# How do you write an RFP for an AI visibility platform?

> Require repeated-run testing because the same brand list appeared in fewer than 1% of repeated ChatGPT or Google AI responses in SparkToro’s study.

Source: https://bestaivisibilityplatforms.com/ai-visibility-rfp-template · Publisher: Platform Ledger · Updated: 2026-11-13

## TL;DR
- An AI visibility RFP should specify the collection method behind every metric because vendors may use controlled prompts, public interfaces, panels, clickstream data, keyword datasets, or modeled estimates.[1]
- Score measurement validity, engine and market coverage, prompt representativeness, accuracy, security, workflow, total cost, and support rather than relying on an undefined industry weighting standard.
- Require repeated-run testing because the same brand list appeared in fewer than 1% of repeated ChatGPT or Google AI responses in SparkToro’s study.[11]
- Compare vendors with identical prompts, engines, markets, languages, time windows, raw-answer exports, and predefined pilot acceptance criteria.[14][12]
- Price prompts, runs, engines, locations, languages, integrations, exports, retention, overages, implementation, support, renewal, and data exit separately.

An effective AI visibility platform RFP defines **what will be measured, how the data will be collected, which engines and markets are included, how accuracy will be tested, and what the vendor will charge**. It should require comparable evidence rather than accepting “AI visibility” as an undefined dashboard metric, because vendors may use controlled prompt runs, public-interface collection, consumer panels, clickstream data, keyword datasets, or modeled estimates.[1]

This guide provides a practical AI visibility RFP template, 40 vendor questions, recommended evaluation criteria, and a pilot structure for comparing GEO platforms.

### What should an AI visibility RFP include?

A complete RFP should cover:

- Business outcomes and scope
- Engine and product-surface coverage
- Data-collection methodology
- Prompt sources, volumes, and refresh schedules
- Metric definitions and validation
- Security, privacy, and governance
- Reporting, exports, integrations, and APIs
- Pricing units, quotas, overages, and contract terms
- Implementation, support, and pilot acceptance criteria

Procurement guidance recommends using SaaS RFPs to expose hidden costs and risks, improve vendor comparability, and reduce negotiation time.[2] Supplier due diligence should also inform risk profiles, response plans, and contract clauses.[3]

## RFP sections and scoring weights

There is no published, AI-visibility-specific scoring standard in the reviewed sources. A Victorian government procurement guide provides a 30% cost and 70% requirements example but does not establish that split as a category benchmark; it also says weighted cost is not recommended in its example.[4] Federal Acquisition Regulation guidance permits several evaluation methods, including numerical weighting, when the method is tied to the stated evaluation factors.[5]

The following allocation is an **editorial starting point**, not an industry standard. Adjust it to your procurement rules and business priorities.

| RFP section | Questions to answer | Suggested weight |
|---|---|---:|
| Measurement validity | Are definitions, denominators, data sources, error rates, and audit trails clear? | 20% |
| Engine and market coverage | Does the platform cover the required engines, surfaces, countries, languages, and models? | 15% |
| Prompt representativeness | Are prompts controlled, observed, panel-derived, clickstream-based, or modeled, and does that fit the use case? | 15% |
| Accuracy and pilot performance | Can the vendor reproduce raw evidence and meet agreed acceptance thresholds? | 15% |
| Workflow and reporting | Are exports, APIs, alerts, integrations, history, and user workflows adequate? | 10% |
| Security and governance | Are privacy, retention, access, residency, subprocessors, and contractual protections acceptable? | 10% |
| Total cost of ownership | Are usage units, overages, add-ons, onboarding, renewal, and exit costs transparent? | 10% |
| Support and implementation | Are onboarding, training, escalation, and response commitments sufficient? | 5% |

Score every requirement using the same scale, such as 0–5, and require written evidence for scores above “partially meets.” Separate vendor claims from pilot results so a polished demo cannot outweigh weak measurement evidence.

## What data-collection questions should you ask vendors?

The phrase “AI visibility” does not identify the underlying dataset. For example, OtterlyAI describes programmatic interaction with public AI-platform interfaces and daily prompt checks, while Spyglasses documents logged-out, anonymous ChatGPT collection and location handling.[6] Peec describes daily refreshes for dashboard results, while Profound distinguishes daily answer-engine monitoring from prompt volumes derived from opt-in consumer panels.[7]

Semrush separately describes a toolkit based on AI-search clickstream and Google keyword data, alongside prompt tracking that queries buyer-selected prompts by platform and location.[8]

Require vendors to complete a methodology sheet for **each metric and module**, not just the platform overall. Ask them to identify:

- Collection source: public UI, API, panel, clickstream, keyword database, or modeled dataset
- Login state, account tier, personalization, chat history, and custom instructions
- Engine, product surface, model version, and subscription tier
- Country, city, language, device, and location method
- Prompt origin, selection, deduplication, translation, and refresh process
- Run frequency, retries, failures, and missing-result treatment
- Raw-answer retention and export options
- Metric denominator and uncertainty or bias information

