Cited, Not Ranked: What Can and Cannot Be Measured in AI Answers
Ask an assistant to name the organisations working on a given file and it will name three or four. There is no results page behind that answer, no position to report and no official tally of how often you were among the names — and the missing tally, rather than the technology, is what an institution has to plan around.
Something has been inserted ahead of the search box. A colleague opens an assistant, describes the problem in a sentence, and reads a paragraph naming a handful of organisations. No ten results, no snippets to compare, no way to see who was left out.
For a body whose work depends on being taken seriously by people who have never met it, that changes the stakes of a layer nobody controls. It also arrives with an unusual measurement problem, and handling that honestly is most of what follows. The generative-search views in the Semalt panel produce useful readings of this layer; what they do not produce, and cannot, is a count.
Your organisation now gets described by something you did not write
Two moments are worth separating. In the first, nobody knows you yet: a researcher at an advocacy organisation is drafting background and needs to know who else works on the subject. The assistant returns a short list of names with a sentence each. Whether you appear is decided inside a system with no public rules, and if you do not, no signal of any kind reaches you.
In the second, your name is already known. Somebody types it in, exactly as they might have typed it into a search engine, but the reply is different in kind: instead of your homepage, they get a three-sentence summary of what your organisation does, assembled from whatever the model absorbed about you. That summary is now the first thing said about you, and you did not write a word of it.
Ask two or three assistants what your organisation does and compare the replies with the description in your own annual report. Four kinds of divergence turn up repeatedly.
A remit you have left behind
The answer describes the file you worked on five years ago, when most of the text about you was written.
- Old text still outweighs new
- Check what still ranks for your name
Merged with a similar name
Two bodies with adjacent names and overlapping files get folded into one description fitting neither.
- Common with acronyms
- Use the full legal name somewhere prominent
A position attributed to you
A stance appears that you never took, usually inherited from a coalition statement you co-signed.
- Check joint statements you co-signed
- State your own position plainly somewhere
Nothing at all
The assistant declines to describe you at all. For a young or small secretariat this is the default, not a failure.
- A corpus problem
- Registers and directories come first
Being named and being summarised are separate problems
Classic search gave you one thing to think about: where you appeared for an expression. The generative layer gives you two, and they respond to different work.
| Situation | What the reader is doing | What you can observe | What actually shifts it |
|---|---|---|---|
| Ranked for an expression | Choosing among options on a page | Position, impressions, clicks | The page, the links, the competing set |
| Named in a list answer | Assembling a field of candidates | Nothing directly reported | Being described, by others, in retrievable text |
| Summarised on request | Checking a name they already hold | Only what you ask an assistant yourself | Clear, current, self-describing pages of your own |
| Quoted as a source | Reading an answer with attributions | Referral visits, sometimes labelled | Extractable claims with dates and method |
The fourth row is where the two worlds touch: when an assistant retrieves live documents, the material it reaches is broadly what a search engine can reach, so indexing and internal linking remain the foundation rather than a superseded discipline. The middle two rows depend instead on how your organisation is written about across the web, and neither responds to keyword work in any recognisable sense.
There is no counter, and it matters that you know it
This is not a temporary gap that better instrumentation will close next year. Assistants generate text; they do not serve documents from an index whose deliveries can be logged. Even where an answer carries links, the count of answers carrying yours sits with the provider. What can be observed is real but indirect, and the honest list is short.
- Referral traffic from assistant domains. Where an answer links out and somebody clicks, the visit arrives with a referrer like any other. A floor, not a total: most answers are read without a click.
- Your own sampled questions. Ask the same twenty questions monthly and record the answers verbatim. Small samples move a lot, so write the sample size next to the finding.
- The retrievable corpus about you. What exists in text, where, and how current it is. Fully inspectable, and the input the models had.
- Modelled position scores. An estimate of how a domain stands in a subject space, derived from what a model reports rather than from any log.
