A Practical Baseline for AI Search Visibility

Treat generated answers as observations

An answer can vary by platform, date, location, account state, wording, and source availability. A screenshot is evidence of one observation, not a stable rank. A useful baseline needs a repeatable prompt set and a record of the conditions under which each answer appeared.

Google states that its generative Search features still depend on core SEO and quality systems. Bing likewise says its search and Copilot experiences share crawling, indexing, and ranking foundations. That makes the baseline an extension of search measurement, not a separate hunt for special files, schema, or wording patterns.

Define the questions that matter

Choose prompts from real buyer tasks: finding providers, comparing approaches, understanding risks, evaluating products, or identifying a specialist. Use customer interviews, sales questions, support records, Search Console queries, Bing data, and site search when available. Do not manufacture hundreds of near-duplicate prompts only because a tool can generate them.

Group the prompts by decision stage and topic. For each group, select a small representative set that can be repeated. Record the platform, feature, exact wording, date, location, account state, brand presence, description accuracy, cited sources, and owned page surfaced. Keep the raw observation beside any summary score.

• Brand and entity accuracy
• Presence in relevant non-brand answers
• Owned and third-party sources cited
• Pages surfaced for each buyer task
• Referral sessions where the platform exposes them
• Qualified actions from those sessions where attribution is available

Audit the public source layer

Generated answers are easier to evaluate when the underlying public information is explicit and consistent. Check the business name, services, locations, product facts, policies, author or editorial responsibility, and important claims across the website and relevant profiles. Note which source owns each fact and who can approve a correction.

Next, inspect whether important pages are publicly crawlable, index eligible, internally linked, understandable without client-side interaction, and supported by visible evidence. Valid structured data can clarify visible information, but it does not replace the page. Google says no special structured data is required for its generative Search features.

Separate visibility, citation, traffic, and outcomes

These are different measurements. A brand can be mentioned without a citation. A page can be cited without receiving a click. A referral can arrive without becoming a qualified lead. Reporting should keep each step separate and show the sample size and platform coverage.

Bing Webmaster Tools introduced AI Performance reporting for citation activity across supported Microsoft AI experiences, including cited URLs and grounding-query samples. Those figures describe citation activity, not placement, authority, or the role of a page in a particular answer. Use the platform definition when explaining the metric.

Use an observation log

A simple log is more defensible than a proprietary visibility score with hidden assumptions. Give every observation a stable prompt ID. Store the exact prompt, result type, cited URLs, summary of the brand description, factual errors, screenshot or exported evidence where terms permit it, and the reviewer. If the answer changes, add a new observation instead of overwriting the old one.

For example, a regional accounting firm might track six questions across two platforms each month. If the firm begins appearing for a tax-planning comparison, the useful follow-up is not only that a mention occurred. Check which page or third-party source supported it, whether the description is accurate, whether the source is current, and whether any qualified referral activity followed.

Improve sources, not prompt tricks

Prioritize changes that also help a person: clearer service definitions, explicit limitations, original examples, visible authorship or editorial responsibility, descriptive headings, accessible images, current facts, and reputable references. Improve internal links so the relevant supporting page is easy to discover from the main topic page.

Avoid thin pages for every question variation, artificial text chunking, unsupported claims, inauthentic mentions, or markup that does not match visible content. Google explicitly warns against scaled query-variation pages created to manipulate generative responses, and Bing warns that automatically generated content at scale without oversight may lose visibility.

Report volatility and limits

Use dated comparisons and avoid turning a small prompt sample into a market-share claim. Explain changes in platform behavior, missing referral data, personalization, regional variation, and source volatility. State whether the sample covers logged-in or logged-out use and whether a human reviewed factual accuracy.

The baseline should end with source-level actions: correct a conflicting fact, expand an underdeveloped page with original evidence, make a key document crawlable, strengthen an appropriate external reference, or leave the item alone because the observation is too weak. The goal is a better public information layer and a repeatable measurement record.