A prompt-volume estimate should not be presented as observed demand unless the vendor can explain how it was measured. Ask whether volume is directly observed, modeled, weighted, or extrapolated, and whether it represents real-user behavior or responses to a fixed monitoring set.[9]

## What engine and coverage questions belong in the RFP?

Coverage should be specified by **engine and surface**, because ChatGPT Search, Google AI Overviews, Google AI Mode, Gemini, Perplexity, Claude, and Copilot can have different collection methods and availability. Product documentation reviewed for this guide lists different combinations of these surfaces across vendors.[1] Finseo's own study (Finseo is built by the team behind this site) found that [Google AI Mode and AI Overviews share only 13.7% of cited URLs](https://www.finseo.ai/blog/ai-mode-vs-ai-overviews) on the same query, so coverage of one Google surface should not be treated as coverage of the other.

Ask:

1. Which exact engines and surfaces are supported?
2. Which model or version produces each result?
3. Which account and subscription tiers are used?
4. Which countries, cities, languages, and devices can be selected?
5. Can location be controlled through IP or request geolocation, prompt context, an engine setting, or a combination?
6. Are personalization, chat history, cookies, and custom instructions enabled?
7. What happens when a surface is unavailable or returns no answer?
8. How are engine changes and model upgrades versioned?
9. Which historical data remains comparable after an engine change?
10. Are regional or language-specific results priced separately?

Do not accept a single “engines supported” checkbox. Require a matrix that states availability, collection method, cadence, location controls, model/version, exclusions, and evidence for every engine.

## What prompt-volume questions should you ask?

Prompt volume can mean the number of tracked prompts, the number of answer runs, the size of an observed panel, or an estimated market opportunity. These units are not interchangeable.

Ask vendors:

1. How many prompts are included in each plan?
2. How many runs does each prompt receive per day, week, or month?
3. Is a “prompt” counted once per project, engine, market, language, or run?
4. Are retries and failed requests billable?
5. Are panel-derived prompt volumes included or licensed separately?
6. How are prompts discovered and refreshed?
7. What proportion is customer-provided versus vendor-researched?
8. How are duplicate, translated, and near-duplicate prompts handled?
9. What sample size exists by engine, country, language, and time period?
10. What bias, confidence, or uncertainty information is reported?

For example, Profound describes panel-derived Prompt Volumes as using real prompts from double-opt-in consumer panels, with adjustments for demographic and geographic bias and rolling weekly refreshes.[7] That is a different product from daily monitoring of a buyer’s selected prompts.[10]

## How should you evaluate accuracy?

Define accuracy before vendors demonstrate their products. Include separate tests for:

- Brand and entity matching
- Mention detection
- Recommendation detection
- Citation and source extraction
- Position or rank extraction
- Competitor identification
- Missing and failed answers
- Historical continuity after model changes

Require vendors to provide the raw prompt, timestamp, engine, model or version, location, raw answer, extracted entities, cited sources, and final metric for each pilot record. This makes it possible to recompute scores and identify whether an error came from collection, parsing, classification, or aggregation.

Repeated answers can vary substantially. SparkToro’s study reported that the same brand list appeared in fewer than 1% of repeated ChatGPT or Google AI responses, while the same list in the same order appeared in fewer than 0.1% of cases.[11] The study involved 600 volunteers, 12 recommendation prompts, and 2,961 runs across ChatGPT, Claude, and Google AI tools.[1][11]

That finding means a single answer or rank is a weak basis for evaluating visibility. Ask vendors to report inclusion frequency separately from rank position and to show distributions across repeated runs.[9]


How can you test data accuracy in an AI visibility pilot?

Create a frozen test set containing the same prompts, engines, markets, languages, and time window for every shortlisted vendor. Run each prompt repeatedly, preserve the raw answers, and independently label a sample for brand mentions, recommendations, citations, sources, and position.

Compare the vendor’s output with the adjudicated labels using precision, recall, false-positive rate, false-negative rate, missing-response rate, and agreement by engine. Also compare the distribution of repeated answers rather than demanding that every run return the same list.

Do not claim that the research establishes one universally correct number of repetitions. One 2026 audit found that ten calls recovered most, but not all, of its observed 15-call organization set and explicitly did not establish a universal stopping rule.[12]



## What should the AI visibility pilot include?

Use the same test conditions for every shortlisted vendor. A practical pilot should produce both a buying decision and a reusable measurement baseline.