People who file consultation responses for a living know the difference between a measured quantity and a derived indicator, and they notice when a supplier blurs it. Stating the limit up front is not a hedge here; it is what makes the remaining numbers usable.
What the generative-research area shows, and on what basis
Six views make up this part of the panel, in two groups: where you stand relative to others in a subject space, and which of your pages are the levers. Both are model-generated readings, and both are worth more as a series than as a single number.
Where a model places you among your peers
A competitiveness score, a circular map of the field, and a written description of your domain as the model perceives it.
- A score for the generative layer. The AI competitiveness score condenses a domain's standing into one figure whose value lies in its direction over months, not its absolute height.
- The market circle. The field drawn as rings: dominant names at the centre, a middle band around them, specialists at the edge. Which ring you sit in is often the most informative thing on the screen.
- A written market context. Positioning, an estimate of reach and a list of openings, expressed in prose. Read it as the model's account of you, not as data.
- An overall visibility figure. One number for standing across the generative landscape, subject to the caveat above and useful chiefly as a tracked line.
Where the work would pay
Question research with intent classification, pages flagged for expansion or linking, and an account of what peers cover that you do not.
- Questions rather than expressions. People type sentences at assistants. The query research and intent classification views work at that granularity, closer to how the layer is used.
- Intent, sorted. A question asked to understand a procedure and one asked to find a supplier want different documents. Sorting by intent stops a page trying to serve both.
- Leverage pages. Documents the model marks as worth extending or worth linking to internally — usually pages that already almost answer something.
- Gaps against the field. Subjects your peers cover in retrievable text and you do not. In an institutional space that list is usually short and specific.
None of the six is a log. They are structured readings of what models report, and their honest use is comparison: this quarter against last, your domain against a named peer. Used that way they inform. Used as evidence of reach, they are not evidence at all.
Which numbers may appear in a document somebody signs
Organisations in and around the European institutions file things: tender responses, grant reports, transparency declarations, submissions to a supervisory body. A figure that enters any of those acquires a status it did not have on a dashboard.
The readings are not worthless. Internal instruments and filed evidence are simply different categories, and mixing them is a governance failure rather than a marketing exaggeration.
| Figure | Fine internally | In a filed document | Defensible wording |
|---|---|---|---|
| Modelled visibility score | Yes, as a tracked line | No — no primary source exists | "An internal indicator we track quarterly" |
| Market circle placement | Yes, for prioritising work | No — a model's classification | Not worth citing externally at all |
| Referral visits from assistants | Yes | Usable, with the caveat stated | "Referred sessions recorded in our analytics" |
| Your own sampled questions | Yes | Usable if the method is described | "In a monthly sample of 20 questions…" |
| Search Console clicks and impressions | Yes | Usable — a primary record | "Google Search Console, period, property" |
The fourth row is the interesting one. A self-collected sample with its size and method written down is weaker than a log and stronger than an estimate, and it is defensible provided you never round it up into a claim it cannot carry. Twenty questions asked monthly, recorded verbatim and described as a small study, holds.
An institutional publisher starts this race ahead
The characteristics that make a document usable as a source are, almost without exception, the characteristics of good institutional writing.
A claim with a date on it
A statement carrying when it was true can be repeated without risk. One that does not is a liability to anything repeating it.
- Date the finding, not the page
- Say what has changed since
A named author with a role
A passage attributed to a person with a stated function is safer to quote than the same words published anonymously.
- Name and function, not a byline
- A page for the person that exists
A stated method or sample
Two sentences on how you know something turn an assertion into a finding, and findings travel further than assertions.
- Sample size and period
- What the method cannot show
Adjectives about yourself
Leading, trusted, innovative: these describe nothing, verify nothing, and protect nothing that reproduces them.
- Replace with a countable fact
- Or delete without replacement
One structural habit belongs on the list too. Answers get assembled from passages rather than from pages, so a document whose central point is spread across nine paragraphs is harder to draw on than one that states it in a single self-contained passage and then argues it. Put the answer near the top, in a form that survives being lifted out of its surroundings.
