DOCUMENTED
Directly supported by the platform provider’s own documentation.
RESEARCH · METHODOLOGY V1.0
HOW WE MEASURE
AI VISIBILITY.
We measure AI visibility as a chain — from technical eligibility through discovery, citation and recommendation to a business outcome. We freeze the prompts before we look at the answers, record every source and competitor, separate documented platform rules from our own experiments, and keep the losing results so that progress can be measured instead of asserted.
This page is the method. It is published in full because a measurement you cannot inspect is a claim, not a measurement. The last section shows what it returned when we pointed it at ourselves.
THE CHAIN
Seven things that are not the same thing
Almost every confusion in this field comes from treating two of these as one. A site can be perfectly crawlable and never be cited. A brand can be cited and never recommended. A recommendation can appear once and be nothing but model variability. We measure the steps separately because they fail separately and they are fixed separately.
- 01 Crawlability
- 02 Indexing
- 03 Retrieval
- 04 Mention
- 05 Citation
- 06 Recommendation
- 07 Commercial outcome
THE LADDER
Seven levels · every observation gets exactly one
“Visible” is not a measurement. Each run against each prompt is scored at one level, and the levels for cited and recommended are deliberately kept apart: a citation means an engine used you as a source, a recommendation means it told someone to hire you. Collapsing the two is the most common way this gets measured wrong.
- L0
Not present
The brand does not appear in the response at all.
- L1
Discovered
Surfaced in a list or retrieval result, without meaningful descriptive treatment.
- L2
Mentioned
Named and described in the response.
- L3
Cited
Used as a source or shown as a linked citation.
- L4
Recommended
Actively recommended for the query, not merely listed.
- L5
Repeated
Reaches L4 repeatedly, across several engines or across repeated clean runs.
- L6
Commercial outcome
A visit, enquiry or client can reasonably be attributed to that visibility.
NO SINGLE SCORE
Why there is no “AI visibility score” here
A composite number is easy to sell and it hides the problem. Consider a site with strong crawler access, strong entity understanding, weak unbranded discovery, a citation rate of zero and a recommendation rate of zero. A score compresses that into something like 62 out of 100 — which tells you less than the five facts it replaced, and points at no action at all. We report the dimensions separately and let them disagree with each other.
FROZEN PROMPTS
Four rules · the reason a result means anything
A prompt is written down and frozen before anyone looks at the answer. After that it is an instrument, not a draft. Without these four rules a benchmark measures the person running it.
- Do not rewrite a prompt after a loss.
- Do not add our own name or language to a prompt to manufacture a mention.
- Do not remove a query because its result is unflattering.
- Do not replace a losing historical run. A better query gets a new ID; the old one keeps running.
WHEN A RESULT COUNTS
A single positive answer is a signal, not a victory
These systems are stochastic. Ask the same question twice and you can get two different answers with nothing having changed. So one good result is never reported as a win. It becomes meaningful when at least one of these also holds.
- It repeats in a fresh session.
- It appears on more than one engine.
- The same page is cited more than once.
- Independent third-party sources support it.
- Referral traffic confirms the discovery.
- An enquiry names AI search as how they found us.
HOW WE LABEL A CLAIM
Five classes · applied before advice reaches a client
Every strategic statement carries one of five labels internally. The point is narrow and practical: it stops industry folklore from being repeated to a client as if it were a platform rule.
OBSERVED
Seen in our testing, but not established as a platform rule.
INFERRED
A reasonable reading of the evidence. Not a mechanism.
UNVERIFIED
Insufficient evidence. Says so, and stays unused.
EXPERIMENTAL
A deliberate test of something unverified, registered before it runs.
WHY YOU LOST
Six diagnoses · six different answers
When a competitor is named and you are not, the useful question is which kind of gap it was. These six have different remedies, and picking the wrong one is how budget gets spent on the wrong work.
A Eligibility
The site cannot be reliably crawled, indexed or retrieved.
B Relevance
No page or entity directly matches what the query asks for.
C Evidence
The service is claimed, but there is no public proof, method, data or example behind it.
D Authority
Competitors carry stronger independent mentions, reviews or ecosystem presence.
E Freshness
Competing material is newer or more actively maintained.
F Variability
No stable evidence explains the difference yet. Often the honest answer.
Two corollaries we apply strictly: an authority gap is not solved with keywords, and an evidence gap is not solved with more schema.
WHAT THIS DOES NOT CLAIM
The list that makes the rest of the page worth reading
Most of what is sold as AI search optimisation rests on one of the following. None of them is part of this method, and none of them is something we will tell you we can do.
- That any action guarantees a citation. No platform offers that, and OpenAI, Google, Microsoft, Perplexity and Anthropic each state so in their own documentation.
- That any action guarantees a recommendation.
- Knowledge of how any ranking or retrieval algorithm actually works.
- A universal list of “GEO ranking factors”. We have not found one that survives testing.
- That schema, an llms.txt file or crawler access secures inclusion. They are eligibility, and eligibility is not selection.
- Any control over training data. It is outside this method’s scope, it is not measurable from outside, and the available tactics are manipulation.
OUR OWN RESULT
Baseline 27 August 2026 · sealed 4 September 2026
5 Engines
12 Frozen prompts
50 Observations
50 At L0 — not present
We ran this method against our own site first. Across 50 controlled observations on 5 engines, Black Oar Studio did not appear once — including on prompts that describe what we actually sell. That is the honest starting position, and it is published here rather than kept in a drawer, because the “before” is what any later movement has to be measured against.
Two things we will not round off. One engine contributed only 2 of 12 runs, because we count a run only when the web search visibly executed — a toggle being switched on is not evidence that a search happened. And a sixth engine was run in full but is excluded from the count entirely, because its source panel was never captured, so we cannot show that it searched at all.
The full dataset — every prompt, every engine, every competitor named, and the source domains behind them — is not published yet. Two of the six engine records are incomplete, and we would rather publish late than publish a reconstruction. It follows as a separate piece.
WHAT CAN MOVE A RESULT
Read this before trusting any single run — ours included
No manual benchmark reproduces every user's experience of an AI system. The same prompt can return a different answer for reasons that have nothing to do with the site being measured.
- Models are updated without notice.
- The search indexes behind them change.
- Location and language change the answer.
- Personalisation and session context change the answer.
- The same engine answers differently in different modes.
- Generation is stochastic. The same prompt is not the same experiment twice.
Which is why this method reports trends and repeated observations, and never presents one run as a permanent ranking.
