Turn ModelSaid's near-100 AI visibility playbook into agency proof

An agency can use the ModelSaid near-100/100 dogfooding playbook as proof of process, not as a promised client outcome. Select an agency-owned site, define a commercially relevant prompt benchmark, publish the rules, capture complete answers, fix one evidence gap at a time and retest the affected cohort. Report misses and null results alongside improvement. The proof is that the agency follows a controlled method on its own brand and can expose the work for scrutiny; the score remains bounded by the chosen prompts, models, modes, market and dates.

State exactly what near 100 means

ModelSaid describes its own performance as near 100/100 within a monitored benchmark. Treat that as a first-party, product-authored claim rather than an independent award or universal ranking. An agency adapting the approach should publish its prompt families, qualification rules, supported assistants, search modes, geography, run window and scoring method. Separate branded recall from non-branded discovery and report factual accuracy independently. Do not imply that the benchmark represents every possible buyer question. A narrow, disclosed score can be useful because another reviewer can understand and challenge it.

• Choose real buyer questions from sales, support, search and product research.
• Define eligible answers and material facts before observing any model output.
• Preserve complete responses, citations, model, mode, date and prompt version.
• Maintain a stable reported cohort and label exploratory prompts separately.
• Publish limitations, unfavorable examples, methodology changes and unresolved gaps.

Run baseline, diagnose, fix and retest

Start with the AI visibility scan, then group failures into access, clarity, entity, evidence, accuracy or eligibility problems. Inspect site foundations with the AI readiness checker. For each diagnosed gap, choose the smallest truthful intervention and predict which prompt family it could affect. Release changes separately where practical, log the date and wait for a defensible retest window. If the answer does not improve, record the null result and revisit the hypothesis. This prevents the agency from crediting a broad redesign for unrelated model movement.

Convert dogfooding into sales evidence responsibly

Create a public methodology page, dated change log and compact case narrative showing the starting observation, approved intervention and later result. Use the agency site because the team controls its facts and can disclose the process without exposing a client. Invite a colleague who did not run the work to audit a sample. Show prospects the evidence chain and let them inspect limitations. Do not extrapolate the agency score into promised visibility, traffic or revenue. Use current ModelSaid pricing to explain which recurring monitoring and Agency capabilities support the process without hardcoding a subscription amount.

1. Freeze a version-one benchmark and publish the definition of every reported metric.
2. Run the baseline across ChatGPT, Claude, Gemini and Perplexity in the defined modes.
3. Prioritize material inaccuracies and qualified discovery gaps over easy score gains.
4. Ship bounded, truthful changes and preserve validation plus deployment evidence.
5. Retest fairly, publish the result and version the benchmark when genuine market needs change.

Set a written rule for benchmark evolution. Add a prompt when customer research reveals a durable buying question, not because the agency performs well on it. Retire a prompt when the offer or market genuinely changes, and preserve the previous version for historical interpretation. Recalculate comparisons only where the cohort remains equivalent. This prevents score maintenance from becoming score manipulation. A public version note also teaches prospects that responsible monitoring adapts to reality without rewriting the past, which is a stronger proof of expertise than a permanently perfect chart. Keep an archived benchmark export so future reviewers can reproduce the published scope. Date every exported version and name its approver.

Build proof around discipline, not perfection

A credible proof process will contain volatility, ambiguous cases and interventions that do not move the benchmark. That is a strength when disclosed. Validate any structured-data work with the schema validator, but never present valid markup as a guaranteed visibility mechanism. Keep screenshots secondary to complete captures and make the measurement window visible. The agency can confidently say it uses the same monitoring, diagnosis and review loop on itself that it offers to clients. It should not say that near 100 is independently certified or transferable. Dogfooding becomes persuasive when it demonstrates intellectual honesty, operational repeatability and a willingness to be measured by rules set before the answer appeared.