Start with a decision, then write the buyer questions
Choose one decision the research should inform: whether buyers can discover a capability, whether a product claim is understood, or which facts need clearer evidence. Name the target market, language and buyer role. A broad goal such as “appear more in AI” does not define a measurable observation.
Write questions in the buyer’s language, keeping discovery, comparison and verification questions separate. Record a stable question ID and the exact wording. A question that names the brand is a different test from an unbranded supplier question. Do not change the wording midway through a comparison.
| Question type | Illustrative question | What to examine |
|---|---|---|
| Discovery | What should a procurement team check when choosing a specialist supplier? | Relevant categories and any spontaneous brand mentions |
| Comparison | How can a buyer compare the evidence behind two service proposals? | Selection criteria, sources and stated uncertainty |
| Verification | What evidence supports the capability described on this page? | Whether the answer identifies the claim and its supporting material |
Keep the observation surfaces separate
An API model response, a search-enabled API response, a consumer product interface and a search AI summary are different observation surfaces. Record the actual surface for every observation. An API result is not evidence that the same answer appeared in a consumer interface or a search result.
The current hosted implementation supports API-model observation code; a real run still needs an authorized model configuration, a confirmed plan and evidence acceptance. Other surfaces must be scoped and observed separately. This guide does not report a production run.
- API model: preserve the provider, endpoint, model identifier, request settings and original response.
- Search-enabled API: additionally preserve the actual tool configuration, retrieved sources and search evidence.
- Consumer interface: identify the actual product, observation time and permitted account context; retain the original visible evidence.
- Search AI summary: identify the actual search experience, query, market and date. Record absence and failures distinctly.
Freeze a small protocol before execution
For each planned observation, fix the question version, language, market, surface, model or product version where available, repetition number and relevant tool settings. Define a spending limit and a stop condition before calling a paid provider. Record any field the platform does not expose as unknown.
Use a run ID to connect the planned observation with its attempt and evidence. Keep provider failures, timeouts and incomplete responses as failed or missing observations. A configuration test only checks that a connection can respond; it is not a completed brand audit.
| Protocol field | Why it matters |
|---|---|
| Question ID and exact text | Makes the question recoverable instead of relying on a remembered prompt |
| Surface and model/product identifier | Prevents pooling different experiences into one visibility figure |
| Market, language and observation time | Keeps the intended buyer context explicit |
| Repetitions, limits and completion status | Shows how much of the planned work was actually completed |
Preserve the answer before interpreting it
Keep the original answer, available citations, timestamp, model metadata and usage or cost evidence. Preserve an immutable version or content fingerprint so that a reviewer can identify the exact material behind a conclusion. Redact credentials before storage or display.
A brand mention, a link and a verified supporting source are separate findings. A citation in a model response still needs review: does the linked page exist, does it support the statement, and was its content available at the relevant time? Do not label a claim verified merely because a model printed a URL.
Report coverage and outcome rates with their denominators
Method example only — synthetic counts, not a client result: plan 24 observations, obtain 18 valid answers, and find the brand in 6 of those answers. Coverage is 18/24 = 75%. The valid-answer mention rate is 6/18 = 33.3%. The 6 missing observations are not 6 negative answers.
Show the run totals alongside question-level evidence. If one surface has only a few valid answers, disclose that limit rather than averaging it into a confident cross-platform score. Preserve the evaluation rule for what counts as a mention or citation.
| Synthetic example metric | Calculation | Interpretation |
|---|---|---|
| Observation coverage | 18 valid / 24 planned = 75% | How much of the agreed plan was observed |
| Valid-answer mention rate | 6 mentions / 18 valid = 33.3% | Mention frequency within the observed valid answers |
| Missing observations | 24 planned − 18 valid = 6 | Unobserved outcomes, not proven non-mentions |
Connect discoverability checks to the correct claim
For supporting links in its AI search features, Google requires a page to be indexed and eligible for a search snippet; meeting those conditions does not guarantee inclusion. Treat a successful page-access test as an access result, not a citation result.
OpenAI separates its search crawler settings from training crawler settings. Review the intended search access without assuming that making a page searchable requires granting every training permission. Keep the operating workspace and customer material protected.
Sources for this section: Google Search Central — AI features and your website · OpenAI — Overview of OpenAI Crawlers
What a reviewable baseline handoff contains
The useful next step is a scoped correction and a comparable remeasurement. Publishing more pages, receiving a model answer or saving a report is not by itself proof of improved visibility.
- The frozen protocol and planned observation list, including exclusions.
- Original answers and citation records for valid observations, plus failures and missing evidence.
- Definitions and explicit denominators for each reported measure.
- A prioritized improvement list linked to the observed evidence, not guaranteed rankings or orders.
- A reviewer’s decision tied to the evidence version, with unresolved questions kept visible.