How to write an RFP for AI visibility work: 30 questions, and the answers that should worry you

Why AI visibility work needs a proper RFP

Because supply has grown faster than standards. An RFP forces every bidder to answer the same questions in writing, which is the only way to compare an agency, a tool vendor and a consultancy that describe the same service in different words.

Demand is pulling agencies into the work quickly. In AgencyAnalytics' 2026 benchmarks survey of 494 agency professionals, 64% of them in the US (February to April 2026), 66% reported increased client demand for AI-search services.¹ In the same survey, 64% cited Google's AI Overviews as their top concern.¹ Many of the agencies bidding for your work will have added the service within the past year.

Tools have multiplied too. G2 listed more than 150 AEO software products by January 2026, up from 7 in March 2025.² The market is also consolidating: Sitecore announced in June 2026 that it had acquired the AI visibility platform Scrunch.³ A bid built on a tool can change when the tool's owner does.

Buyers are already sceptical. Digiday reported in May 2026 on marketers questioning the price of AI visibility tools as results varied between runs.⁴ Our view: that scepticism is healthy, and an RFP is where you turn it into specific questions instead of a vague unease.

Who will bid

A young market, growing fast

66%of agency professionals report more client demand for AI-search servicesAgencyAnalytics, Feb–Apr 2026 ¹

64%name Google AI Overviews as their top concernAgencyAnalytics, Feb–Apr 2026 ¹

150+AEO software products listed on G2, up from 7G2, Mar 2025 to Jan 2026 ²

How should the RFP be structured?

Keep it short and make it comparable. Send a two-page brief, the 30 questions below, a pricing template and your scoring sheet. Ask for answers in the same order, with a word limit per answer, so you compare substance and not presentation.

The brief should state your category, the countries and languages that matter, three to five competitors, and the outcomes you care about. Name the decision you need to make at the end of the first phase, such as whether to fund a programme. Bidders who understand the decision write better proposals.

Give bidders a short window for clarifying questions and share every answer with all of them. Then allow two weeks for written bids and shortlist two or three for a call. Our view: a paid pilot or audit is a better final test than a pitch meeting, because it shows the method working on your prompts.

A four-week RFP

1. Week 1

   Send the brief, the 30 questions, the pricing template and the scoring sheet.

2. Week 2

   Answer clarifying questions and share every answer with all bidders.

3. Week 3

   Receive written bids; score them independently, then compare scores.

4. Week 4

   Call two or three shortlisted bidders; agree a paid audit or pilot with one.

What should you prepare before sending it?

Your own first look at the answers, and a draft prompt list. Bidders write sharper proposals when they can see what you already know, and you read bids better when you have seen the problem yourself.

Spend an afternoon asking ChatGPT, Gemini, Perplexity and Google's AI Mode the questions your buyers ask. Note whether you are named, which competitors are, and what is said about you. Do it on two different days and you will see for yourself why run counts matter: some answers change.

Then write twenty to thirty prompts in your buyers' words, grouped by stage: problem, comparison and vendor choice. Share the list in the brief and ask bidders to improve it. How a bidder edits your prompts tells you more about its judgement than its case studies do.

Finally, decide who scores. Two people from different teams, such as marketing and sales operations, catch different weaknesses. Agree in advance what a 1, a 3 and a 5 look like for each area, so scores reflect the bids and not the scorers' moods.

Scope: questions 1 to 5

Scope questions find out what the bidder will actually do, as opposed to what it will measure. AI visibility work has three parts: measurement, on-site fixes and off-site work such as earning coverage. Many bids cover one well and gesture at the others.

1. Which of measurement, on-site work and off-site work does your bid include, and which do you leave to us or another supplier?
2. Which of our products, services or locations will the prompt set cover, and which are excluded?
3. What will you deliver in the first 30 days, and what decision will it let us make?
4. Who on our side needs to be involved, for how many hours a month, and with what access?
5. How does your work fit alongside our SEO agency and content team, and who owns which pages?

A good answer to the first question names the parts not covered. A bid that claims to do everything for a small monthly fee is usually doing measurement and calling the rest a recommendation.

Method: questions 6 to 12

Method questions are where weak bids fall apart. AI answers vary between runs, engines and countries, so a measurement that does not say how it handles that variation cannot be compared with anything, including itself next month.

6. How many prompts will you track, who writes them, and will we approve the final list before measurement starts?
7. Which engines do you measure, and how do you reach each one: an official API, a licensed data provider or scraping?
8. How many times is each prompt run, on how many separate days, before a result is reported?
9. Which countries and languages do you measure from, and how is location set for each engine?
10. How often is the full prompt set re-measured during the engagement?
11. What does your headline score count, and how is it calculated from individual answers?
12. Do you report a margin of error or confidence interval, and how do you tell a real change from noise?

MaxAEO, a vendor in the category, has published guidance on how sample size and clustered runs affect AI visibility measurement, which is a useful primer before you read bids.⁵ An April 2026 preprint by Schulte, Bleeker and Kaufmann, "Don't Measure Once", argues for measuring AI visibility repeatedly, as a distribution rather than a single reading.⁶ Our buyer's spec for an AI visibility score sets out what a credible score should disclose.

Our view: question 8 is the single best filter. A bidder who runs each prompt once and reports the result as a share is reporting one roll of the dice. Ask to see a sample report with run counts on it.

Data ownership: questions 13 to 17

You should own the prompt set, the raw answers and the reports. Without them you cannot switch supplier, check a claim or compare this year with last year. Write ownership into the contract, not just the proposal.

