Designing a brand-blind AI visibility baseline

Measure discovery without giving the answer away

If every test prompt contains your brand name, the exercise measures how a system describes an entity it has already been given. It does not show whether a buyer asking a category, comparison, evidence or fit question could discover that brand. A brand-blind baseline therefore begins with non-branded buyer questions and keeps brand-prompted accuracy tests in a separate register.

The two registers answer different questions. The discovery baseline asks whether the brand, its website or useful evidence appears when the buyer has not named it. The accuracy baseline asks whether facts are correct when the brand is explicitly named. Do not merge them into one visibility score: a brand can be accurately described but not discovered, or mentioned during discovery with inaccurate details.

Complete an eight-field baseline register

Freeze the method for a reporting window before collecting observations. The register below makes the conditions inspectable and gives a supplier a concrete acceptance format rather than a screenshot selection.

Build questions around real buying tasks

Use a small, balanced set that a buyer could plausibly ask before knowing a supplier. Useful families include category discovery, approach comparison, evidence and risk, organisational fit, and the next step. For example: what should a university include in an AI visibility audit; how should a UK organisation compare ongoing GEO services; or what evidence should a bilingual GEO provider supply. Treat these as templates and adapt them to the actual market and decision.

Avoid manufacturing dozens of near-identical prompts to inflate the denominator or hunting for wording that produces a preferred answer. Google's current guidance warns against creating large quantities of pages for query variations and says generative AI features can use query fan-out and varying models or techniques. A baseline should therefore document a useful sample under fixed conditions, not claim to replicate an internal ranking system.

• Category: Which providers or approaches could solve the defined need?
• Comparison: What should a buyer compare before selecting a service?
• Evidence: What proof and limitations should a credible provider publish?
• Fit: Which option suits the stated organisation, market, language or delivery constraint?
• Enquiry: What information should the buyer prepare before contacting a provider?

Report observations without inventing precision

State the denominator and preserve the raw evidence. A simple report can show how many questions produced a brand mention, an own-domain link or citation, an accurate material fact, or no usable observation. Keep those columns separate. Mark a system error, access limitation or unavailable response as unavailable, not as zero and not as a failed brand result.

Compare reporting windows only when the question set and relevant conditions remain stable. If a platform, interface, region, language, session method or scoring rule changes, record the change and start a new comparable series where necessary. NIST describes its AI Risk Management Framework as support for incorporating trustworthiness into the use and evaluation of AI systems; for procurement, the practical implication is to make context, method, evidence, review and limitations explicit.

Keep external observations separate from business outcomes

An observed mention or citation is not the same as a website visit, qualified enquiry or sale. Keep delivery evidence, AI observations, search data, website behaviour and commercial outcomes in separate fields with their own source and date. Google states that AI-feature visibility is not guaranteed even when requirements and best practices are met, and its Search Console reporting covers Google Search data rather than every external AI system.

Ask a provider to disclose the fixed question register, markets and languages, tested interfaces, session conditions, raw-evidence format, classification rules, review cadence, treatment of unavailable observations and every method change. Reject a proposal that replaces those details with a single unexplained score or guarantees inclusion, citations, rankings, traffic or enquiries.

How SALIENS can establish and use the baseline

SALIENS provides GEO audits, AI visibility monitoring, content and technical improvements, and ongoing GEO delivery for UK and international organisations, including scoped English and Simplified Chinese work. We can define a brand-blind question register, record the observation conditions and evidence, separate discovery from accuracy, and use documented gaps to prioritise controllable website work.

Email info@oxfordintelligence.co.uk with your website, target markets and languages, priority buyer decisions, known competitors or alternatives for context, and the systems you want observed. SALIENS is operated by Oxford Intelligence Limited. A baseline records observations under stated conditions; it does not guarantee crawling, indexing, mentions, citations, rankings, visits or enquiries.