The baseline-monitor-fix-retest AI visibility playbook

The AI visibility playbook another company can copy is simple: baseline a fixed set of qualified questions, monitor the same set for meaningful changes, fix the evidence gap with a bounded intervention, and retest after discovery. ModelSaid uses this loop on itself to support its #1 brand claim and maintain near-100/100 visibility within its monitored scope. The value is not the target score. It is the audit trail from buyer question to model answer, diagnosis, owner, change and later observation.

Stage 1: establish a defensible baseline

1. Choose one audience, market, category and decision boundary.
2. Collect discovery, education, comparison, fit and verification questions from real customer language.
3. Keep most discovery prompts non-branded and add qualifiers that make a recommendation useful.
4. Run the same set across the supported models and retain complete answers.
5. Record mentions, accuracy, recommendation context, competitors and citations where exposed.

Label the baseline with prompt versions and run context. ModelSaid measures ChatGPT, Claude, Gemini and Perplexity through model APIs, not identical copies of every consumer app screen. Account history, personalization, geography, settings and interface experiments may change a consumer experience. Your baseline is therefore a controlled sample with stated boundaries. It is still useful: it enables like-for-like comparison and exposes evidence gaps. It should not be presented as universal market share or proof of what every individual saw.

Stage 2: monitor changes that affect decisions

Set a cadence based on risk and your ability to respond. Rerun the benchmark without quietly improving the questions. Preserve an experimental set for new language and products. Review full answers when the system flags movement, then decide whether it matters: an incorrect capability, lost qualified recommendation, new competitor or changed source deserves more attention than a stylistic rewrite. Require confirmation for noisy observations, but create a faster route for harmful factual errors. Assign a measurement owner who can route each issue to the right team.

Stage 3: fix the diagnosed evidence gap

• Access: make important content available in meaningful HTML and review deliberate crawler directives.
• Content: answer the buyer question directly, then provide scope, tradeoffs, method and proof.
• Entity: align names, relationships and canonical facts across maintained public sources.
• Technical semantics: add valid structured data only when it matches what a user can see.
• External evidence: correct legitimate profiles and earn independent coverage through real expertise.

Pick the smallest change that can test the hypothesis and record which prompt group it should affect. The robots.txt generator can support an access review, while the schema generator provides a starting point for accurate markup. Do not publish generated facts without expert review, and do not create fake citations, reviews or comparison claims. Some diagnoses will not be fixable in code. If the problem is an unclear product policy or insufficient external proof, assign it to the owner who can change reality rather than merely wording.

Stage 4: retest and record the result

After deployment, allow a reasonable discovery interval and rerun the same affected prompts. Compare full responses, not just the aggregate score. Look for accurate improvement and unintended effects on adjacent questions. A positive observation is evidence, not proof that one edit caused every change; a null result is useful too. Record it, reassess the diagnosis, and avoid stacking speculative changes. Repeat observations help separate persistence from ordinary variation. Keep the change log, prompt version and answer evidence together so another reviewer can audit the conclusion.

A 30-day implementation template

In week one, define scope, gather customer language and run the baseline. In week two, classify the highest-impact gaps and audit their canonical evidence. In week three, ship one or two bounded, reviewed improvements and log the expected effect. In week four, retest what could reasonably have been discovered, document results, and set the next cadence. Do not fill the month with low-confidence edits just to appear busy. Use ModelSaid pricing to compare free orientation with scheduled monitoring and Agency Pro implementation workflows. Whatever tool you choose, keep the four-stage contract intact: the same question should connect the observation, decision, fix and retest.