Two further things cost nothing: a page saying who the organisation is in plain declarative sentences, on a URL that has not moved in years; and agreement between that page and the registers and membership listings others maintain. Where those disagree, a model has no way of resolving the disagreement, and resolves it anyway.
Which of your three versions gets used as the source
An organisation in Brussels commonly publishes in French, Dutch and English, and in the generative layer the arithmetic of that differs from classic search. An assistant answers in the language of the question but does not restrict itself to sources in that language: a question asked in Spanish about a European procedure can be answered from English documents and rendered into Spanish on the way out.
The failure mode follows directly. Where the English pages are a thinner rendering of the French originals — a summary, an abstract with a download link — the substantial text sits in a language most of the continental audience will not search in, and the thin version is what is available to be drawn on. The remedy is not translation volume.
- Pick the subjects, not the pages. Three or four subjects where you genuinely want to be the reference get full English documents; everything else can stay a summary without harm.
- Do not let a summary carry a claim. If the finding, the method and the date live only in the French original, the English page is not a source. It is a signpost to one.
- Keep the self-description identical across versions. Who you are should not vary by language. Differences here resurface later as contradictory answers about what you do.
- Check the answers in more than one language. Ask the same question in French and in English and compare. Divergent answers point straight at which version holds the substance.
Frequently asked questions
Can we find out how often an assistant cited us?
No. No provider publishes that figure and no third party has access to it. You can count referral visits from assistant domains, which is a floor rather than a total, and you can run your own repeated sample of questions. Anything presented as a citation count has been estimated, and should say so.
Does classic search work still matter if answers are replacing results?
It matters more, not less. When an assistant retrieves live material, it reaches broadly what a crawler can reach. A document that is not indexed cannot be retrieved either, so indexing and internal linking remain the foundation the generative layer sits on.
An assistant describes our remit incorrectly. What can we do?
Work on the inputs rather than the output. Make sure a clear, current, plainly written description exists on a stable page of your own, and correct the registers and membership listings that describe you differently. Expect this to take months and to be partial. There is no correction mechanism to appeal to.
Should this figure go in our annual report?
Not as a headline number. A modelled score has no primary source and will not survive a reader who asks how it was produced. If the subject belongs in the report, describe the work and cite figures that do have a source: referred sessions, Search Console clicks and impressions, or your own sample with its method stated.
How quickly does any of this respond to changes we make?
Slowly and unevenly. Retrieved answers can reflect a new document within weeks, in line with the four to eight weeks it typically takes to see any first movement in search. Answers drawn from what a model absorbed during training may not reflect your changes until that model is replaced.
Working usefully with a layer you cannot count
That sounds like a reason to ignore the layer entirely. It is not. A quantity that cannot be measured can still be influenced, and the work of influencing it is well matched to what this readership already does: publish dated findings, attribute them to named people, state methods and their limits, and keep the description of the organisation current wherever it appears.
A workable routine fits in an hour a month. Ask your fixed set of questions and record the answers verbatim. Read the market circle and note whether the ring has changed. Turn one gap from the list into a document that stands on its own in English. Then leave it until next month. Our notes on content and technical work set out how we run it, and the other articles here cover the measurement side. The market context and content-gap reports are the parts that reward attention; the score is the part to hold lightly.
If you want a first reading for your own domain, connect the property and let the generative-research views populate before drawing conclusions: open the panel and run the first market analysis. Then write down, before you look, what you expect the answer to your own name to be. The gap between that sentence and the one that comes back is the most useful finding this layer offers, and it costs nothing.
Besoin d'Aide Avec Votre SEO ?
Notre équipe d'experts est prête à vous aider à améliorer votre classement dans les moteurs de recherche et à développer votre présence en ligne.
Consultation Gratuite