Settle the entity
One clear name, one clear description, one address. Actively ruling out confusion with similarly named companies.
LLMO agency for large language model optimisation
Into the sources the models draw from.
LLMO works on the picture of you that exists across the web rather than on your website. Who writes about you, which details circulate, whether they agree. That is the material language models build an answer from.
Definition
What large language model optimisation means here.
Lead KPI
LLMO
Citation frequency per model
How often a given model names you, and which source it leans on when it does. Kept separate per model, because the numbers differ sharply.
- Target the models themselves
- Format long term
- Audience brand and comms
A language model knows your company from two directions. From what sat in the training data, and from what it retrieves at the moment of the question. You have little influence over the first and a lot over the second.
So LLMO is work on sources. Does your company exist as a clearly bounded entity, or does it get confused with a similarly named firm. Do address, founding year and service description agree everywhere they appear. Is there anything outside your own domain a model can lean on.
The result is unglamorous and lasts: a clean set of facts, confirmed in several places. The technical side of that runs under AI Brain.
Scope
What we work on.
Clear contradictions
Founding year, headcount, services: where two versions exist on the web, a model picks one. Usually not yours.
Build mentions
Trade media, associations, industry directories, speaking slots. Sources that exist independently of you and therefore weigh more.
Structured data
One connected graph instead of scattered fragments, with stable anchors that can be referenced.
Access
A clear statement per provider in robots.txt, plus an llms.txt. Both consistent with whatever the website says about AI use.
Measure citations
Recorded per model, with the source named. Only that breakdown shows where the work landed.
Boundaries
When LLMO is the wrong place to start.
LLMO is the slowest of the four disciplines. Mentions do not appear on request, and models pick up new sources with a delay. If you need to see something within a few weeks, start elsewhere.
If your content does not answer questions clearly, AEO moves faster. If you do not know where you stand at all, the audit is the better first step. LLMO pays off once the groundwork is in place and still nobody names you.
Common questions
About LLMO.
What does LLMO stand for?
Large language model optimisation. It covers the work on the material language models use to form claims about you: sources across the web, structured facts, and whether you are recognisable as a distinct entity.
How does our company get into a model's sources?
Through text that is not yours but talks about you: trade articles, association pages, press, directories, conference programmes. And through your own set of facts, which has to agree with those texts. Together they make a coherent picture.
What is an entity and why does it count?
An entity is a uniquely identifiable thing: this company, this person, this product. Systems work with those rather than with strings of characters. Anything not cleanly bounded gets mixed up with similarly named things, and then half the answer is wrong.
Do we need a Wikipedia or Wikidata entry?
Helpful yes, necessary no. Wikipedia has hard notability criteria many companies do not meet, and a bought entry gets found out. Wikidata is more open and often the more practical route for structured facts. We assess that honestly rather than promising it.
How much do mentions on other sites matter?
A great deal. A claim about you carries more weight when it does not come from you. That is why digital PR is not a side topic in this field. In Austrian search results for agencies of this kind, third-party publications rank visibly.
What if contradictory details about us are circulating?
Collect them first, decide which version is correct, then correct downward from the most important source. It is tedious detail work across directories, profiles and old press texts. It also works more reliably than almost anything else in this field.
How do you measure citation frequency?
With the same question set as the audit, but evaluated by model and by the source named. The distribution matters more than the total: if one model knows you through a single external source, your visibility hangs on that one page.
Can we influence models directly?
No. There is no channel for feeding content into a model. What can be influenced is the material a model finds, and whether it is allowed to fetch your pages. Anyone claiming otherwise is selling you something.
What about training data cut-offs?
They explain why some models repeat older details about you. Current changes land first where a system retrieves live. Anything tied to the model's own knowledge only shifts with a new version, and nobody outside has influence over that.
How is LLMO different from classic link building?
Link building counts the link, this counts the claim. A mention without a link can be valuable when it confirms your facts. Conversely, a link from a page nobody treats as a source does little here.
How long does the effect last?
Longer than with the other disciplines. A clean set of facts and mentions in trade media do not vanish with the next update. In exchange, building them takes months. That trade has to be made deliberately.
Difference
How others do this, and how we do.
Not a judgement on individual agencies, but a description of two approaches. The right column is checkable, because this website itself is built that way.
| Difference | Usual agency approach | agenticsearch.at |
|---|---|---|
| Starting point | A proposal after a first call | A measurement with a date, before anything gets recommended |
| Basis | Keywords from classic search | A question set of real prompts, clustered and prioritised |
| Proof of success | Rankings and a visibility index | Mentions per question and system, with the raw data |
| Reporting | A monthly dashboard with curves | The same measurement repeated, change called only after three runs |
| When unsure | Try it and see what happens | We say what is not evidenced, and write it on the page |
| Tooling | Agency licences, access ends with the contract | The question set and the data are yours, afterwards too |
| Scope | As many channels as possible from one supplier | A capped number of concurrent projects, otherwise we turn it down |
Your guides
You play the lead.
You know your business, your customers and your offer better than any agency does. What is usually missing is solid numbers on how AI systems talk about it, and an order in which to tackle that.
Who is behind this
two people
- Florian Hieß measurement, technology, structure
- Romana Hieß day to day support and implementation
- Operated by Digital Wings GmbH, founded 2010
- Client projects 100+
The number of concurrent projects is capped. When a piece of work needs more capacity than exists, we turn it down rather than take it on.
01
You know where you stand
It starts with a measurement, not a proposal. You see in black and white which AI answers name your company and which do not, with a date and a question set.
02
You know what comes next
Instead of a list of everything, you get an order. Every step has a result you can verify yourself, including without us.
03
You keep the decisions
Your offer and your positioning stay yours. We tell you what stands in the way of visibility and take on the technical part of it.
Next step
Let us find out where you stand in AI answers.
One conversation, 30 minutes, no sales pressure. You leave with an honest read on whether this is worth the effort for you.
