Named on the Shortlist: AI Answers in a Crowded Field

Someone has been in Zurich for six weeks, needs a supplier for something specific, and has no colleague to ask. They type the question in English, read the paragraph that comes back, and three names go on a shortlist. That paragraph is now doing the job a personal network used to do — and it is doing it for the fastest-growing part of the city's buying population.

For a provider, this is a different competition from ranking. A list of ten links rewards the strongest domain; a generated paragraph rewards the clearest description. Those are not the same property, and firms that have spent years on the first frequently discover they have done nothing for the second.

The situation is sharpest where the field is crowded and the offerings are genuinely similar. Forty firms in the canton do roughly the same work to roughly the same standard. Nothing in the search results distinguishes them, so the answer layer distinguishes them instead — on the basis of what each has written down. The Semalt workspace measures that layer separately from position, and the two pictures rarely agree.

Four separate things, routinely treated as one

«Are we visible in AI search» is four questions wearing one coat. They have different answers, different causes and different remedies, and conflating them is why so many conversations about this topic go nowhere.

• Position Where does the page sit in the list of links? The old question, unchanged.
• Citation Is your text used in the answer, with your page credited as the source?
• Mention Does the company name appear in the answer, even with no link back?
• Characterisation What does the answer say you do — and is it what you actually do?

For a newcomer buyer the mention is worth more than the click. They are not looking for a page to read; they are looking for three names to compare. Being one of the three is the entire outcome, and whether they arrive via your site or via a directory afterwards barely matters.

What decides who makes a three-name shortlist

In a field of forty comparable providers, the selection is not made on quality — the model has no access to quality. It is made on which firms can be described concisely and confidently from public material. That is a written-description contest, and most of the field has never entered it.

The second pattern is more expensive precisely because it feels like success. The firm is being named; only somebody who reads the description notices that it is attracting the wrong enquiries and repelling the right ones. Reading the description is therefore the first measurement, before any count of mentions.

Write for the question, not for the machine

The material a model can use is the material that survives being lifted out of context. A sentence that needs the surrounding page to make sense is unusable. A sentence that states a fact, bounded and checkable, is usable. This is the whole of the technique, and it is why the advice sounds suspiciously like ordinary good writing.

There is a ten-minute test for any paragraph on your site: substitute a competitor's name. If the paragraph remains true, it cannot distinguish you and will not be used. Running that test across a homepage and three service pages typically produces a longer list than anyone expects, and the list is the brief.

Agreement across sources beats emphasis on one

A claim made once, on your own website, is a claim. The same claim appearing on your site, in an association profile and in a conference listing is a fact as far as any summarising system is concerned. This is the mechanism that surprises firms most, because it reverses a decade of advice about directories being irrelevant.

The practical route is a list: every place your firm is described that you do not control directly. For a typical Zurich provider that list has between eight and twenty entries, most of which can be corrected with an email. The work is administrative rather than creative, and it moves the needle faster than a new page, because it removes contradiction rather than adding volume.

For firms operating under more than one legal entity — common where a consultancy and its licensed arm share a name — the corroboration problem compounds. Systems merge the entities or attribute the work to the wrong one, and no ranking report will ever show it. The check has to be run per domain and the results compared side by side, which is why multi-entity groups tend to end up on the managed tier, where projects and attributions are set individually.

An eight-week routine that produces evidence

The temptation is to change everything at once and then be unable to attribute any change. A narrow, dated sequence is more useful, even though it moves fewer things.

• Week one Write down ten questions a real buyer would ask, in English, and record today's answers verbatim.
• Week two Read the characterisation of your own firm and list every factual error, with its likely source.
• Weeks three to five Write one page for each of the two most valuable questions, answering them completely.
• Weeks six to eight Ask the same ten questions again and compare. Record the differences, not your impressions of them.

One detail about the two pages written in weeks three to five is worth stating, because it decides whether the exercise works. They should answer the question completely, including the parts that are inconvenient — what the work costs to structure, how long it takes, and when a different provider would be the better choice. Pages that answer completely are the ones that get summarised; pages that answer partially and invite a call are the ones that get skipped, because the summary would leave the reader where they started. The same applies to the market view in the generated analysis: it can only reflect what has actually been written down somewhere.

The discipline of recording answers verbatim matters more than it sounds. Answers vary by phrasing, by day and by system; memory smooths that variation into a story. A dated record does not, and after two rounds the difference between noise and movement becomes visible without any statistical work.

Two habits keep the record straight afterwards, and neither takes long. The first is to treat the outside-source list as part of onboarding and offboarding: when a service is added or dropped, the list is walked. The second is to date the list itself, so that anyone picking it up later knows how stale it is. Both are administrative, and both are the kind of thing that quietly stops happening unless somebody owns them by name.

The market circle, and why the result stings

The analysis places providers into a leading, middle or niche tier. It is not a judgement of the work; it is a statement about how present and how clearly described a firm is in the material available publicly. For established firms with excellent client lists, landing in the niche tier is a common and unwelcome result.

It is also the most useful line in the report, because it converts «we are well known» into «known by whom, and written down where». The first is unfalsifiable and ends the conversation. The second is a work item. Firms that survive that first reading tend to make more progress in a quarter than in the preceding two years, precisely because the discussion finally has a subject.

Where the two layers meet, and where they do not

Position and citation correlate loosely. A page at the top of the list is cited more often than one at the bottom — but only if it answers the question. Pages from the middle of the list get cited routinely because they explain something cleanly, while the top result is a sales page that explains nothing.

Treating the two layers as one project is where budgets get wasted. They are watched together and worked on separately: ranking through campaign automation, characterisation through the editorial and corroboration work described above. Which of the two is the current bottleneck is usually obvious within a month of watching both.

How much precision this deserves

Not much, and pursuing more is a reliable way to waste a quarter. The systems change faster than any measurement series; a figure that meant one thing in spring can mean something else by autumn because the answer length or the source selection changed. Building a team target on such a figure means measuring the platform rather than your own work.

The durable measurement sits on your side of the line: how many of your services have a page that states audience, scope and boundary. That count rises only through your effort, never falls on its own, and holds its meaning across any platform change. Mentions follow it, with the usual lag and the usual scatter, and that is as much certainty as this subject offers.

What to do first

The first move costs an afternoon: write the ten questions, record the answers, read your own characterisation and mark the errors. At the end of it you will know whether you have a recognition problem, a description problem or no problem — and those three lead to entirely different plans, only one of which involves writing anything new.

If you want to see how your domain is currently described, connect it in the Semalt workspace and read the generated market context before looking at any number. In a crowded field there is nearly always one statement in it about your firm that you would not let stand — and that sentence is the most productive starting point this subject has to offer. Further pieces are in the blog, with the underlying editorial work covered under content and page revision.