Buyers ask AI who uses your product and what they got. Your customers page is the answer.

/ 6 min read / By Faz

Somewhere between the shortlist and the contract, a buyer wants proof. Not your claims, other companies’ results. Who like us uses this, and what did they actually get? That question used to send people to your case studies page, a reference call, or a review site. More of them now ask an AI engine.

The advice you will find on this says AI engines love case studies, so publish more of them with hard numbers. I wanted to see what the answers actually do with them. In November I asked Google AI Mode three proof questions about three well-known products: a corporate card and spend platform, an AI support agent, and a data enrichment platform for sales teams. Each asked for named customers and real numbers. One added “not marketing claims”. Then I read every source behind each answer.

The customers page became the answer

The spend platform was Ramp. Its answer used five sources: Ramp’s homepage, its customers page, a blog post rounding up its case studies, a LinkedIn post by a co-founder, and a McKinsey interview with its CEO. The body of the answer closely followed Ramp’s customers page. Named companies, one number each: a company that saved over $2 million, one that reclaimed 400 hours a month from expense reports, one that halved its month-end close. The engine even built a table from it.

The data enrichment platform’s answer looked similar. Its table listed seven well-known customers, each with a metric, and the numbers came from the cards on the vendor’s customers page.

So the advice is partly right. A page that pairs a named customer with a specific result is exactly what a proof question wants, and the engine will lift it whole. But the other two findings were less comfortable.

The engine discounted the headline number

The AI support agent’s answer used 12 sources, five of them the vendor’s own: two customers pages, two individual case studies and a thread in its community forum.

The answer opened by setting the vendor’s headline figure, an average resolution rate across thousands of customers, against third-party benchmarks. It called the headline the marketed number, then said B2B companies with complex products realistically see a lower range, and cited a third-party benchmark site and a comparison post for that. My question had asked for no marketing claims, so the framing is partly my doing. But it shows what happens to an aggregate claim when the buyer is sceptical: it gets treated as the vendor’s opinion.

Then the answer listed named customers, and four of the five came from the vendor’s own case studies. Each had a specific number for a specific company, with context: how many brands one customer merged onto the platform, how long another had been live. Those were treated as evidence, not marketing. The caveat that followed, about how the vendor counts a resolution for billing, came from a thread in the vendor’s own community forum.

The difference between the two kinds of proof was not who wrote it. It was whether a sceptic could check it. “Our customers average X” cannot be checked. “This named company got Y after Z weeks” can.

The engine mixed up whose number was whose

Back to the data enrichment platform. One row in its table credited a well-known design software company with three results: double the cold email performance, 50 percent more qualified meetings per rep, and 140 percent growth in outbound pipeline.

I checked the vendor’s customers page. That company’s card carries a headline about building its go-to-market infrastructure, and no number in its text. The three numbers sit on neighbouring cards. The page lays out results and customer names in separate blocks, and the answer appears to have paired them by position. A second customer in the same table was also given the double cold email figure.

I cannot tell from the outside which company each number really belongs to. Neither can the buyer, and that is the problem. A proof page the engine misreads hands out your best results to the wrong logos, and some buyer will repeat them on a call.

The rest of the answer’s detail, the “how they did it” part, came from elsewhere: an agency’s post about a campaign it ran for a client, a third-party site that republishes case studies, and a LinkedIn post from a practitioner. The vendor’s own case studies were in the source list, but the engine went to other people for method.

What to do

Put the customer name and the number in the same sentence. “Acme cut month-end close from 20 days to 12” survives being lifted. A card with a logo, a separate stat block and a “Read story” link does not. Check your customers page as plain text, without the layout, and see whether every number still sits next to the right name.

Lead with specific results, not averages. An aggregate figure across all customers reads as marketing to a sceptical buyer, and the engine will set it against third-party benchmarks. A named customer, a number, a timeframe and the starting point read as evidence. If you publish an average, publish how it is calculated, because someone else will explain it if you do not.

Write the method into the case study. The answer went to an agency and a practitioner for how results were achieved. If your case study says what the customer actually did, in what order and how long it took, the engine can quote you for that too.

Cover the segments buyers ask about. The questions asked for mid-size companies and B2B companies, and the answers described each customer by what it does and picked ones that fit. Tag your case studies by company size, industry and use case in the text itself, so a question with a segment in it has something to match.

Make your proof checkable by someone else. Proof that appears outside your site, in a customer’s own post, an interview or a review, carried weight in these answers. It is the same reason review sites feed so many AI answers. Ask happy customers to say it in their own channels, not only in your PDF.

This is a smaller version of a problem we have written about before: most AI search statistics are vendor marketing, and readers are learning to discount them. Engines are doing the same thing to yours.

What I got wrong

I expected the engine to treat all vendor-published proof the same way, either trusting it or discounting it. It did neither. It discounted the vendor’s average and trusted the vendor’s individual case studies in the same answer. The line it drew was between a claim and a checkable example, not between the vendor’s site and everyone else’s.

I also assumed that if the engine quoted a customers page, it would quote it accurately. On one of the three, it did not.

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 proof questions before you act on mine.

Where this leaves you

Your buyers are asking an engine who uses your product and what they got. The engine will answer from your customers page if it can read it, and it will believe the specific examples more than the averages. Make every result a sentence with a name in it, and check that the engine reads it the way you meant.

If you want to see what the engines tell your buyers when they ask for proof, and where it comes from, that is part of the engagement, and it runs on a documented method.

Want this run for your B2B SaaS?

Founding pricing for the first 5 clients. Methodology fully public. Month-to-month, cancel anytime.

Apply to work together

Leave a comment

Your email address will not be published. Required fields are marked *