Blog · Content that gets cited

Writing case studies an engine can quote

Published 14 April 2026 · Updated 12 October 2026 · 8 min read

Writing case studies an engine can quote
Photograph: Annie Spratt / unsplash

The short answer: Write each case study so that one self-contained sentence carries the customer name, the number, the time period and the condition that produced it. Engines retrieve passages, not pages, and a passage that says "a 38 percent reduction in handling time" with no named subject anywhere near it cannot be quoted without the engine inventing the subject. Put the whole claim in one place, repeat it near the top, and name the client.

That is the entire mechanism, and it is the part most often missing from the case studies we read during audits. We have not counted how common the failure is across the web, so treat that as what we keep running into rather than a measured rate.

Why the narrative format fails retrieval

Case studies are usually written as stories. Background, challenge, solution, result. The result sits in the last fifth of the page, often in a pull quote or a styled stat block, while the customer is named in the first paragraph and then referred to as "the client" for the next nine hundred words.

A passage retrieval system does not read it that way. Systems of that kind split the document into passages, score them against the query, and keep a handful. The engines in our prompt set publish neither their chunk size nor their scoring, so we cannot tell you where your page will be cut. What we can tell you is what the cut costs: the passage holding your headline number contains the words "reduction", "handling time" and a percentage, and it does not contain the customer or the year. Pull it out of the page and it is an orphan statistic, with no subject for an engine to attach it to.

Reading retrieved passages during audits points the same way: where a quoted passage carries the claim, it carries the subject with it. We have not scored that across a counted sample, so it is a pattern we notice, not a rate we can give you. It is the same property discussed in what makes a page citable by an AI engine, applied to the one page type where companies have real proprietary facts and still manage to lose them.

The evidence that quotes and figures pull their weight

There is one measured version of this, from outside our own work. The KDD 2024 paper that introduced the term generative engine optimisation tested nine content edits against a 10,000 query benchmark and reported that its three best performers were adding credible quotations, adding relevant statistics, and citing sources, at a relative improvement of 30 to 40 percent on its Position-Adjusted Word Count metric. Keyword stuffing, the tactic carried over from an older era, did not perform well at all. Read the scope with the result: those headline figures were measured in 2024 on a generative engine the authors built for the benchmark, and a separate run against Perplexity put the best method at 22 percent on the same metric. The authors also report that the effect varied by domain. None of that makes the ranking of edits a law about engines in general, and we have not reproduced the experiment ourselves.

A case study is the one asset where those three edits are native rather than bolted on. You already have the quotation, because the customer said something in the interview. You already have the statistic, because that is why the study exists. The source is you, and you can describe the method.

Google's own guidance, last updated in December 2025, is worth reading next to that. It states plainly that no new markup, AI text file or special optimisation is required to appear in AI Overviews or AI Mode, and it describes the query fan-out technique that turns one user question into a wider set of searches. Read together: markup will not buy you a citation, but a claim that answers one of those fanned-out sub-questions in a single retrievable sentence might.

The claim sentence, and where to put it

Build one sentence per study that stands alone. It needs five parts.

The named customer. The direction and size of the change, as a figure. The metric, in the words a buyer would use. The time window over which it happened. The condition or caveat, because an uncaveated number reads as marketing and gets treated as such.

A working example of the shape: "Across the nine months after the migration, Acme Logistics cut average first-response time on support tickets from 14 hours to 4, measured on their own helpdesk export, on a ticket volume that grew 20 percent in the same period."

Put that sentence three times. Once in the page description. Once as the first line of the body, before the background. Once in the results section where it belongs narratively. The repetition is not for the reader, who will skim past it, but for the chunker, which has no idea which slice of your page will be scored.

Then repeat the customer name in the results section, in the FAQ, and in any pull quote. Pronouns are cheap to a reader and expensive to a passage.

Five failure modes we keep finding

The anonymised client. "A leading European fintech" gives an engine nothing to resolve against, so there is no named entity for a claim to be corroborated against. We cannot see what the engines do with an unattributable claim, only that the page supplies no subject. If legal will not let you name them, treat the study as weak retrieval evidence and spend the budget elsewhere.

