Case Study Pages AI Cannot Quote a Word From

Open the case study page on your website and it probably reads like this: a customer in some country, buying product X, annual volume Y, extremely satisfied with quality and service.

Thirty case studies, thirty near-identical paragraphs.

Then ask Doubao which manufacturers in China have supplied European brands with product X. AI names three. Not you.

Case pages fail mostly not from having too few stories, but from being written as praise with no facts. "The customer was very satisfied" is an evaluation, and it matches no specific query. What AI extracts is a unit of information that meets a set of conditions, and not one of those sentences qualifies. This is the hardest layer of GEO (Generative Engine Optimization): it does not care how moving the copy is, only whether you match the conditions.

The boundary first. Not every business should put the case page first: for standard parts sold in volume, where buyers order from a photo and a price, the case page does little. Rewriting it will not make AI mention you tomorrow either. What it affects is something else: when a buyer asks who you have worked for and how far you took it, whether your site has anything worth extracting. Everything below assumes B2B, contract manufacturing or project work.

What AI looks for when asked who you have worked for

Google's documentation describes query fan-out. When you ask a question, AI does not search your wording; it splits the question into several sub-queries, retrieves each, then combines the results into an answer.

"Which manufacturers in China have supplied European brands with product X" might become: a list of domestic producers of X, manufacturers exporting to Europe, certification status, production capacity. Each sub-query is really an intersection of several conditions.

After retrieval there is a second gate: AI extracts sentences from the pages it found. The preprint arXiv:2410.22349 measured citation completeness at 20–25% (not peer-reviewed; conclusion bounded by the query set tested). Even when a page is cited, often only a small part of it reaches the answer. So in generative engine optimization, every paragraph is better off standing on its own — a description that makes no sense out of context is unlikely to be extracted.

One set of figures often misread. Semrush analysed more than 540,000 AI answers (vendor-published, not independently verified): when a brand is mentioned in the answer, its content is cited 53.1% of the time; when the brand is absent, 10.6%. A further 62% of citations never produce a brand mention at all. Do not read it as causation, but it settles one thing — being cited and being named can happen separately. A case page happens to supply both at once: project detail that can be cited, and an industry identity that can be named.

Doubao, DeepSeek and ChatGPT draw on different source mixes, but they agree on specificity: your page has to match a country, a category, an order of magnitude before citation is even in question. That is why "the customer was very satisfied" yields nothing on any of them. Those are exactly the sentences GEO has to change.

Four ways of writing that give AI nothing to quote

All four before-and-after pairs below are illustrations of structure, not real customer cases; they demonstrate sentence form only. From a GEO standpoint all four share one fault: praise without facts.

Nothing but adjectives

Before: "The customer is extremely satisfied with product quality and service, and has indicated continued cooperation."

After: "Continuous supply since 2023, 42 batches to date. Two batches arrived with plating oxidation from sea-freight moisture; we replaced them within 5 working days and changed the packaging to double-layer moisture barrier film."

The first sentence contains nothing checkable. The second carries timing, batch volume, a specific problem, the action taken and how long it took — each item answers a sub-query on its own.

The customer becomes "a well-known brand"

Before: "Long-term supplier to a well-known appliance brand."

After: "Supplying a European appliance brand, with the part used in the water inlet assembly of its commercial dishwashers, at an annual volume in the 300,000-piece range."

"Well-known" says nothing: no country, no category, no order of magnitude — all three conditions miss. Anonymising is not the same as deleting every detail; that gets its own section below.

Only the result, never the process

Before: "Helped the customer cut costs by 30%."

After: "The original design used imported seals: high unit cost, 45-day lead time. Switching to a domestic substitute lowered the unit cost and cut lead time to 20 days, at the price of a maximum service temperature dropping from 180°C to 150°C — suitable for ambient-temperature duty only."

With only the result, neither the reader nor AI knows where the 30% came from or when it applies. Naming the trade-off and the limits is what makes it credible — nobody invents drawbacks for their own solution.

The numbers do not reconcile across the site

The case page claims 500,000 pieces a year while the product page states monthly capacity of 30,000; the case page dates a new line to 2024 while the About page says 2022. This is common, because different pages were written by different people at different times.

Faced with conflicting numbers, AI usually takes the conservative one or drops the figure entirely. So reconcile the site before rewriting the case page: capacity, plant area, founding year, certification validity, headcount. In generative engine optimization and AI search optimization alike, this is the dullest step with the most reliable payoff.

Break "satisfied" into six fields

The most useful move in GEO is breaking a case into checkable fields. A case needs at least six to be extractable; missing three or four and there is nothing worth taking.

Industry and application: which industry the customer is in, and where the product sits in their process. The same pump in a wastewater plant and on a food line tells two entirely different stories.

Country or region: the first thing a buyer checks for export experience, and one of the most frequent qualifiers in sub-queries.

Specification and model: material, bore size, power rating, certification level. The most frequently omitted field, and precisely what peers and engineers search for.

Volume and frequency: annual usage, minimum batch, replenishment rhythm. The order of magnitude tells the buyer whether you are a supplier of comparable scale.

Lead time: how many days for standard orders, how few for urgent ones. This field carries heavy weight in a buyer's decision.

