivals sit
Companies with something to sell write about themselves, not the category.
- Narrow slice of the space
- Closest to your own profile
Not a consolation prize
Deep coverage of one problem is what a narrow question pulls in.
- Cheapest tier to enter
- Often the highest intent
Missing entirely
A rival you lose deals to but who appears nowhere in the circle.
- Wins on price or relationships
- Text cannot capture either
Read the circle as a map of the conversation, not a league table. You cannot be cited in a discussion you never joined, and the tiers show which parts of it are crowded.
A portfolio figure is only as complete as the portfolio
The global visibility figure runs across every connected domain. Most technical companies own more domains than they remember and connect fewer than they own, so the figure describes a subset chosen by accident.
The pattern is predictable. Marketing is connected, because marketing thinks about it. The documentation subdomain runs on another stack and was verified by an engineer with whatever Google account was open. Careers lives on the applicant-tracking vendor's domain and was never treated as a web property. The status page belongs to a third vendor.
| Property | Usually verified by | Does a model read it | Belongs in the portfolio figure |
|---|---|---|---|
| Marketing site | Marketing, company account | Yes, and it is the shortest text | Yes |
| Documentation | An engineer, personal account | Heavily — the bulk of your prose | Yes, and this is the omission that matters |
| Careers subdomain | Nobody, vendor-hosted | Yes, job adverts describe the company | Yes, if you control the hostname |
| Status page | Nobody, vendor-hosted | Sometimes, for reliability questions | Optional, and tag it separately |
| Old pre-rename domain | Whoever ran the rebrand | Yes, if it still serves content | No — fix the redirect instead |
Two panel features address this. Several Google account groups can be linked and read together, so documentation need not be re-verified merely to appear beside marketing. Individual sites are shared with named e-mail addresses, so a contractor gets one domain rather than the group. Site tags filter every view at once, scoping a portfolio figure to the properties that sell something.
The work that has a defensible mechanism behind it
Nobody can promise a citation. What can be stated is which actions have a plausible causal path into a generated answer. The list is shorter than the industry would like, and every item is worth doing whatever the answer layer does next.
- Be indexed at all. Nothing can be drawn on if it was never crawled. The least interesting item on the list, and the one that disqualifies the most pages.
- Say the same thing everywhere. One canonical description repeated across documentation, careers and marketing gives a model one story instead of three.
- Answer the question, not the keyword. Intent classification in the query research view earns its place because generated answers respond to questions rather than search strings.
- Link internally to the pages carrying the argument. Pages flagged as levers are usually substantive and orphaned, which is a fixable combination.
- Exist on other people's sites. A model reading about your category meets your name where other domains mention it, which is what placement work is for.
That last point connects to the campaign side. Automated placement draws on a partner network of more than 230,000 sites, and the two tiers differ in how much of the selection a person controls; the tier comparison covers that, and it is the subject of its own article. The mechanism is the point here: mentions on third-party domains are text a model can read, and text is the entire substrate of this layer.
The assistant bound to your own project data
For questions that fall between two views and would otherwise become a spreadsheet.
- One chronological feed per project. Answers, generated reports, newly placed links with donor rating and traffic, to-dos and campaign news in a single column.
- A router decides what data to load. Between zero and three data blocks are pulled per question by relevance, so a content question does not drag campaign figures into the reply.
- Answers stream token by token. Up to twenty messages of history are retained, enough to refine a question rather than restate it.
- Lists go in as batches. Keyword and URL lists are accepted wholesale, and to-dos carry an active, deferred or dismissed state.
An assistant belongs in an article about honesty with numbers because this one is bound to project data rather than general knowledge. Ask it something the loaded blocks cannot support and the answer should be thin — correct behaviour, worth testing early so everyone knows where the boundary sits.
How to write this down for people who will check
Here is what makes the subject usable inside a technical company. Since no observed count exists, define your own sampled measure, state its method completely, and report it as what it is. Designing it takes an afternoon; running it takes twenty minutes a week.
Fix a set of prompts — thirty to fifty questions a real buyer might ask, written once and frozen. Run them on a stated day, against a named model, from a session with no history. Record for each prompt whether your domain was named and whether a competitor was. The output is a fraction with a documented denominator, which is exactly what the slide in the first paragraph was missing.
Reported that way, the monthly update reads: in our fixed panel of forty prompts, run the same day each week against the same model, we were named in nine answers, up from six a month ago. Every part of that is checkable, including the part where forty prompts is a small sample.
None of this argues for skipping the layer. It argues for holding it to the standard the rest of your monitoring already meets — which your colleagues will apply anyway, the first time a figure turns up without a method behind it. Other notes on measurement sit in the English blog archive, and the reporting side of the same estate is covered in the tooling overview.
Frequently asked questions
Can we find out how many times a model has cited our site?
No. No engine exposes that count, and no product on the market reads it directly, because there is nothing to read. What exists is your own sampling, which you control and can document, plus inferred scores derived from public signals. Treat a vendor claiming a true citation count as a prompt to ask precisely how the number is produced.
Our documentation describes the product differently from our website. Does that matter here?
More than it does for ordinary ranking. A generated answer synthesises across everything it read, so two competing vocabularies produce a blurred description, or a choice you did not make. Pick the canonical term, apply it in the docs first because that is where the volume is, and let the marketing copy follow.
The competitor circle shows domains we have never heard of. Is the tool wrong?
Usually not. It maps who occupies the explanatory space around your category rather than who competes for your deals. Reference sites, aggregators and adjacent documentation legitimately hold that ground. Read it as a map of what a model is likely to have read.
We have almost no Swedish content. Should we build some for this reason alone?
Only if it has substance of its own. A Swedish page mirroring the English one word for word adds little and may not survive assessment. A Swedish page answering a question your Swedish-speaking buyers actually ask is different, and in a small language market the absence of a competitor doing the same is often the opportunity.
How long before work on this shows up anywhere?
First measurable movement in ordinary search typically appears after four to eight weeks. The answer layer lags that: your text has to be indexed, then read, then reflected in generated output. Set the review date at three months and resist the weekly noise in between.
The layer above classic ranking is real, it is growing, and it is poorly instrumented. Those three statements are compatible. A company that reports what it sampled, labels what it inferred and admits what it cannot count decides better than one presenting a percentage nobody can reproduce — and among people who read graphs for a living, it is also believed.
To see the generative views against your own domains, start by connecting the properties that carry your text, documentation included. You can open the panel and connect a first domain through a single Google consent flow, then read the generative research section against a portfolio you know is complete. Background on how the workspace is assembled is worth ten minutes beforehand, and the generative analytics section itself goes further into the individual views.