1. **Freeze the scope.** Select the brands, products, competitors, markets, languages, engines, surfaces, and user intents.
2. **Freeze the prompt set.** Include customer prompts, category prompts, competitor prompts, and discovery prompts; document their source and selection rules.
3. **Define the data dictionary.** Specify what counts as a mention, recommendation, citation, source, position, valid answer, failure, and duplicate.
4. **Set repetition rules.** Require identical prompts to run repeatedly under the same declared conditions.
5. **Require raw evidence.** Obtain prompts, timestamps, engine and model details, raw answers, source URLs, extracted entities, and error statuses.
6. **Measure separate outcomes.** Compare inclusion frequency, rank distribution, citation accuracy, source coverage, competitor detection, latency, and missing responses.
7. **Recompute a sample independently.** Have the buying team or an external evaluator check vendor scores against raw answers.
8. **Apply acceptance criteria.** Define minimum coverage, export completeness, data-quality thresholds, support response times, and delivery dates before the pilot starts.
9. **Record operational friction.** Score onboarding, configuration, permissions, documentation, support, and analyst usability.
10. **Close the pilot contractually.** Require a reproducible results file, issue log, export, deletion confirmation, and clear conversion or termination terms.

The pilot should not reward a vendor merely for producing the highest visibility score. It should reward transparent collection, defensible definitions, reproducible evidence, and useful decision support.

## What are the 40 questions for an AI visibility vendor?

### Outcomes and coverage

1. What business decisions should the platform inform, and what measurable outcomes define success?
2. Which brands, products, competitors, business units, and query intents are in scope?
3. Which exact engines and product surfaces do you cover?
4. Which countries, cities, languages, and device or account contexts can be selected?
5. For each engine, which model or version and subscription tier produced the result?

### Data collection and prompt volumes

6. Is each metric based on a public web UI, API, opt-in panel, clickstream, keyword database, or modeled prompt set?
7. Are UI requests logged in or out?
8. Are personalization, chat history, and custom instructions enabled?
9. How is location applied?
10. How often does each prompt run?
11. How many repeated runs are included?
12. Are exact prompts, timestamps, raw answers, retries, and failures retained?
13. Do you distinguish real-user demand from a fixed tracked-prompt list?
14. For panel data, what are the panel source, opt-in basis, anonymization method, and sample size?
15. Are prompt volumes observed, modeled, weighted, or extrapolated?
16. How are prompts selected, deduplicated, clustered, translated, and refreshed?

### Metrics, accuracy, and auditability

17. What is the exact formula and denominator for every metric?
18. How do you define a mention, cited source, recommendation, and content used without a citation?
19. How are aliases, repeated mentions, multiple citations, and omitted brands handled?
20. Can you report repeated-run distributions rather than only average rank?
21. What human review or ground truth validates entity and citation extraction?
22. What measured false-positive, false-negative, and missing-response rates can you provide?
23. Can evaluators inspect and export raw prompt-answer-source records?
24. How are engine and model changes versioned?
25. How are historical breaks flagged?
26. How are failed, blocked, stale, and unavailable answers represented?
27. Can you report inclusion frequency separately from rank position?

### Security, pricing, support, and pilot

28. What user, panel, or account data do you collect?
29. What consent, anonymization, retention, deletion, and residency controls apply?
30. Which subprocessors are used?
31. Will you provide a data-processing agreement and security documentation?
32. What SSO, role-based access, audit logging, and incident-notification terms are included?
33. Who owns prompts, answers, exports, and annotations?
34. May customer data be reused to train or improve the product?
35. Quote pricing by prompts, runs, engines, locations, languages, domains, seats, API calls, and retention.
36. What quotas, overage rates, add-ons, and refresh limits apply?
37. What fees apply to onboarding, integrations, exports, APIs, historical data, and premium support?
38. What are the renewal, price-change, termination, export, and deletion terms?
39. What are the pilot duration, included engines and markets, prompt volume, resources, and acceptance thresholds?
40. At pilot close, will you provide raw evidence, a reproducible results file, an issue log, and documented data exit and deletion?

## How should pricing be compared?

Compare vendors using the unit that actually drives your use case. A low subscription price may become expensive when every engine-market-language combination creates a separate billable run.

Request a pricing workbook with:

- Base platform fee
- Tracked prompts
- Answer runs
- Engines and surfaces
- Countries and languages
- Domains, brands, and projects
- Seats and permissions
- API calls and exports
- Historical retention
- Panel or prompt-volume data
- Integrations
- Onboarding and implementation
- Premium support
- Overage rates
- Renewal increases
- Termination, export, and deletion fees

Ask vendors to price three scenarios: the initial pilot, the first-year production workload, and a 2–3x growth case. Require them to state whether failed, retried, blocked, and duplicate requests consume usage.

## What security and governance terms should the RFP require?

The RFP should ask what data the vendor collects, where it is stored, how long it is retained, who can access it, and whether it is reused for product improvement. Require documentation for subprocessors, data-processing terms, access controls, incident notification, deletion, export, and business continuity.

