# AI Search Baseline Starter
By Rob Knott | https://robertknott.com
Version 1.0 | 6 September 2026

A practical starting worksheet for a small, documented AI-search pilot. Copy the templates into your preferred document or spreadsheet. This is a free sample, not a full audit or a promise of representative market coverage.

## 1. Define the decision before the prompts

Complete this brief before collecting answers:

- Business decision this research will inform:
- Brand, products and official names:
- Audience and real buying questions:
- Market and language:
- Platforms and search modes:
- Competitors and why they are relevant:
- Person responsible for interpreting the research:
- People who can act on findings:
- Evidence we already have:
- Access or evidence we are missing:

## 2. Build a small pilot sample

Begin with an exploratory pilot across discovery, comparison and validation. There is no universal correct number of prompts. Sample design depends on your questions and the variation you need to understand. Record the selection rule so a later comparison cannot quietly change the test.

Example patterns, to adapt to actual customer needs:

- Discovery: What options help [audience] solve [problem] in [market]?
- Comparison: How do [option A] and [option B] compare for [specific constraint]?
- Validation: What should I check before choosing [brand/product] for [use case]?

These are prompt patterns, not observed customer queries or measured results. Separate branded from unbranded tests. A prompt that names your brand cannot establish spontaneous discovery.

### Prompt register template

- Prompt ID:
- Exact prompt:
- Journey stage:
- Branded or unbranded:
- Market and language:
- Reason for inclusion and supporting customer evidence:
- Stable comparison sample or exploratory test:
- Planned platforms, run dates and repetitions:

## 3. Preserve the evidence

Use fresh conversations for independent tests. If you deliberately test a conversation, preserve its full context. Record the account state, location and search mode where visible. Do not record a hidden model version as if it were known.

### Evidence log template (one record per answer)

- Run ID and prompt ID:
- Date, time and timezone:
- Platform, displayed model/version and search mode:
- Account state, language and location where known:
- Exact input, including conversation context:
- Full answer or durable capture reference:
- Brand mentioned: yes / no / unclear:
- Brand recommended: yes / no / unclear:
- Owned source cited: yes / no / unclear:
- Third-party sources discussing the brand:
- All relevant citation URLs:
- Competitors included:
- Representation: accurate / inaccurate / mixed / insufficient evidence:
- Claims requiring verification:
- Reviewer and review date:
- Collection error or missing answer:

Keep mention, recommendation and citation separate. Define recommendation as an answer presenting the brand as an option for the user's task, not merely mentioning it. Verify factual claims against reliable source material. Use 'unclear' when the evidence does not support a judgement.

## 4. Report denominators and uncertainty

For any percentage, show the numerator, eligible observed answers, exclusions and question being measured. An unsuccessful collection is not an absent brand. If 3 of 10 observed answers mention a brand, that is 30% of those answers, not 30% of all AI search. Repetitions of a prompt are related observations; do not treat them as independent customer demand.

Check individual answers before explaining an aggregate. Preserve variation between platforms, journeys and repetitions. Citation frequency is an observation, not proof that the cited source caused a recommendation.

## 5. Turn an observation into an action

### Findings and action template

- Finding ID and question:
- Observed behaviour:
- Evidence references:
- Size and boundaries of the affected sample:
- Interpretation:
- Alternative explanations:
- Additional evidence needed:
- Proposed action and expected mechanism:
- Owner and dependencies:
- Priority and reason:
- Acceptance criteria for the implementation:
- Repeat-measurement plan:

### Worked example (fictional)

An answer to a comparison question cites a competitor's implementation guide and omits the fictional brand. Save the exact answer and citation. Check whether this pattern recurs and whether the brand has accessible, accurate implementation information. If a gap is confirmed, a content owner can improve the relevant page using product evidence. Validate the published page, then repeat the unchanged prompt sample. Do not claim that one missing page caused absence or that writing a guide guarantees inclusion.

## 6. Plan the next measurement

Freeze the comparison prompts and document changes separately. Record the intervention date, implementation checks, collection window and known platform changes. Explain what changed in the sample and what remains uncertain. A before-and-after difference alone does not prove that your intervention caused it.

## Use and limitations

You may adapt these worksheets for your own internal work and client delivery. This permission does not cover reselling or redistributing the starter as a standalone product. Keep client information in an appropriate private workspace. No confidential client examples are included here.

For current platform guidance, consult:
- https://developers.google.com/search/docs/appearance/ai-features
- https://help.openai.com/en/articles/12627856-publishers-and-developers-faq

## Want Rob to design the baseline with your team?

Scope, process and deliverables:
https://robertknott.com/services/ai-search-visibility-audit

Contact:
https://robertknott.com/contact?service=AI%20Search%20Baseline%20and%20Action%20Plan