The number inside an image. A stat block baked into a background image carries no text to extract at all, and SVG text survives some extraction paths and not others. Check the rendered HTML, not the design file.

The gated PDF. A study behind a form is a study a crawler is not logged in to fetch. Publish the page, gate the appendix if you must.

The number without a method. "Revenue up 3x" with no baseline, no window and no measurement source is the kind of claim a cautious model will skip rather than quote.

The results page with no date. Google documents that AI Overviews and AI Mode may fan one question out into multiple related searches, and in the prompt runs we read, those sub-questions often carry a year. We cannot see how the engines weight recency. We can see that an undated 2022 study gives a year-bearing sub-question nothing to match on, and a dated 2026 one does.

The same discipline makes a comparison page work, for the same reason, and the two page types support each other when you cross-link them. See comparison pages that get cited for the other half of that pattern.

A check you can run in ten minutes

Fetch your own case study as a crawler would, strip the tags, and slide a window across the text. Any window that contains your headline figure must also contain the customer name.

`` curl -s example.com/case-studies/acme | \ sed -e 's/<[^>]*>/ /g' -e 's/ */ /g' | \ fold -w 1200 -s | grep -n '38%' ``

Read the matching blocks. If the customer is not in them, your number is an orphan, and markup will not put the subject back into the passage. Fix the prose first.

For the markup, a Quotation node from schema.org is the right type for the customer testimonial, with text, creator and isBasedOn filled in, nested inside the Article. If you publish the underlying numbers, Dataset markup is defined for it, and Google's Dataset documentation lists name and description as required properties, with the description between 50 and 5000 characters. None of this creates a citation on its own. It makes the relationship between quote, speaker and claim explicit for anything parsing the page rather than reading it.

The sample audit report shows how we record this kind of finding, with the retrieved passage attached, so an argument about whether a page is quotable can be settled by looking at the passage instead of the page.

What this does not do

Writing a quotable case study does not make an engine cite it. Retrieval still depends on whether the page is reachable, whether your domain is one the engine already pulls from, and whether anyone asks a question your study answers. We have seen well-built studies sit uncited for months because nobody searches for the problem they solve.

We also cannot tell you how many citations a fixed study will earn, because the engines change their retrieval behaviour without notice and two identical prompts can return different sources. If your category is small enough that nobody is asking an assistant about it at all, this is premature, and when answer engine optimisation is not worth doing yet is the more useful page.

What the work does reliably is remove a self-inflicted failure. A study that cannot be quoted will not be quoted. That is one of the few things in this field you control completely, and it costs an afternoon per page.

Honest limits

This post does not establish that the engines chunk pages the way described, only that passage retrieval is the published design of the systems involved and that an orphan number survives it badly. The audit observations here are not counted: we have no sample size or date behind "most case studies we read", and we have not run a controlled before and after on a fixed study. The 30 to 40 percent figures come from one 2024 benchmark, largely on an engine the authors built, and we have neither reproduced them nor tested them on a case study page. We cannot see inside the engines' indexes, so where this post says a gated or anonymised study fails, it is describing what the page supplies, not what an engine was observed to reject.

FAQ

How long should a quotable case study be? Length matters less than density. Six hundred words with a complete claim sentence near the top beats two thousand words of narrative with the number at the end.

Should I use the customer name in every section? In the opening, the claim sentence, the results section and any quote, yes. Reusing the name costs a reader nothing and keeps each passage self-contained.

Does structured data make a case study more likely to be cited? Google states that no special markup is needed to appear in its AI features. Markup helps anything parsing your page resolve who said what, which is a different and smaller benefit than a citation.

What if legal will not let us name the client? Then the study is weak evidence for an engine as well as for a buyer. Consider publishing an aggregate instead: a named method, a sample size, and a range across clients, which you can describe honestly without naming anyone.

Is this different from ordinary content work? It overlaps with the structural part of what is answer engine optimisation, and differs in that case studies carry facts no other site has, which makes them worth more per page than general articles.

Sources

Related reading

What makes a page citable by an AI engineComparison pages that get citedWhat is answer engine optimisationWhen answer engine optimisation is not worth doing yetChunking: how your page is cut up before it is readPruning thin content to raise AI visibility