Customer Proof as a Discovery Asset: Getting Case Studies Into AI Answers

Customer evidence used to be late-stage persuasion. In AI-mediated discovery it also has to work early, which means it has to be on an ungated page, written so one passage survives being lifted out, and specific about sector, size and constraints.

A case study earns its place in an AI answer when three things are true: it sits on a public HTML page that no form stands in front of, one passage in it states the situation and the result well enough to be quoted on its own, and it names the sector, the size of the business and the constraint it was working under. Most case studies fail all three, because they were built for a different job: persuading a buyer who had already chosen to talk to you.

That job has not gone away. What has changed is that the same material now has a second one, earlier in the process, and the format that suits the first job actively prevents the second.

What the research actually says

Two pieces of work are worth reading on this, and both need reading carefully.

LinkedIn’s Buyability research surveyed 750 senior B2B buyers. It reported that peer recommendations, from colleagues, customers or companies like the buyer’s own, were around three times more influential in moving a vendor from the shortlist to being bought than a cheaper price or a claim of better performance. It also found that what the study calls defensibility, the buyer’s ability to justify the decision afterwards if it goes wrong, ranked at the top of the emotional dimensions it measured.

Separately, WARC, LIONS Advisory and LinkedIn Marketing Solutions analysed 700 award-entered B2B campaigns from 2010 to 2025, drawn from more than 30 awards programmes. Each case was coded for seven signals: customer proof, expert or peer validation, celebrity or influencer validation, long-term relationships, category leadership, customer insight and brand working style. Campaigns carrying three or more signals were classed High Buyability, those with none to two Low. The High group was 1.91 times as likely to report improvements in brand health metrics and 2.09 times as likely to report incremental revenue growth.

The same analysis found the average campaign carried only 1.6 of the seven signals, with 22 per cent carrying none and 30 per cent carrying one.

The caveat belongs in the finding, not the footnote

The guide itself states that the findings show associations, not causation. Two further limits matter before anyone reorganises a content plan around this. Campaigns entered for awards are not a random sample of campaigns; they are the ones their agencies thought were worth entering. And the outcomes are reported by the entrants rather than measured independently.

So the honest reading is narrower than the headline: among campaigns their own teams considered good enough to enter for awards, the ones that reported better results tended to carry more of these signals. That is useful. It is not evidence that adding a testimonial produces revenue.

One more note on sourcing. A summary of this research published on LinkedIn’s own marketing blog organises the signals under a different framework and gives a specific split for customer testimonials, use cases and recommendations appearing in 64 per cent of High Buyability campaigns against 16 per cent of Low. We could not find that split in the published guide, so we have not used it here. Where our own numbers come from is listed at the end.

Why discovery changed the job of a case study

Most AI assistants that show source links work by retrieval-augmented generation: the assistant turns the question into one or more searches, retrieves candidate pages, selects passages from them, and writes an answer from those passages. The unit that competes is the passage, not the page and certainly not the PDF.

The second change is what buyers put into the question. A buyer can now open with their own situation in full: the sector, the headcount, the systems already in place, the budget, the thing that went wrong last time. We call the conditions a buyer leaves unstated but still expects an answer to respect latent constraints, and when the buyer states them outright they stop being latent and start being filters.

Put those two together and the mechanism behind the research becomes concrete. The reason a relatable customer story helps in AI-mediated discovery is not that recommendation is persuasive, although it is. It is that a story naming a sector, a size and a constraint gives the retrieval step something to match against a buyer who named the same three things. A page that says “we help businesses grow” matches nothing in particular.

The three reasons customer evidence stays invisible

What is in the way What the crawler or assistant gets What to do instead
The case study is behind a form The form. Crawlers do not fill in forms, so the story behind it is not in any index and cannot be retrieved for any query. Publish the story on an open page. Gate the extras if you must: the detailed figures, the template, the recording.
It exists only as a PDF A file that may well be indexed, but which carries no internal links, no heading structure a retriever can use, and no stable section anchors. Write the case study as an HTML page. Offer the PDF as a download of that page rather than as the only version of it.
It is written as a narrative with the result at the end A passage that is on topic but does not answer anything until its final paragraph. State the situation and the outcome near the top, then tell the story underneath.

The third one is the least obvious and the most common. Many retrieval systems re-rank a shortlist with a model that reads the question and a passage together and scores how directly the passage answers it. A passage that builds to its point loses to a shorter one that leads with it, even when the longer passage is better writing.

What a retrievable case study contains

The test we apply is quotability: lift any single passage out of the page and ask whether it is still correct, complete and attributable. It is our own checklist rather than anything a search engine publishes, but it predicts what gets used.

In practice that means each case study carries these, in roughly this order.

  • The situation in the first two sentences, with the sector, the approximate size of the business, the location if it matters, and the constraint. “A 40-bed hospitality property with no website of its own, taking every room booking through online travel agents” is matchable. “A client in the hospitality space” is not.
  • One result sentence that survives on its own, carrying its number, its unit, its time period and its source. A sentence that starts with “this” or “it” cannot be quoted.
  • What you actually did, named specifically, including the stack. Naming the technology is what lets the page be retrieved for a query about that technology.
  • What you deliberately left out, and why. This is the part almost nobody writes, and it is the part that carries information gain, because no competitor’s page contains it.
  • Where the result came from, dated. If the project is too new to have results, say that instead of implying them.

