Every B2B deal of any size hits a security review. Somebody on the buyer’s side, often not the person who found you, has to approve the tool. Before they open your questionnaire or request your SOC 2 report, more of them now type the question into an AI engine: is this vendor SOC 2 compliant, where is our data stored, do they train AI on it.
The answer they get shapes the review. If it says you are missing a certification, the reviewer starts sceptical. If it says your data residency does not fit their policy, the deal may stall before anyone talks to you.
So in November I asked Google AI Mode three of these questions about three software products: a scheduling tool, an AI notetaker, and a newer CRM. Each question was written the way a reviewer would write it, naming the product and saying it was for a security review. Then I read every source behind each answer.
One vendor owned its answer almost completely
The scheduling tool was Calendly. The answer drew on 11 sources, and 9 of them were Calendly’s own pages. Its security page, a blog post on enterprise compliance, a page for IT teams, and six separate help centre articles: one on SOC compliance, one on data storage and international transfers, one on where its AI assistant’s data is stored, one on platform security, and two more on privacy and compliance.
The two outside sources were a HIPAA explainer and a GDPR explainer from other sites. Everything else the reviewer read came from Calendly, in Calendly’s words. SOC 2 Type 2, ISO 27001, US-only hosting, no HIPAA business associate agreement, how to request the full report. Even the limits were Calendly’s own statements, which is the best case for a vendor: the bad news arrives accurately and without spin.
What made the difference was not a trust centre. Calendly’s full reports sit behind an NDA on a third-party platform. What the engine used were plain public pages, each answering one question a reviewer asks.
The other two vendors shared their answers
The AI notetaker’s answer used 9 sources, 4 of them its own: its trust centre, the trust centre’s FAQ, a help article and a product page. Those covered the headline facts, SOC 2 Type II and the rule that its AI subprocessors cannot train on customer data.
The details a reviewer cares about most came from elsewhere. Which plans include a HIPAA business associate agreement was attributed to a LinkedIn post. Where the data is hosted came from a third-party blog post about the product and GDPR. That the vendor trains its own models on de-identified data by default unless you opt out came from a security blog. How long recordings are kept came from another LinkedIn post. And under the answer, a rival notetaker had bought an ad pitching itself as the compliant alternative.
The CRM’s answer used 7 sources, 3 of them its own: its trust centre and two help centre pages. Two more were from a rival CRM’s blog, one framed as a security comparison of the product and one about how it handles customer data. The answer’s section on data residency, the part that tells an EU buyer whether they can host in Europe, cited the rival’s article first.
I have not checked whether those third-party claims are accurate. That is the point. The reviewer reading them cannot easily check either, and the vendor did not write them.
Why the gaps get filled by someone else
A security reviewer’s questions are specific and predictable. Which certifications, and which type. Where the data lives, and whether you can choose. Which subprocessors, and whether any of them train on customer data. Retention. Who gets a business associate agreement. SSO and SCIM, and on which plan.
When the vendor’s public pages answer each of those plainly, the engine has no reason to go elsewhere. When the answer is behind an NDA, inside a PDF, or only in the questionnaire your team sends back, the engine finds whoever did write it down. That is often a reviewer blog, a LinkedIn post, or a competitor who has every reason to stress your gaps.
This is different from the security software category, where certifications are the trust signal the engine sorts vendors by. Here the product is not a security product at all. A scheduling tool or a CRM still has to pass a review, and the review now starts with an engine.
What to do
Ask the reviewer’s questions yourself. Take them from the questionnaires your sales team fills in most often. Ask in a reviewer’s words: “[you] SOC 2 type”, “where does [you] store data”, “does [you] train AI on customer data”, “[you] HIPAA BAA”. Run each several times on more than one engine, and list every source behind each answer.
Publish one plain page per question. The pattern in the strong answer was a help centre with a separate article for each question, each one specific. A trust centre home page with badges and a request button does not answer “where is our data stored”. A short public article that says the region, the cloud provider, whether residency is available and on which plan does. Your security team already has these answers in the questionnaire. Most of them can be public.
Say the limits yourself. If you do not sign business associate agreements, or residency is Enterprise only, or your SOC 2 is Type I while the Type II audit runs, say so on your own page with dates and what is coming. The engine will report the limit either way. The only choice is whether the reviewer reads it in your words or a rival’s.
Put your AI data practices in writing. Training on customer data was the point where these answers leaned hardest on third parties. If you train on de-identified data by default, publish what that means, how to opt out, and who can do it. If you do not, say it in one plain sentence.
Keep it current and correct what is wrong. Certifications renew, subprocessors change, residency options get added. Date your pages. If a third-party review has your facts wrong, get it corrected where it lives, because your page will not outvote it on its own.
What I got wrong
I expected the trust centre to be the page that mattered, and that the fix would be making it crawlable. It was cited, but it was not what won. The vendor with the strongest answer won through ordinary help centre articles, one question each. A gated trust centre mostly gives the engine a sentence saying reports are available on request. The actual answers came from public text.
These were single runs of three questions about three products on one engine, and Google offered to rerun them without personalisation. Treat this as a read of the pattern, not a measurement. Run your own reviewer questions before you act on mine.
Where this leaves you
Your buyer’s security reviewer is asking an engine about you before they ask you. Where your own public pages answer the question, the engine quotes you. Where they do not, it quotes whoever wrote it down, sometimes a competitor. The questions are known in advance, because your team answers them every week. Put the answers where the engine can read them.
If you want to see what the engines tell a security reviewer about you, and who wrote it, that is part of the engagement, and it runs on a documented method.
3 responses