Pre-register
Freeze the core questions, markets, surfaces, entities, truth sources, treatments and exclusions before the baseline.
Version 1.0 · 30 August 2026
Evidence before
applause.
Humenta's AEO measurement protocol separates broad directional monitoring from a smaller frozen core designed to make the first 90 days interpretable. Every observation records its denominator, timestamp, surface, collection method, repetitions, sources, missingness and confidence so a readiness score or sampled answer is never misrepresented as universal visibility or causation.
Measurement principlesTransparent enough
01
02
03
04
Transparent enough
to challenge.
A Humenta AEO result is an observation with a denominator, timestamp, surface, collection method, source set and preserved evidence. It states successful and failed samples, repetitions, location or account context where known, material unknowns and confidence. It is not a claim about every user, every location, every provider experience or every future answer.
Repeat
Run multiple observations because model output is stochastic. Preserve differences rather than choosing the best answer.
Label
Keep API output, consumer-interface observation, Google Search Console data and analytics in separate evidence classes.
Disclose
Report missingness, access changes, provider changes, reviewer decisions, uncertainty and method revisions.
The 90-day protocolA controlled core inside
A controlled core inside
the operating program.
Scope the decision
Define the commercial question, organization and product entities, target audience, markets, languages, comparison set, approved surfaces and data boundaries.
Build the truth pack
Approve at least ten material claims and their primary sources. Record names, descriptions, offers, proof, exclusions, risky ambiguities and facts the answer must not invent.
Freeze 12 core questions
Select buyer-relevant informational, evaluative and recommendation questions. Randomly assign eight treatment questions and four holdout questions before execution begins.
Capture the baseline
Run three repetitions per core question and approved surface under documented account, locale and collection conditions. Store full answers, citations, timestamps, errors and cost.
Run the broad panel
Monitor the plan's 150 to 1,000 prompts for directional visibility, competitor, source and content signals. Do not treat every movement in this changing panel as a causal result.
Execute treatments
Apply technical, entity, content and legitimate authority changes only to the treatment set's mapped source system. Record owner, approval, publication evidence and effective date.
Re-run and compare
Repeat the frozen core under matched conditions at endline. Compare treatment and holdout movement, repetitions, answer quality, cited sources and missingness.
Publish the readout
Separate observed facts, verified facts, inferences and vendor statements. Report delivery, directional signals, controlled-core findings, limitations and the next decision.
Why 12 questions?
The core is intentionally small enough to inspect answer-by-answer and stable enough to repeat. It does not represent all demand. The larger plan panel supplies breadth; the 12-question panel supplies discipline.
Observation recordPreserve the evidence
Preserve the evidence
behind every chart.
| Field | What is recorded | Why it matters |
|---|---|---|
| Question identity | Frozen text, intent, treatment/holdout assignment, market and language | Prevents silent prompt drift and post-hoc selection |
| Surface identity | Provider, product surface, model when exposed, account state and locale | Consumer products and APIs can produce different outputs |
| Collection | Permitted method, timestamp, repetition, session conditions and errors | Makes the observation reproducible enough to audit |
| Answer evidence | Full response, citations or links, source domains and captured artifact | Prevents a score from replacing the underlying answer |
| Evaluation | Brand presence, position, relevance, sentiment, claims, factual status and reviewer | Separates automated extraction from human judgment |
| Missingness | Blocked run, unavailable surface, no answer, no citation, parse failure or exclusion | Keeps the denominator honest |
| Cost | Usage, operator and specialist effort where available | Connects evidence quality to an economically viable program |
Metric definitionsA numerator needs
Presence rate
Citation rate
Share of voice
Accuracy rate
Source readiness
Delivery
A numerator needs
a denominator.
Brand mentioned
Valid observations that mention the in-scope entity divided by all valid eligible observations. Missing runs are reported separately.
Owned source cited
Valid observations citing an approved owned or controlled source divided by valid observations that exposed citations.
Eligible presence share
In-scope entity mentions divided by eligible mentions across the frozen comparison set, with the exact inclusion rule stated.
Verified material claims
Material claims consistent with the truth pack divided by all verifiable material claims about the entity. Opinions and unverifiable claims are separate.
Audited source rubric
A disclosed account-level rubric across technical access, entity consistency, evidence quality, answer coverage and external corroboration. It is not a provider ranking.
Approved work shipped
Completed and verified work divided by approved in-scope work due in the period, with client/access blockers reported separately.
Traffic, pipeline and revenue may be reported when the client has approved analytics and attribution. Correlation is not labeled causation unless the design supports that conclusion.
Surface and method boundariesName what
Name what
was actually observed.
“ChatGPT visibility” is incomplete unless the product surface, access mode, locale, session state, date and method are known.
Provider APIAPI observation
Consumer productUI observation
Google Search ConsoleSearch performance
GA4 or server logsReferral/crawler signal
Third-party vendorVendor method
Humenta uses only collection methods approved for the program. A provider API response is never presented as a consumer-interface observation. If access or terms prevent a surface from being collected responsibly, it is marked unavailable rather than silently substituted.
Known limitationsWhat this method
Stochasticity
Product change
Small core
Attribution
What this method
cannot promise.
Answers vary
Repetitions reduce cherry-picking but cannot remove randomness or guarantee the next user sees the same answer.
Surfaces move
Providers change models, interfaces, retrieval and citations. Method changes are versioned and breaks are disclosed.
Not all demand
Twelve questions support close inspection, not universal market representation. The broad panel remains directional.
Many forces act
Holdouts improve interpretation but cannot isolate every market, competitor, provider or search change in a live environment.
No outcome guarantee
Humenta does not guarantee a ranking, citation, answer, traffic level, lead volume or revenue result. Delivery commitments and remedies belong in the signed order form.
Primary guidanceSearch fundamentals
Search fundamentals
still apply.
The protocol follows the evidence, not folklore about special AI files or guaranteed markup.