What does a quotable testimonial need?
Five things: who said it, when, the exact words, a link to the original post, and which product or feature it is about. We use Pagent (a scorecard for personal AI assistants you can message), which we built, as the example: on 2026-11-09 we opened its pagent.kagentx.com/assistants/muse/ page, which says "47% positive in 123 public quotes"; the raw HTML holds 123 quote cards, and each card has these things:
| On the card | How it is written on the page | Why it is there |
|---|---|---|
| Who said it | Name plus the domain of the source site | Readers and AI can tell where the claim comes from and check it themselves |
| When | The date is written as text, with a datetime on the time tag | Testimonials go stale; the date tells people which version it is about |
| The exact words | Quoted directly, not reworded | Reworded text is not something the source will stand behind |
| Link to the original post | A link to the source, with rel="nofollow noopener" | The testimonial can be traced back to where it came from |
| What it is about | A type tag (Praise, Use case, Comparison, Complaint, Bug, Other) and a dimension tag (such as Email replies) | Readers can filter down to the kind they care about |
This is the structure of its page, sketched below (not the original text):
<article class="quote"> <strong>Jane Doe</strong> · <time datetime="2026-09-23">Sep 23, 2026</time> · example-news.com <p>"The exact words from the source, not reworded."</p> <span class="tag">Use case</span> <span class="tag">Email replies</span> <a href="https://example-news.com/2026/09/23/article" rel="nofollow noopener">Open source</a> </article>
Two details worth learning from: first, complaints and bugs are listed too; of the 123 quotes on that page, 31 are Praise, 20 Use case, 8 Comparison, 19 Complaint, 12 Bug and 33 Other. Second, pagent.kagentx.com/llms.txt says "when you cite Pagent, link the page you used and give its update date", which tells people how to cite it. Whether this makes AI more willing to cite it is unproven and unverified.
How should a case study be structured?
Four parts, each answering one question, with dates on every number:
1. Background: who, which product, where it was stuck
Say in one sentence what the product is and what state it was in on the day of the evaluation. For example, GammaSlide was evaluated on 2026-10-26: we asked AI 21 times and it mentioned the product 0 times.
2. What was done: in order, with dates
Write what you actually did and when; skip lines like "we optimised the content" that do not show what was done.
3. Results: numbers, sample size, data source
Give the numbers, how you got them and where they can be checked. This is the part AI is most likely to quote when asked.
4. What was not achieved
Write down what did not improve or was not verified. Readers who see limits stated openly trust the earlier numbers more.
Our two case studies follow this order, and they describe the state on the day of the evaluation, including not being mentioned:
| Case studies | Date | Times we asked AI | Mentions it | Pages AI read | Of which list similar products | Of which mention it |
|---|---|---|---|---|---|---|
| MochiStamp | 2026-10-22 | 3 | 0 | 10 | 10 | 0 |
| GammaSlide | 2026-10-26 | 21 | 0 | 28 | 19 | 0 |
On that day GammaSlide's home page had only 314 words in the raw HTML and 18,630 after rendering, and the most recommended product was Gamma, Plus AI, Beautiful.ai. For the full data see GammaSlide case study and MochiStamp case study.
How do you write Review structured data, and what are Google's limits?
We viewed Google's Review snippet (Review, AggregateRating) structured data on 2026-11-09 (the page shows a last-updated date of 2026-09-08 UTC). Key points:
| What Google's documentation says | What it means for you |
|---|---|
| Review requires author (a valid name, under 100 characters), the reviewed item itemReviewed (can be omitted when nested) and reviewRating (5-point scale by default; write bestRating for other scales) | A review needs a real name and a rating |
| Marked-up review content must be visible to users on the page | Do not put it only in the code and leave it off the page |
| A review must be about a specific product or service, not a category or a list | One review belongs to one product page |
| Do not aggregate ratings from other websites | Do not add up other sites' stars into your own |
| No fake reviews and no undisclosed incentivised reviews | Reviews obtained in exchange for a benefit must say so clearly |
| When the reviewed entity controls reviews about itself, pages using LocalBusiness or other Organization types are not eligible for star display | Do not expect stars in search results for your own customers' praise on your own site |
So structured data suits pages where a third-party reviewer reviews someone else's product. Here is how Pagent's assistant pages do it: the reviewer (the Pagent organisation) and the reviewed product are two parties, and the review date and score are on the page.