For panel-based products, ask specifically about consent, anonymization, sampling, demographic and geographic weighting, and the separation between panel data and customer-tracked prompts.[7][13]

For all vendors, require the right to audit a representative sample of records and to receive sufficient metadata to reproduce the reported score. A dashboard without an evidence trail should not receive full marks for measurement validity.

## Which contract and support questions matter?

The contract should convert pilot promises into measurable obligations. Include:

- Named implementation owner
- Onboarding milestones
- Training and documentation
- Support channels
- Severity definitions
- Response and resolution targets
- Escalation contacts
- Engine or model outage handling
- Data-quality incident handling
- Historical-data correction process
- Export format and timing
- Renewal notice period
- Price-change limits or notice
- Termination rights
- Data deletion certificate

Ask whether support commitments differ by plan and whether API access, raw exports, historical data, or advanced locations require separate fees. Gartner procurement guidance emphasizes examining missing terms, risks, and hidden costs before signing, while supplier due diligence should flow into contractual protections.[2][3]

## FAQ
### What is an AI visibility RFP?
An AI visibility RFP is a structured request for vendors to explain how they measure a brand’s presence in AI-generated answers. It should cover engines, prompts, collection methods, metrics, accuracy, security, pricing, support, and pilot requirements rather than asking only for a product demonstration.[1]

### How many questions should an AI visibility vendor RFP include?
Forty questions provide practical coverage across outcomes, engines, data collection, prompt volumes, accuracy, governance, pricing, support, and pilots. The exact count matters less than requiring vendors to answer each question with evidence, definitions, limitations, and commercial assumptions.

### What is the most important question to ask an AI visibility vendor?
Ask what data source produces each metric. Vendors may use controlled prompt runs, public web interfaces, opt-in panels, clickstream data, keyword datasets, or modeled estimates, and those sources answer different business questions.[1]

### How should an AI visibility platform be tested during a pilot?
Give every shortlisted vendor the same frozen prompts, engines, markets, languages, time window, and repetition rules. Retain raw answers and compare inclusion frequency, rank distributions, citations, source extraction, missing responses, and independently recomputed scores.[12]

### How many repeated AI prompts are enough for a vendor pilot?
No universal repetition count is established by the reviewed research. A 2026 audit found that ten calls recovered most, but not all, of its observed 15-call organization set and explicitly did not establish a universal stopping rule.[12] Set the number in advance, apply it equally to every vendor, and report the resulting uncertainty.

### Should rank position be the main AI visibility KPI?
No. Rank can change substantially across repeated runs, so the RFP should request inclusion frequency, rank distribution, citation accuracy, and source coverage separately.[9] Rank can remain useful, but it should not be treated as a stable ground truth from one answer.

## Sources
[1] NEW Research: AIs are highly inconsistent when recommending brands or products; marketers should take care when tracking AI visibility - SparkToro — https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility
[2] Tool: Fast-Track SaaS RFPs by Evaluating Key RFP Terms and Conditions — https://www.gartner.com/en/documents/6868667
[3] Successful Sourcing and Procurement: Key Strategies & Steps — https://www.gartner.com/en/supply-chain/topics/sourcing-and-procurement
[4] Evaluate and select offers: Goods and services guide | buyingfor.vic.gov.au — https://www.buyingfor.vic.gov.au/evaluate-and-select-offers-goods-and-services-guide
[5] 15.305 Proposal evaluation. | Acquisition.GOV — https://www.acquisition.gov/far/15.305
[6] AI Visibility Data Collection Methodology — https://www.spyglasses.io/api/docs-markdown/en/methodology/ai-visibility-data-collection
[7] Prompt Volumes FAQ — https://docs.tryprofound.com/prompt-volumes/faq
[8] Where does the data in Semrush’s AI Visibility Toolkit come from? — https://semrush.com/kb/1607-ai-visibility-data
[9] A statistical framework for evaluating the repeatability and reproducibility of large language models - PMC — https://pmc.ncbi.nlm.nih.gov/articles/PMC12637745/
[10] How often does OtterlyAI check AI search engines? — https://help.otterly.ai/monitoring-interval
[11] AI Recommendations Change With Nearly Every Query: SparkToro — https://www.searchenginejournal.com/ai-recommendations-change-nearly-every-query-sparktoro/566242
[12] Measuring Brand and Source Discovery under Repeated LLM Queries:A Finite-Sample Audit — https://arxiv.org/html/2609.05059
[13] FAQ | Profound Help Center — https://help.tryprofound.com/articles/4340126061-faq
[14] Is one run enough? Reproducibility of flagship large language models across temperature and reasoning settings in biomedical text processing | Journal of the American Medical Informatics Association | Oxford Academic — https://academic.oup.com/jamia/article/33/6/1179/8559659
