Method

How the scan is run, and why it is run that way.

Every scan runs the same four steps.

The pipeline

One scan: the name goes in, the finding comes out

A scan is one pass with two lanes. Four models are asked about the company by name, while the site and the public record are read with no model involved. The six checks are decided where the lanes meet.

The question below is all we say about you. Four calls, the same question, and one standing instruction that never names your company.

Four public signals, read directly. No model is involved in this lane.

Input

Sent to the model

Company name
Marlbeck Fluidics Ltd
Country
United Kingdom given here, optional

The prompt, in full

What can you tell me about Marlbeck Fluidics Ltd (United Kingdom)?

That is the whole question. Behind it sits one standing instruction, the same on every scan: answer from your own knowledge, and say plainly when the name means nothing to you.

Not sent

  • The website address. The model is never shown marlbeck-fluidics.example.
  • Web access. No search and no retrieval of any kind.
  • Tools. Nothing to call mid answer.
  • Any earlier turn. Each call starts from an empty conversation.

Hand a model the address and it reads the page, then answers from the page. What comes back is a summary of your own site, and the gap we came to measure has closed.

Model calls

Four bare calls to two providers, all issued at once.

Model AProvider onecold, no tools
Model BProvider onecold, no tools
Model CProvider twocold, no tools
Model DProvider twocold, no tools

One provider is one opinion, so we use two.

Signals

Read in the same pass as the model calls.

  • Company websiteFetched at scan time. Title, description, and any structured data the page publishes.no model
  • Public entity recordSearched by name in the open reference bases models lean on, then read for what the entry actually describes.no model
  • Domain registrationThe date the domain was first registered, read from the registry.no model
  • Use of the nameA search for the company name next to the word scam, read for whether other people are using the name to approach yours.no model

Nothing here is generated. It is fetched, and the four answers are held against it.

The six checks

clear review issue

01Recognition, result: clearcomputedmodel answers
02Identity accuracy, result: issuejudgedmodel answers + site
03Cross-system consistency, result: issuejudgedmodel answers, side by side
04Entity anchoring, result: issuecomputedentity record + your site
05Machine-readable identity, result: reviewcomputedyour site
06Name contention, result: issuecomputedentity record + model answers

Computed means the result is worked out in code, by counting the answers and reading what the site and the public record say. Judged means one further model made the comparison, with the four answers and the signals in front of it and nothing else. Checks 04 and 06 are computed but put one question to that model: does the record filed under this name describe this company.

Output

Verdict

Two of the four models describe a different business under this name, and the public entry carrying the name describes that same other business.

Four issues, one to review, one clear.

02 Identity accuracyIssue

Two models describe a supplier of dental laboratory equipment. Marlbeck Fluidics builds flow control valves for water treatment plants.

“They supply dental laboratory equipment and consumables to clinics across the north of England.”

Model B, cold call, site withheld

04 Entity anchoringIssue

One public entry carries this name and it describes the dental supplier, not the valve maker. It is still the nearest reference the models have, and it is where two of them got their answer.

“Marlbeck Fluidics. Dental laboratory supplier, north of England.”

Public entity record, matched on name

05 Machine-readable identityReview

The site publishes structured data, but none of it declares an organisation. Nothing on the page states in a form a machine can read which company this is.

“Marlbeck Fluidics Ltd, flow control for municipal water treatment.”

Company website, meta description

Marlbeck Fluidics, every quotation and every result above are invented for this example.

Step 01

We ask without helping

Each model gets the same question: your name, and your country if you gave one. It never sees your website address and it cannot search.

Why

Nobody forming an opinion about you looks up your site first. The answer we want is the one they get.

Step 02

We fix a reference point

Your own site is treated as correct. We read its title, its description, any structured data it publishes, and the date the domain was registered.

Why

Two models disagreeing is only noise until there is something true to measure them against.

Step 03

We find the record they lean on

We read the public entity records filed under your name and check what each one actually describes. A record that matches the name but describes another business is the most serious result the scan returns.

Why

Models lean on a small number of structured, citable sources. When the nearest one is about somebody else, that is where the wrong answer comes from.

Step 04

We compare, and we quote

Every answer is set against the others and against your site. Each finding in the report carries the sentence or record it came from.

Why

An unsourced claim about a company is worth nothing. We will not make one.

What the scan will not do

It does not rank you against competitors, and it does not judge whether anything said about you is deserved. The report gives what the models answered and what public records exist, then stops. Where a model recalled nothing, it says so.

Run a scan