A worked rewrite

The figures below are invented to illustrate the shape of the sentence, not drawn from any project.

Weak passage Quotable passage
“The client was delighted with the results and we saw a significant improvement in their online visibility over the course of the engagement.” “Between March and August 2026, organic enquiries to this 25-person industrial coatings distributor in Delhi NCR rose from 4 to 19 a month, measured in Google Search Console and their call tracking log.”
“We implemented a comprehensive technical SEO strategy alongside a content programme.” “We rebuilt the site on a custom WordPress theme with server-rendered HTML, added Organization and Service schema in a single JSON-LD graph, and published 70 articles over 14 days.”

Neither right-hand sentence is better prose. Both are better evidence, and either one can be lifted into an answer without the reader losing the thread.

Structured data, used honestly

A case study page benefits from JSON-LD that names the organisation, the service the project concerned and the author, connected through stable @id values rather than sitting in separate blocks.

Review and rating markup needs more care. Google’s review snippet documentation excludes what it calls self-serving reviews, meaning reviews about an entity placed on that entity’s own site, from review rich results. Marking up your own testimonials as reviews of yourself will not produce stars, and invalid or misapplied structured data is worse than none, because it undermines the markup that is correct. Mark up what the page genuinely is. Let the ratings live on the platforms that host them.

Evidence on sites you do not control

An assistant weighs the same claim differently depending on where it sits. Your own page saying you are good at something is a claim. A review platform, a directory entry, a trade publication or a customer’s own post saying it is corroboration, and mentions that do not even link to you still put your name next to the topic in the text these systems retrieve and train on.

This is the part of the LinkedIn finding that translates most directly into ordinary work: being named, by real customers, in the places that get cited for your category. The practical route to it is unglamorous and is covered separately in how to get more genuine customer reviews, including the practices that break platform rules and India’s IS 19000:2022 standard.

What to measure

Agree a fixed set of prompts that match how your buyers would describe their own situation, including the sector and size you want to be found for. Run each one several times across the assistants your buyers use, because answers vary between runs. A single run is a sample, not a measurement.

Record four things each time: whether you were named, which URL was cited, whether the description of you was accurate, and which competitor was named instead. The fourth is usually the most useful, because it tells you which page you are losing the passage to.

Traffic alone will understate this. A buyer who reads your customer’s situation inside an AI answer and never clicks has still encountered the evidence, and a page that is working this way may show no visit at all.

How we handle our own

Our case studies name the stack, the trade-offs and the parts we deliberately left out, so that another business can judge whether the approach transfers rather than taking our word for it. Where results exist they are stated with their source and date. Where a project is too new to have results, the case study says so. We do not publish invented figures or describe a client’s numbers without their agreement, which occasionally means a good project goes unwritten.

That last constraint is worth accepting. A case study that a reader can check is the only kind that does the job the research describes, and a page of unverifiable claims about delighted clients is not customer evidence at all.

Frequently asked questions

Should we stop gating case studies entirely?

Gate the extras, not the story. The narrative, the situation and the result belong on an open page so they can be retrieved and quoted. A detailed appendix, a template or a recording is reasonable to ask for an email address for, because the buyer has already seen enough to decide whether it is worth it.

Our clients will not let us name them. Is the case study still worth writing?

Yes, and it loses less than you would expect. What makes a case study matchable is the sector, the size, the location and the constraint, not the brand name. “A 25-person industrial coatings distributor in Delhi NCR” does the retrieval work. Confirm with the client what you may describe before you publish, and keep a written record of it.

How many case studies do we need?

Fewer, more specific ones beat a long list of thin ones. Each case study is competing to be the passage that matches one buyer’s situation, so coverage of distinct situations matters more than volume. Two well-written studies in sectors you want more of are worth more than ten that could describe anyone.

Does putting paid media behind customer stories help with AI visibility?

Not directly. Paid distribution does not make a page more likely to be retrieved, because retrieval works from the index rather than from ad spend. What it can do is produce the second-order effects that matter: people writing about the story, linking to it, and mentioning your name next to the topic. Treat that as the mechanism rather than assuming the spend itself is the signal.

Is this different from ordinary SEO for a case study page?

It overlaps heavily. Clean HTML, a stable URL, server-rendered content and sensible structure serve both. The differences are the emphasis on each passage standing alone, naming the unstated conditions a buyer brings, and treating the PDF as secondary to the page rather than as the deliverable.

Sources

  1. WARC, LIONS Advisory and LinkedIn Marketing Solutions, “Building More Buyable B2B Brands”, for the 700-campaign analysis, the seven signals, the 1.91x and 2.09x figures, the 1.6-signal average and the associations-not-causation caveat
  2. LinkedIn, “Buyability: The Future of B2B Marketing Success”, for the 750-buyer survey, the roughly threefold influence of peer recommendations and the finding on defensibility, as reported in trade coverage of the study
  3. Google Search Central, review snippet structured data documentation, on self-serving reviews
  4. Bureau of Indian Standards, IS 19000:2022 Online Consumer Reviews