How to Write Case Studies AI Engines Cite
A case study that AI can cite states the result in the first sentence, then gives the starting point, what was done, the measured outcome and the time it took. Use real numbers with their source, get written permission to name the client, and keep the structure the same every time. Vague praise gives an assistant nothing to quote.
- Lead with the result and the time frame. A headline like 'Great partnership' cannot be quoted.
- Show before, what you did, after, and how you measured it. State what you do not know.
- Get written permission to name a client. Where a result is unusual, say what is typical.
- Use one fixed layout so every case study is easy to find, compare and quote.
Why do case studies matter for AI search?
Case studies are first-hand evidence, which is something competitors cannot easily copy. Google asks whether content offers original information or analysis, and recommends non-commodity content built on unique expertise or first-hand experience 24. A real project is exactly that.
They also answer the questions buyers ask most near a decision: does this work, for whom, and how long does it take. Assistants answering those questions look for specifics. Ahrefs found that for comparison questions such as 'A vs B', only 46.4% of ChatGPT citations went to official websites, so most went to other sites 3. That suggests assistants lean on other sites for comparison questions. A case study will not change that by itself, but it gives other sites and buyers something specific to cite.
That study covers one model at one point in time. It shows where citations went, not why. Treat it as a hint that proof beats claims.
What do you need before you write one?
You need a finished project with a measured result, the numbers behind it, a person who can confirm the facts, and written permission to publish. In our experience, writing takes two to three hours per page once the facts are in hand. Getting approval is usually the slow part.
If you have no measured result yet, do not publish a case study. Publish a project note that says what you did and what you will measure. A case study without an outcome is a brochure.
What structure should every case study follow?
Use the same layout every time: result summary, background, problem, what was done, results, what was not measured, and what to do next. Same layout means a reader and a machine know where to find each fact. Microsoft advises headings that act as chapter titles, defining clear slices of content 5.
Open with a two-sentence summary that holds the client type, the problem, the result and the period. That summary is the part an assistant is most likely to lift, so it must make sense on its own.
| Part | What to write | Quotable fact it supplies |
|---|---|---|
| Summary | Client type, problem, result, time period | 'X went from A to B in N months' |
| Before | The starting numbers and situation | A baseline to compare with |
| What we did | Three to five actions, in order | A repeatable method |
| Results | Numbers, source of each number, date range | The measured outcome |
| Limits | What was not measured, other causes, what is not typical | An honest boundary |
| Lessons | What another team could copy | A decision rule |
How specific do the numbers need to be?
Specific enough that someone could check them. Give the metric, the starting value, the end value, the dates and where the data came from, for example Search Console clicks between two dates. 'Traffic grew a lot' is unusable. 'Search Console clicks rose from 1,200 to 4,800 between March and August' (an invented example) can be checked and quoted.
Show the measurement source next to each number. If a result is a ranking, say which query, which country and which date. If it is revenue, say whether it is tracked or self-reported.
Never round up, borrow numbers from another project, or present a modelled figure as measured. If the data is partial, say so in the Limits section.
- SUMMARY: A local trophy shop in Bengaluru, India went from near-zero search visibility to 27,300 Google clicks in 16 months, with page-one rankings for 'trophy shop in bengaluru'.
- BEFORE: A local trophy shop with near-zero search visibility.
- WHAT WE DID: An AI SEO programme built around the high-intent local searches buyers type right before they order.
- RESULTS: 27,300 clicks and 2.3 million search impressions over 16 months; average position 6.9.
- LIMITS: These are Search Console clicks and impressions, not revenue. We do not claim other factors played no part. One local shop is not a forecast for another business.
- LESSON: For a local seller, pages that match late-stage, high-intent searches can matter more than broad topics.
How do you get permission, and what must you disclose?
Get written permission before you name a client, quote them or publish their numbers. Agree the exact text, the logo use, the numbers you may share and whether they can withdraw. Keep the email or signed form.
