Measurement contract at a glance
| Prompt bank | 150 named buyer questions with stable IDs, exact wording, intent group and owner-approved priority |
|---|---|
| Surfaces | ChatGPT, Perplexity, Claude, Gemini, Google AI Overviews and Microsoft Copilot |
| Scheduled volume | 150 prompts × 6 engines = 900 prompt-engine checks per weekly collection |
| Cadence | Weekly collection; monthly trend report |
| Primary unit | One named prompt on one named engine at one timestamp |
| Missing-data rule | Report unavailable separately; exclude it from eligible-rate denominators |
| Change control | Version prompt wording, engine coverage and classification rules |
| Commercial context | Starter is $2,900 per month, month-to-month; verify the current pricing and scope |
1. Build and freeze the prompt bank
The company and strategist agree on 150 questions that represent how buyers discover, compare and evaluate the category. Each prompt receives a permanent ID, exact wording, intent label, priority and relevant product or audience. The bank covers branded, category, comparison, problem, feature and recommendation intents.
Prompt wording stays fixed inside a reporting period. Adds, removals and edits create a new bank version, because changing the question and calling the answer a trend would confound the measurement.
2. Run the engine panel on a fixed cadence
Each weekly collection schedules every named prompt on each of the six listed surfaces. Where the surface permits it, the check uses a clean session without prior conversational history. Locale, account state, model or product label, web-grounding availability and any exceptional settings are recorded when observable.
A scheduled check is not automatically a completed answer. Rate limits, product availability, safety responses, missing web access and technical failures can make a run ineligible. The coverage metric exposes that gap.
3. Capture the answer evidence
For each eligible answer, the record stores the prompt ID, engine, timestamp, answer text or auditable reference, cited URLs in displayed order, whether Quoted First or the client brand appeared, whether it was recommended, named competitors and materially inaccurate descriptions. Citations and brand mentions are separate fields: an answer can mention a company without citing its site, or cite a page without recommending the company.
4. Calculate the metrics
| Metric | Formula | Interpretation |
|---|---|---|
| Coverage rate | eligible completed checks ÷ scheduled checks | How much of the planned panel produced usable evidence |
| Mention rate | eligible answers naming the brand ÷ eligible completed checks | How often the brand appeared at all |
| Citation rate | eligible answers citing the target domain ÷ eligible completed checks | How often the site's pages appeared as sources |
| First-citation rate | answers with the target as first displayed citation ÷ eligible answers with citations | How often the target led the displayed source list |
| Recommendation rate | eligible comparison or recommendation answers positively selecting the brand ÷ eligible answers in those intents | Selection frequency on buyer-decision questions |
| Material-error rate | eligible brand-bearing answers with a documented material error ÷ eligible brand-bearing answers | How often the brand was described inaccurately |
Every rate carries its numerator, denominator and unavailable count. A percentage without those values can make a partial collection look complete.
5. Report trends without pretending one run is a ranking
AI answers vary. One favorable output is a dated observation, not a permanent rank. Weekly rows show movement; the monthly report separates repeated changes from one-run noise, identifies which prompts and sources moved, and maps recurring gaps to technical fixes, entity work, briefs or rewrites.
The panel is a controlled synthetic measurement of named prompts. It is not a log of every real user question, a substitute for referral analytics or proof that a particular change caused a result.
What the open files establish—and what they do not
The machine-readable specification makes the promised measurement contract inspectable before purchase. Quoted First also publishes a separate native-search visibility experiment with dated negative and positive controls. That experiment demonstrates the evidence policy; it is not a customer case study or proof that the Starter service has produced outcomes for clients.
Quoted First currently has no customer results represented on this page. The methodology defines how Starter is intended to measure future work. No agency controls whether an independent AI system crawls, indexes, retrieves, cites, ranks or recommends a page.
Frequently asked questions
Does 900 scheduled checks mean 900 successful answers?
No. Coverage is reported separately. Unavailable or failed checks are not silently converted into negative answers.
Why keep the prompt wording stable?
Because changing the input and comparing the outputs would mix prompt edits with actual movement in the answer systems.
Does this methodology guarantee an AI citation?
No. It measures independent systems; it cannot guarantee crawling, indexing, retrieval, citation, ranking or recommendation.
Who publishes this specification?
Quoted First, which also sells the Starter service. It is a first-party commercial specification, not an independent endorsement.
Inspect the exact service scope.
Compare the public price, included work, exclusions and cancellation terms before deciding.