Method
First we measure, then we fix. How the QuorumLab measurement runs, what the report covers, what we guarantee and where the limits are.
First we measure, then we fix
QuorumLab does not start with assumptions and does not start with edits. First we need to know what search engines and AI assistants already say about a company, where they take it from and how stable those answers are.
One answer proves nothing. Ask an assistant the same question about your company on different days and in different sessions, and the answers will differ: sometimes close in meaning, sometimes at odds on the facts, sometimes about another company altogether.
This is not a fault. The model builds the answer again on every request, and the wording of the question, the history of the conversation, the moment in time and the sources it managed to read all feed into it. A screenshot of one answer shows only what happened once.
So we work with a measurement, not with an answer.
What we do and why
The job of QuorumLab is to make the data a machine builds its answer from complete, consistent and backed by sources.
Wikidata, owned surfaces, Reddit and an article in the English Wikipedia all affect what AI assistants say, and they do it in different ways. We treat them without preference: they are instruments, not the goal. Which set a company needs is what the measurement shows.
What the system does with corrected data is up to the system. We answer for the input and for the fact that it agrees with itself.
What the client gets
Not an opinion about how visible the company is in AI, but a map: which answers are stable, where the machine gets it wrong, which sources the error comes from, what can be fixed inside the site, what needs external records, and what nobody can control directly.
The map carries a date. Three or six months later you can come back to it and see what changed.
What a baseline is
A measurement fixed to a date. Not a forecast and not an analyst opinion, but a state you can repeat and compare the next measurement against.
The date sits on the title page of the report and works as the point of reference. Without it no claim of improvement can be proved.
How the measurement runs
Each question runs at least five times over several days and in separate sessions. The second part matters: five requests in a row inside one chat are not five runs, they are one conversation with accumulated context.
The result is recorded as a share: five of five, four of five, three of five and so on.
| Result | Zone |
|---|---|
| 5 of 5 and 4 of 5 | green |
| 3 of 5 | yellow |
| 2 of 5 and below | red |
The thresholds are printed in the report itself, so you can check the conclusion instead of taking our word for it.
Three of five is neither bad nor good, it is a fact. Two of five means your buyer gets the right answer roughly forty per cent of the time and hears something else in the other sixty. The red zone tells you the error repeats often enough to rule out chance.
Questions are asked in the language your market checks you in. For Finland that is Finnish. English comes in when there are international buyers, partners or investors. Answers differ between languages because the model has a different pool of sources for each, and the error is often present in only one of them.
What we look at
We look at more than the answer itself. What matters is why the system answers the way it does.
The report covers:
- the company website and its structured data;
- Wikidata and external identifiers;
- the link between the brand and the legal entity;
- the link between the company and its products, executives and former names;
- public sources that assistants can draw on;
- Reddit and other open discussions, if they turn up in the answers;
- contradictions between your own sources and outside ones.
The question is not whether the AI got it wrong. The question is which layer the error came from.
Why the problem usually starts outside the model
A machine does not read names. It looks for a record of the company in its sources and answers from that.
If there is no record, the answer gets assembled from scraps: reviews, directories, stray mentions, a competitor page with a similar name. If a record exists but went stale after a rebrand, the earlier version of the company answers. If there are two records and they contradict each other, the machine picks one, and not necessarily the right one.
Sometimes the problem sits on the website. Sometimes in external records. Sometimes in the missing link between sources that already exist. Sometimes the company is not recognised as a separate stable record at all.
An error in the answer nearly always arrives two steps earlier, from the record or from its absence. So the thing to fix is not the answer but what the answer is built from.
Why the order of work matters
The order matters, and here is why.
Start with owned surfaces before the measurement and you may build something nobody needed. Edit the site before you understand the external sources and you may fix the wrong layer. Go into Wikidata without checking the sources and the work turns out impossible or unstable. Start with Reddit without knowing whether it appears in the answers and you end up doing ordinary content management.
The basic order: measure, examine the sources, fix the records, align the site, then owned surfaces, Reddit or a Wikipedia eligibility check if they are needed, then a second measurement.
Work done out of turn usually gets done again. Coverage published before the machine recognised the company piles up against a string of text rather than against you, and then has to be reconnected.
We do not decide in advance which instrument is needed. Wikidata, owned surfaces, Reddit and a Wikipedia article affect assistant answers differently and suit different companies. The measurement shows what to apply. Sometimes the answer is "none of the above, fix the markup on your site", and that is a result too.
Reddit comes in separately and not always: when public discussions are already visible, already cited by assistants, or important for your market.
What we fix and what we only measure
| Layer | What QuorumLab does | Limit |
|---|---|---|
| Wikidata, site markup, external identifiers | Corrects and aligns records | Project rules apply |
| The client website | Checks it and sets out what to change | The client team or its contractor makes the edits |
| Owned surfaces | Builds open sources under client control | The owner is named openly |
| Sets up the official account and your own community | Outside communities are not controlled | |
| Google Knowledge Graph and assistant answers | Measures change against the date of measurement | Nobody has direct access |
The last row limits all the ones above it, and it sits here on purpose.
We do not make an outside system say a particular sentence. We fix what can be fixed in the source layer: records, links, markup, sources and public context.
What we do not promise
- A Google Knowledge Panel.
- That ChatGPT, Claude, Gemini or Perplexity will say what you want.
- Pushing negative material out of search results.
- That outside systems are obliged to accept our data.
What we do guarantee is the work performed and a measurable change in the input data against the date of the first measurement.
The absence of guarantees is a position, not a line of small print. A supplier who promises a result inside a system it does not control is promising something outside its control.
About open projects
Wikidata and Wikipedia are edited by a community, and any edit of ours can be revisited by other contributors. Keeping 85 to 90 per cent of what was added after six months counts as a good result, though it is not a guarantee that the record is fully protected.
Three months of monitoring are part of the work on records. Monitoring covers our own edits. It does not cover defending the record against anything a third party does: no such service exists in an open project.
When the work is not possible. When there is nothing verifiable to support the facts about the company. When the work would be read by the community as self-promotion.
When the effect is limited. If a company works through direct referrals and is rarely checked through search or assistants, the effect will be limited. The measurement shows this straight away, and that result is useful too: it saves you unnecessary work.
If the company was registered recently and nothing has been written about it yet, there is nothing to verify. That work starts somewhere other than with us.
What happens after the report
You get the report in Word and PDF. It is yours, show it to whoever you like.
On Pro and Premium we go through it together: the red zones, the sources of the errors, the order of action. The point of the meeting is to explain the document, not to sell the next stage.
If work is needed, we set out the scope and the next step separately. Nothing obliges you to continue.
A second measurement runs on the same method and shows the change against the first date. You order it separately, when you need it.
The report is useful on its own. Some clients go on to fix the site, the markup or their internal data themselves. That is a normal outcome.
Where to start
Send us your company website and describe the task briefly. Our first reply contains an observation about your data, not a request for a call.