27 questions to ask an AI visibility or GEO vendor before you sign, and our own score against them

A procurement checklist is only useful if each item can be answered with evidence rather than reassurance. These 27 questions are written so that every one of them has a document, a number or a demonstration behind it, and a vendor who cannot produce it has answered the question anyway.

The 27 is our count, not an industry standard. At the end we score ourselves against it, because a checklist published by a vendor who will not take it is a sales tool.

Measurement and reporting, questions 1 to 7

1. Which engines and surfaces do you read, named individually, and is Google AI Overviews counted separately from Gemini. They are different surfaces.
2. For each engine, is the reading taken from the consumer app, an official API or a browser session. The three can route to different models with different retrieval.
3. What counts as a mention, a citation and a recommendation in your reports, written out as three separate definitions.
4. What is the denominator for each rate you report, and are readings that named no company at all excluded from it.
5. Can we inspect raw observations, including the full answer text and the cited URLs, and export our history.
6. How many times is each prompt run, and how do you separate a real change from ordinary answer variation.
7. What is your coverage rate, meaning successful observations divided by scheduled observations, for the last complete month.

Method and data quality, questions 8 to 13

1. Where does the data come from. Your own collection, an official API, or a licensed third-party provider.
2. How are prompts selected, who approves them, and how often are they refreshed.
3. Are branded and unbranded questions reported separately, and is our brand name excluded from the unbranded set.
4. How are failed runs and missing data reported, and are they ever counted as a zero mention.
5. How is the market or country fixed, and how is it verified in the answer rather than trusted to a setting.
6. Has your detection been checked against answers a human read, and what was the agreement rate on what sample size.

The work being delivered, questions 14 to 18

1. What exactly is produced each month. Content, technical changes, outreach, or measurement only.
2. Who has editorial review and approval before anything is published in our name.
3. What are the deliverables, the deadlines and the acceptance criteria, in the contract rather than the pitch.
4. What evidence sits behind any projection you have shown us, and what was the comparison group.
5. Show one engagement where the work did not produce a change, and what you did next.

Security, privacy and ownership, questions 19 to 24

1. Do you offer single sign-on, and is SAML supported.
2. Do you hold SOC 2, ISO 27001, or any independent audit, and can we see the report or its date.
3. Will you sign a data processing agreement, and do you publish your subprocessors.
4. What is the data retention period, and who inside your company can access our account data.
5. Who owns the content, the data and the account at termination, and how is it exported or deleted.
6. Is white labelling or multi-brand access part of the plan, and how many brands does one account genuinely cover.

Commercials, questions 25 to 27

1. What is the published price of the plan that includes every engine we need, with the date the price was published, and which allowances it carries.
2. What does the cheapest plan actually read. A low headline price often reads nothing.
3. What are the cancellation, renewal and exit terms, in writing.

Our own score, 9 verified of 27

Verified here means a figure recorded in our own product-fact check of 22 October 2026, tested against our published pricing page rather than remembered.

• Question 1, engines. Verified. We read five surfaces, not nine.
• Question 19, single sign-on and SAML. Verified absent. Our pricing page, 119,203 bytes, renders neither, tested by word boundary across the whole document. A loose search for sso on that page matches crossOrigin seven times and subprocessors twice, which is why a careless check reports this wrongly in both directions.
• Question 20, SOC 2 and ISO. Verified absent. Neither string renders on that page.
• Question 21, data processing agreement. Verified absent as a published term. No DPA renders on that page.
• Question 24, white label and multi-brand. Verified absent, and contradictory. No white label renders. The Agency plan is described as covering 10 client brands while the comparison table answers No to more than one business in all eight columns. Both render today, and the unflattering reading is the one to plan around.
• Question 25, published price with allowances. Verified. SEO Start 149, Growth 299, Scale 599, SEO Scale 699, Inbound 1099, Outbound 1299, Agency 1299, Everything 1799 dollars a month, read on 22 October 2026.
• Question 26, what the cheapest plan reads. Verified, and this is the one to insist on everywhere. SEO Start at 149 reads 0 questions, 0 credits and 0 engines. The cheapest plan that reads all five surfaces is Growth at 299. Anybody writing that our AI visibility starts at 149 dollars has read the price and not the allowance.
• Question 4, denominator. Verified in our published measurement. We report eligible readings as the denominator, and we publish the count of readings that named no company at all, which was 184 of 635 on 22 October 2026.
• Question 17, evidence behind projections. Verified, and it is a negative result. Between 20 and 22 October 2026 we went from 134 recommendation-shaped pages to 641 and our naming count went from 0 to 0, over 1,260 and then 635 readings. We publish our own null result rather than a projection.

One item is partial. Question 25 also asks about free terms, and our pricing page renders a free scan on your own domain and a free plan that stays free, while it does not render that scan's exact question limit, brand count or engine coverage. We will not state a number our own page does not carry.

The remaining 17 are not evidenced on a published page in this check, which is the honest answer and the one we would want from anybody else. A checklist is worth more when the vendor publishing it fails part of it in public.

Where a competitor passes items we fail

Read on 22 October 2026 at tryprofound.com/pricing, Profound publishes SSO and SAML with SOC 2 compliance on its Enterprise plan and up to 9 answer engines tracked, naming ChatGPT, Perplexity, Google AI Mode, Gemini, Microsoft Copilot, DeepSeek, Anthropic Claude, Google AI Overviews and Exa Search. It publishes no numeric price for that plan, and its trial covers ChatGPT, Gemini and Google AI Overviews with 50 prompts daily for 7 days. On questions 1, 19 and 20 they are ahead of us on published evidence, and if a security review decides your purchase, that is where to start.

How to score a vendor without a spreadsheet war

Make four of the 27 pass or fail and treat the rest as weighting. The four we would make mandatory are question 4, the denominator, question 5, access to raw observations, question 11, how failures are reported, and question 26, what the cheapest plan actually reads. A vendor who cannot answer those four has a dashboard rather than a measurement, however good the demonstration looks.

Two red flags worth more than any score. A promised ranking or placement outcome, which nobody controls. And a case study without a comparison group, which cannot separate the vendor's work from the market and the model updates that happened at the same time.

Questions buyers ask before signing a GEO contract

Is a 27-item checklist overkill for a small purchase

Use the four mandatory ones. They take one email and they eliminate most of the risk.

What if the vendor will not share raw observations

Then no number in the report can be checked, including the good ones. Treat it as a pricing question rather than a principle, because unverifiable reporting should cost less.

Should I expect SOC 2 from an AI visibility vendor

Larger buyers increasingly do, and we do not publish it. That is a real reason to buy elsewhere and we would rather you learned it here than three weeks into a procurement review.

How do I tell a measurement from a dashboard

Ask for the coverage rate for last month and the agreement rate of the detector against answers a person read. Both are single numbers, and both are hard to produce if the product does not actually measure.