Last Tuesday I ran the same query in Perplexity three times: “what do customers say about Decagon AI versus Sierra.” Twice, Perplexity surfaced direct quotes from a Decagon case study page. Once, it quoted a G2 review verbatim. Sierra got two paraphrased mentions. No direct quote. In a category where the two companies are within 8% of revenue, that asymmetry is worth a quarter.
Testimonials for AI search are a different beast than testimonials for a landing page. Landing-page testimonials exist to convert a visitor who already arrived. AI-search testimonials exist to be retrieved, parsed, attributed, and reproduced inside an answer your visitor never even saw. The encoding rules are different. The placement rules are different. The volume rules are different.
What follows is the working rule set, based on six months of running comparative queries across ChatGPT, Perplexity, Claude, and Gemini for B2B AI tooling, SaaS, healthcare, and consumer brand searches.
1. Use the customer’s actual words, in quote marks, attributed by full name
LLMs retrieve attributable content. Anything inside double quotes followed by a name and title gets weighted as quotable. Anything written in the third person (“Customers love how we”) gets discarded as marketing copy.
Compare two pages on the same site. Page A says “Acme cut onboarding time by 40% with our platform.” Page B says: “We used to spend 9 days getting a new analyst productive. Now it is under three. The platform paid for itself in the first month.” (Maria Sosa, VP Analytics, Acme Corp.)
In a head-to-head retrieval test using Perplexity’s API, Page B was cited in 8 of 10 relevant queries. Page A was cited zero times. The retrieval engine reads quotation marks and proper-noun attribution as a verbatim claim it can reproduce without rewriting.
If your testimonials are not in quote marks with a full name, title, and company, they are invisible to AI search. Rewrite them or you will not get pulled.
2. One quote per page, not a wall of them
The instinct most marketing teams have is to dump 12 testimonials into a single wall-of-love page. That wall is structurally invisible to retrieval. LLMs treat a clustered list of quotes as a single chunk and either return all of it or none of it. Usually none.
Spread the testimonials across the site. One quote per case study. One quote per product page. One quote per pricing page. One quote per industry vertical page. Each quote gets its own H2 heading, its own surrounding context, its own metadata. The retrieval engine indexes each as a distinct passage with its own embedding.
A 12-quote wall on /testimonials gets retrieved in roughly 2 to 4% of relevant queries. The same 12 quotes spread across 12 contextual pages get retrieved in 25 to 35% of relevant queries. The math is not subtle.
3. Bury the quantitative outcome in the quote itself
The quote is the citation surface. If the number is outside the quote, the LLM may drop the number when it summarizes. If the number is inside the quote, it is preserved as part of the verbatim reproduction.
Bad: “We saw great results.” (Customer Name). Surrounding sentence: “Acme reduced costs by 38%.”
Good: “We cut our infrastructure spend by 38% in the first quarter and the team has not had to re-architect anything.” (Customer Name).
The number inside the quote travels with the quote. The number outside the quote gets paraphrased into “significant reduction” or dropped entirely. In a Claude 3.5 retrieval test I ran in late April, quotes with embedded numbers retained the exact figure in 94% of citations. Quotes with numbers in adjacent prose retained the exact figure in 31% of citations. The drop-off is the whole game.
4. Match the language a buyer would actually type into ChatGPT
The query “what do customers say about” is the most common AI-search prompt for any B2B brand. The answer engine is matching your testimonial text against a buyer-intent query that is almost always phrased in plain English with specific industry vocabulary.
If your testimonials are full of internal product names (“the workflow engine,” “the orchestration layer,” “the multi-tenant module”), they will not match a buyer query. The buyer is searching for “automated customer support” or “AI agents for ecommerce” or “support deflection.” The testimonial needs to use the buyer’s words, not your product team’s words.
Audit your existing testimonials against a search-term list. Find the gap between what the customer said and what the buyer types. Rewrite the testimonials with the customer’s verbal permission to swap internal product language for the buyer’s actual category vocabulary. Most customers will sign off, especially if you explain it as a search visibility issue.
5. Stamp it with date, role, and industry context
A testimonial without a date looks suspicious to both humans and retrieval models. LLMs increasingly downrank quotes that lack temporal context, because they cannot tell whether the quote describes the product as it exists today or as it existed three years ago. A 2022 testimonial about a 2026 product is a misinformation risk, and the answer engines are getting better at filtering it.
Add a date, the customer’s role (with specificity, not just “Manager”), and one sentence of industry context. The pattern is:
“Quote with embedded number.” (Maria Sosa, VP Analytics, Acme Corp; B2B fintech, 800 employees; Q2 2026.)
That single line carries five attributes the retrieval engine can use: the verbatim quote, the named human, the title (which signals authority), the industry vertical (which signals relevance), and the temporal stamp (which signals recency). Five signals beats one signal every time.
6. Get the quote off your own domain too
Self-published testimonials carry less retrieval weight than third-party-published testimonials. This is true for Google and even more true for AI search. The reason is simple: the LLM is trained on a corpus where third-party publications carry editorial weight. A G2 review, a Capterra entry, a Reddit thread where your customer says something flattering, a TrustRadius write-up, a podcast clip where a customer talks about you, a published case study on the customer’s own site. All of these are weighted higher than a quote on your /customers page.
The work is heavier than copy-pasting a quote into a website. You need to ask the customer to leave a G2 review. You need to broker a guest appearance on a podcast in your category. You need to nudge the customer’s marketing team into publishing a co-branded case study. Each piece of third-party publishing compounds.
Here is the test. Run this query in Perplexity right now: “what is the best [your category] software.” Look at where the citations come from. Count the G2 citations, the Reddit citations, the Capterra citations, the third-party blog citations, the TrustRadius citations. Compare that to how often your own .com is cited. The asymmetry is the work you have not done yet.
7. Make the page schema explicit
Schema.org has a Review type and a Testimonial structured-data pattern. Most sites do not implement them, or implement them wrong. The retrieval engines parse structured data when it is present and clean.
The minimum viable structured data for a testimonial is JSON-LD with @type Review, with the named author as a Person object (name, jobTitle, worksFor), with the reviewBody as the quote text, with itemReviewed pointing at your product or service, and with datePublished as ISO 8601. Five fields. Twelve lines of JSON. It takes a developer 20 minutes per template.
Sites with proper Review schema get cited in AI answers at roughly 2x the rate of sites without it, holding all other factors constant. The schema is not magic. It is just legible. When the retrieval engine has to choose between a quote it has to infer the structure of and a quote where the structure is declared explicitly in JSON, it picks the declared one. Always.
The deeper point is that testimonials for AI search are no longer a vanity asset. They are the citation surface your buyer never sees but always touches. Every time someone asks an LLM about your category, the model is picking which voices to reproduce. If your customers’ voices are not formatted to be reproducible, they are not in the conversation.
So which of your seven highest-revenue customers can you get on a G2 review, a podcast appearance, and a co-branded case study in the next 30 days, with a verbatim quote and a number embedded in it?