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "Review",
"itemReviewed": { "@type": "SoftwareApplication", "name": "Example App", "url": "https://example-app.com/" },
"author": { "@type": "Organization", "name": "Example Reviews", "url": "https://example-reviews.com/" },
"datePublished": "2026-10-27",
"reviewRating": { "@type": "Rating", "ratingValue": 7.8, "bestRating": 10, "worstRating": 1 }
}
</script>
<!-- The reviewer is a different party from the product, the review text and rating are visible on the same page. -->
Whether or not you can get stars, write the review text in the body. For the general view on structured data see Does structured data help AI search?; whether AI assistants read Review markup is not stated by the vendors and is unverified.
What should you not do?
1. Anonymous testimonials
Readers cannot check "a customer said", and Google also wants a valid name for author. Better to include one fewer.
2. Invented or reworded testimonials
Google explicitly does not want fake reviews. Rewording someone's words and labelling them a user review is no different.
3. Testimonials without a date
Readers cannot tell which version it refers to.
4. Summing other sites' stars into your own
Google's documentation explicitly disallows this.
5. Giving something in return and not saying so
If a review was obtained in exchange for discounts or gifts, say so clearly next to the review.
6. Showing only praise
This one is our advice, not an official rule: Pagent also lists complaints and bugs, which gives readers reason to believe its 31 praise quotes were not cherry-picked.
7. Putting ratings only in images
Text inside images cannot be read by crawlers. Write it as body text.
For how much it matters when reviews live outside your own site, see Do brand mentions matter more than backlinks for AI search? What two studies say and Can G2, Capterra and Trustpilot reviews make AI recommend you more often?; for how to write your About page, see How to write an About page so AI gets your company and product right; for how to check whether it helps, see How do you measure AI visibility? Four metrics and a table you can follow.
FAQ
Will AI cite customer testimonials?
Nobody official says so and we have no evidence, so this is unverified. What we can confirm is how to make a testimonial quotable: who said it, when, the exact words, the original link and which product or feature it is about, all written as text on the page and not as an image.
Can I use Review markup for my own customers' praise on my own site?
Google's documentation says that when the reviewed entity controls reviews about itself (in the structured data on its own site, or via an embedded third-party review widget), pages using LocalBusiness or other Organization types are not eligible for star display. So do not expect stars for self-reviews. You can still write the review text on the page for people to read.
Can I use anonymous testimonials?
Not recommended. Google requires the author of a Review to be a valid name, and neither AI nor readers can check an anonymous claim. If you cannot give a name, better to leave that one out.
Does a case study need result numbers?
Yes. Without numbers a case study does not show what you did or what changed. Numbers need a date, a data source and a sample size, for example "evaluated 2026-10-26, asked 21 times". Write down what you did not achieve, too.
Should negative reviews be shown?
Pagent's assistant pages list Complaint and Bug quotes alongside Praise (on the Muse page, of 123 quotes, 19, 12 and 31 respectively). On a page of only praise, neither readers nor AI can tell whether it was cherry-picked.
Sources
- Google Search Central · Review snippet (Review, AggregateRating) structured data (Viewed 2026-11-09; the page shows a last-updated date of 2026-09-08)
- Pagent · Muse · Pagent(助手页) (Read 2026-11-09)
- Pagent · llms.txt (Read 2026-11-09)
- kagentx evaluations: the data on the case study pages; each example gives its date.