What has to be true before we publish it.
We ask buyers to hold AI-visibility vendors to a standard of evidence. This page is the standard we hold ourselves to, written plainly enough that you can check whether any given page on this site meets it.
Version 2026-08-28
Who writes this, and who checks it
Research is produced by the team that runs the engagements
Articles bylined Plastorium Research are written by the strategists and analysts who design the prompt panels, run the measurement, and read the results for clients. They are not commissioned from writers outside the practice, and they are not produced by a content agency.
Every published claim has a named internal reviewer
Before an article ships, someone other than its author checks it against the evidence tiers below and against the measurement methodology. Where a claim cannot be supported at the strength it is written, the claim is weakened or cut — the review is not advisory.
Named bylines are the next step, and they are not here yet
Publishing under an organisational byline is a genuine weakness in this library, and we would rather say so than dress it up. Our differentiator is expert interpretation, and a reader currently has no way to evaluate whose expertise that is. Named author and reviewer profiles — with the background that qualifies each person to make the claims on the page — are in progress. Until they exist, treat every article here as the practice's collective position, reviewed under this policy.
What a claim has to rest on
Claims on this site are written at the strength their evidence supports, and the page says which kind of evidence that is.
| Kind of evidence | What it is | How it may be written |
|---|---|---|
| First-party measurement | A figure we produced ourselves under the published measurement methodology, on a named panel, providers, market, and window. | Stated as a measurement, with its window and scope attached. |
| First-party exploratory analysis | Internal analysis whose study design we have not published — sample, target definition, validation, and uncertainty are not yet inspectable. | Labelled exploratory on the page itself, framed as a hypothesis to test, never as a finding. |
| Cited external source | A claim resting on a provider's own documentation, a named study, or a platform's published guidance. | Linked to the primary source, not to a secondary article summarising it. |
| Practitioner judgement | What we have seen across engagements, where we have no measurement that generalises. | Written in the first person plural as experience ("we have usually seen"), not as a rule about how models behave. |
| Market statistics we cannot source | Adoption percentages, market-size figures, and "X% of buyers now…" claims where no study, date, geography, and definition can be pointed at. | Not published. Removed on sight, including from older pages. |
Rules for causal claims
Most of what goes wrong in this field is a sentence that is one verb stronger than its evidence. These are the rules that verb has to pass.
- 1
Association is written as association
Where we observed two things moving together, we say so. "Visibility rose after the work" is a measurement; "the work raised visibility" is a causal claim, and a fixed-panel before/after in a live market does not support it.
- 2
A citation is evidence, not a mechanism
A source displayed alongside an answer tells us what was surfaced with it. It does not establish that the source caused the brand selection, and an uncited source may still have contributed.
- 3
Provider behaviour is described per provider and per mode
Retrieval, internal knowledge, ranking, synthesis, and displayed citations differ between products and between modes of the same product. We do not generalise one product's behaviour into a statement about "how AI works".
- 4
No guaranteed outcomes
We do not guarantee rankings, mentions, citations, traffic, leads, or revenue, because AI outputs are controlled by the providers and vary over time. What we commit to is the agreed implementation deliverables, the reporting cadence, and the remeasurement — things we control.
- 5
The strongest claim wins the caveat
If one page states a claim more carefully than another, the careful version is the site's position and the looser one is a defect to fix — not a difference of tone.
How client results are published
Anonymised by default
Client names and domains are withheld unless a client asks to be named. Case studies state the category, the market type, and the window instead.
Disclosed and withheld are both listed
Each case study names what it is showing you and what it is holding back — including the verbatim prompt panel and per-run answer text, which would identify the client. Clients receive all of it in their own reporting.
Quotations are approved, attributions are labelled
Testimonials are real quotations from real engagements, published with the client's consent and with names changed where anonymity was requested. Where a client attributes leads to AI-assisted discovery, that is their attribution and is labelled as client-reported.
Illustrative material is marked
Example reports, sample answers, and demo figures used to explain the product are labelled illustrative on the surface that shows them. They are never presented as a client result.
Where AI is used to produce this site
We sell AI-visibility work, so it would be strange to be coy about this.
- Research design, the measurement rules, the interpretation of results, and the decision to publish are human work.
- AI assistance is used in drafting and editing prose, in code, and in bulk data processing — the same way a spellchecker or a script is.
- No figure, quotation, case-study result, or citation on this site is generated by a model. Numbers come from measurement runs; quotations come from named clients who approved them.
- A model is never a source. If a claim rests only on what a language model said, it does not meet the evidence standard above and is not published.
When we get something wrong
- 1
Report it
Anything on this site that looks wrong can be sent to us through the contact page. Errors found in research are worth more to us than the embarrassment costs.
- 2
Fix the claim, everywhere
A wrong claim is corrected on the page that carries it and on every other surface repeating it — article, homepage copy, llms.txt entry, feed description. Half-corrected is still wrong.
- 3
Say what changed
A correction that changes a number, a conclusion, or the strength of a claim is noted on the page rather than silently overwritten, and the page's modification date is updated.
- 4
Fix the process, not just the page
A defect that could recur becomes an automated check in our build. The release guard that blocks unfinished drafts and duplicated page furniture from shipping exists because both once reached production.
Found something on this site that does not meet the standard above? Tell us — including if it is one of our own older pages.