The US Federal Trade Commission (FTC) says an endorsement must reflect the honest opinion of the endorser, and that a material connection, such as payment or an unexpected relationship, should be disclosed clearly 1. It also says that where an ad shows exceptional results and you lack proof they are typical, you must make clear what results are generally expected 1. Those rules are about advertising in the United States. If you sell elsewhere, check local rules.
In practice: do not offer clients a reward for a good story without saying so, and never publish the best result as if it were the norm. If your case is a best case, write that.
What if the client will not let you name them?
Publish an anonymous case study, but describe the client with enough detail to be useful, such as industry, size band and market. Do not name them, hint at a name, or give details that make them easy to guess. Keep the numbers real.
Anonymous cases carry less weight, because nobody can check them. Say so in the page: 'Client name withheld at their request.' Prefer a named case with fewer numbers over an anonymous one with impressive ones, if the client allows only that.
How do you make the page easy for search engines and AI to read?
Put the text in plain HTML, give the page a clear title that says who, what and result, and show author and dates. Google says that for Article markup you should include every author shown on the page, each in its own field, and mark datePublished and dateModified in ISO 8601 format 7. Markup must describe content visible on the page 6.
Schema is optional. Google says structured data is not required for its generative AI features 4. What matters more is that the page is crawlable and that the key facts are written as sentences. Our guide on writing answer-first intros shows how to shape the summary.
Link each case study to the service or method it proves, and link back from those pages. See internal linking for AI search.
What mistakes make case studies useless?
The common mistakes are praise with no numbers, results with no dates, hidden caveats and hiding the case study behind a form. A gated PDF is harder to quote or read.
Another mistake is a page that is all story and no method. Readers and assistants want the steps, so someone else can judge whether the method fits their own case.
Frequently asked questions
How long should a case study be?
As long as it takes to give the summary, the numbers, the method and the limits, in our experience 600 to 1,000 words. Length itself does not help. Google says there is no ideal length 4. Cut anything that does not add a fact a reader can check.
Do case studies need schema markup?
Can I use a case study with no client name?
Yes, but it carries less weight, because no one can check it. Describe the client by industry, size and market, keep the numbers real, and state that the name is withheld at their request. Never invent or blend clients.
What if my results are unusually good?
Say so. The FTC says that where an ad shows exceptional results and you cannot prove they are typical, you need to make clear what results are generally expected 1. Add a Limits section and state what a normal outcome looks like.
How many case studies do I need?
Start with three good ones across different client types, rather than ten thin ones. Quality and honesty matter more than count. Add one when you have a new measured result and permission to publish it.
Sources
- FTC's Endorsement Guides: What People Are Asking · US Federal Trade CommissionOpened October 20, 2026. Supports: Endorsements must reflect honest opinion; material connections must be disclosed clearly; exceptional results require stating generally expected results. The Guides were revised in 2023.
- Creating helpful, reliable, people-first content · Google Search CentralOpened October 20, 2026. Supports: Original information, reporting, research or analysis; expertise; Who, How and Why. Page showed last updated 2026-10-05.
- How ChatGPT Does Query Fan-Out (15 Million Queries Studied) · Ahrefs Blog, 12 October 2026Opened October 20, 2026. Supports: For comparison questions, 46.4% of citations went to official websites, so most went to other sites; for factual questions, 63.4% went to official sites.
- Optimizing your website for generative AI features on Google Search · Google Search CentralOpened October 20, 2026. Supports: Non-commodity content with unique expertise or first-hand experience; structured data not required; no ideal page length. Page showed last updated 2026-07-10.
- Optimizing Your Content for Inclusion in AI Search Answers · Microsoft Advertising (Bing)Opened October 20, 2026. Supports: Headings act like chapter titles that define clear content slices.
- General structured data guidelines · Google Search CentralOpened October 20, 2026. Supports: Do not mark up content not visible to readers; markup must be a true representation of the page.
- Article structured data · Google Search CentralOpened October 20, 2026. Supports: Include all authors shown on the page, each in its own author field; only the name in author.name; datePublished and dateModified in ISO 8601. Page showed last updated 2026-09-08.