13. Who owns the prompt set, the raw answers captured and the reports at the end of the engagement?
14. In what format can we export the raw answers, and how often?
15. Which third-party tools hold our data, and what happens to it if a tool is sold or shut down?
16. How long do you keep our data after the contract ends, and how is it deleted?
17. Can we see the raw answer behind any number in a report, on request?

Question 15 matters more than it did a year ago, given acquisitions such as Sitecore's purchase of Scrunch. Our note on AI visibility tool contracts lists the clauses to negotiate: export rights, notice of price changes, engine coverage and exit terms.

Reporting: questions 18 to 21

Reporting questions tell you whether you will be able to act on what you receive. A monthly PDF of scores with no prompts behind them is a status update, not a report.

18. How often will we receive reports, and what does each one contain?
19. Will reports show results per engine, per country and per prompt group, as well as a total?
20. How do reports connect visibility to outcomes we already track, such as enquiries or pipeline?
21. Who explains the report to us, and how quickly do you flag a sudden change?

Ask for a redacted sample report from a current engagement. A bidder who cannot show one has either no clients or nothing they are proud of. Our view: a weekly dashboard with a monthly written review is the cadence that suits most buyers, because engines change faster than monthly reporting can catch.

Be realistic about question 20. AI referral traffic is small and attribution is weak, so a bid that promises to prove revenue impact in the first quarter is overpromising. Expect leading measures first, such as being named and described accurately, and pipeline measures later.

Pricing: questions 22 to 26

Pricing questions should expose what the price includes, how it can change and what you pay if results do not arrive. Ask every bidder to complete the same pricing template so totals are comparable over twelve months.

22. What is the total cost over twelve months, including tools, set-up and any per-prompt or per-engine fees?
23. Which pricing model do you use: fixed project, monthly retainer, per-prompt tool fee or fees linked to outcomes?
24. If part of the fee is linked to outcomes, which measure is used, who measures it and how are disputes settled?
25. What is the minimum term, and what does it cost to leave early?
26. Which costs can rise during the contract, and with how much notice?

Published prices make this section easier. Our survey of published GEO prices shows what agencies and tools list openly, which gives you a reference before bids arrive. If you are also scoping the monthly work, our checklist for retainer deliverables separates monitoring, on-site fixes and off-site work line by line.

Team and delivery: questions 27 to 30

The last four questions find out who does the work. In a young market, the person who pitches and the person who delivers are often different, and the gap shows up in month two.

27. Who will work on our account, what have they delivered before, and how many accounts does each person hold?
28. Which parts of the work are done by AI tools, which by people, and who reviews the output?
29. Can you show a dated example of a change you measured for a client, with the method behind it?
30. What will you do if measurement shows no improvement after the first phase?

Question 29 asks for evidence, not testimonials. A dated before-and-after with prompt counts and run counts is worth more than a logo slide. If a bidder has no such example yet, an honest "not yet" is a better answer than a vague success story.

Which answers should worry you?

Answers that avoid specifics on method and data. A worrying answer is not always wrong; it is unverifiable, and an unverifiable answer cannot be scored against a better one. The comparison below pairs common answers with the ones you want.

Answers that worry, and answers that reassure

Should worry you

• "Our proprietary score" with no formula
• Prompt list shared only after signing
• One run per prompt, reported as a share
• A promised position in a named engine
• Raw answers stay in the vendor's tool
• Revenue impact promised in the first quarter

Should reassure you

• The score's formula stated in the proposal
• A draft prompt list you approve before measuring
• Several runs per prompt on separate days
• Targets with the method to measure them
• Raw answers exported to you on a schedule
• Leading measures first, pipeline later

Two more patterns deserve a second look. A bid that measures only ChatGPT, or only Google, will miss engines your buyers use. And a bid that leads with tactics, such as adding llms.txt or schema everywhere, before it has measured anything has the order wrong.

How to score three bids

Agree the weights before bids arrive, and have two people score each bid independently. Weight method highest, because everything else depends on it, and keep price below method so a cheap bid cannot win on price alone.

Illustrative. Three fictional bidders, each area scored 1 to 5, multiplied by its weight and divided by 5. Weights and scores are invented.

In this example the cheapest bid, Northwind Digital, scores lowest. It runs each prompt once and keeps raw answers in its own tool. Contoso Growth wins on method and data terms despite a higher price. That is the outcome the weights are designed to produce, and it is why the weights are fixed before anyone sees a price.

Set a floor as well as a total. Our view: a bid that scores 1 or 2 on method should be rejected whatever its total, because you will not be able to trust anything it reports.

How we would answer these questions

A disclosure: Sigzen AI bids for this work, so weigh our answers accordingly. Here is how we answer the method and pricing questions, as published on our pricing page.

On our published prices, the AI Visibility Audit costs $3,500 (₹1.5L), one-time, over two to three weeks. It covers 150 buyer prompts across six engines: ChatGPT, Gemini, Perplexity, Claude, Copilot and Google AI. Each prompt runs on every engine on three separate days, and results are reported with confidence intervals against five competitors. Our AI answer monitor reaches engines through official APIs where they exist, and a licensed AI-answer data provider where they do not.

Also on our published prices, the Growth programme costs $3,500 (₹1.25L) a month with a six-month minimum, a weekly dashboard and part of the fee linked to share of answer. Audits are credited 50% against a sprint booked within 30 days. You keep the audit report. If a bid you receive answers the 30 questions better than we do, take it.