Problems encountered and how they were handled: what went wrong, how it was fixed, how long it took. The hardest field to write, and the one that separates companies most.

A case with all six runs about 150 to 250 words. Ten of them is roughly 2,000 words — far more useful than thirty paragraphs of "very satisfied". This is also where GEO parts company with copywriting: copy aims to read well, GEO aims to extract whole.

Anonymising is not the same as deleting what matters

Contract customers refusing to be named is normal, not an obstacle. What can be removed and what cannot needs separating.

Can be removed: the customer's full name, contacts, contract pricing, drawings and formulations, project codes.

Cannot be removed: industry, country or region, specification, timing, order of magnitude, the problem and how it was handled.

"A well-known brand" is useless not because it was anonymised but because nothing survived the anonymising. Replace it with "a European appliance brand, used in the water inlet assembly of commercial dishwashers, 300,000 pieces a year" — not a word gives the identity away, yet all three search conditions hit. If even the industry cannot be named, use customer type plus region, such as "an OEM of commercial kitchen equipment in northern Europe", which still carries far more than "a well-known enterprise".

One more detail: do not alter the numbers while anonymising. An order of magnitude can be a range ("in the 300,000 range"), but 300,000 must not become 500,000 and 2023 must not become 2024. Once a case page carries an unverifiable figure, its value to AI does not decline — it drops to zero, with no exception in GEO or AI search optimization.

A case page is not a testimonial page, and neither is a review platform

In GEO discussions these three get mixed up constantly. Each does a different job.

The case page: what you have done. A project record, on your own site, fully under your control. It answers: can you do it?

Customer testimonials: how customers rate you. Also on your site, but essentially reported speech. They answer: is anyone satisfied?

Third-party review platforms: what others say where you have no control. They answer: is it credible?

Third-party material is the most credible, and there is data behind that — read with care. Trustpilot's May 2026 figures show citation rates of 1% for brands with no review profile, 53.5% with a profile, and 75.3% with more than 80 active reviews. This is platform-published data; the three figures are differences between groups, not cause and effect.

NP Digital interviewed 100 marketers in June 2026 and concluded that third-party citation is the only signal ranked highly on all six platforms, scored 4.5–4.8. AirOps' "State of AI Search" 2026 states that 85% of brand mentions in AI answers come from third-party pages. Both come from organisations with a commercial interest and can only be read as observations.

To conclude: third-party material is the ceiling; the case page is the foundation. Without the foundation there is nothing for third-party reviews to sit on — nobody leaves a detailed project review on a site that cannot explain what it has built. The order in AI search optimization is own content first, external second; the same holds for GEO.

Three common questions

Our contract customers forbid naming them. Can we still publish cases?

Yes, and you must. Drop the name, keep the other six fields. Industry, country, specification, magnitude, lead time and problem handling do not identify anyone; what actually needs permission is the customer's name and trademark. If even the industry is off limits, write "an OEM in a certain region" — but do not delete the volume and the dates along with it.

Is three to five cases enough?

Enough. Three cases written in full beat thirty that say only "very satisfied". The test is simple: read one case on its own and ask whether it answers three questions — for whom, how far, and what went wrong along the way. Answer all three and it passes. This is also the highest-return rewrite in GEO: low cost, no new data required.

Should old cases be cleared out?

Yes, but not by deleting everything. A case loses value over time: specifications, capacity, certifications and equipment go stale within a few years. Handle it by sorting — where the current state can be confirmed, update the timestamp and capacity; where a product is discontinued or the customer has left, rewrite it as a record of experience or take it down.

The worst option is leaving everything untouched. A pile of cases with unclear dates and figures that do not reconcile is not an asset to AI; it is noise. That is why, with the same effort at AI search optimization, some case pages get cited and others go a decade without a mention.

Where this leaves you

The question in the title — a case page that says only "the customer was satisfied", so why does AI extract nothing — is not answered by writing technique. The purpose of the page was mistaken. It should be evidence a buyer uses to vet a supplier, and it was written as a letter of praise. A letter of praise is for people who already believe you; evidence is for people still deciding. AI stands with the second group.

So start with one small thing: open the case page, pick one case, delete "extremely satisfied" and replace it with the six fields — industry, country, specification, volume, lead time, what went wrong. When you finish you will notice you remembered all of it; you had simply never written it down.

What GEO and AI search optimization actually change is this layer: moving facts out of your head and out of a salesperson's chat history onto the page, written as a sentence someone can extract. It requires no budget and no tools.

Yihang Network (Shenzhen) previously worked in cross-border B2B and Alibaba.com-related services, and has been extending into GEO since 2024. We read a client's case pages in the same order: first count the checkable fields, then look at how nicely it is written.

Data and effect boundaries: the citation-completeness figure (arXiv:2410.22349, a preprint, not peer-reviewed) is academic research bounded by the query set it tested; the Semrush figures of 53.1%, 10.6% and 62% are vendor-published observational statistics, not independently verified, and support no causal claim; Trustpilot's 1%, 53.5% and 75.3% are platform-published data reflecting differences between groups, not causation; NP Digital's six-platform signal scores (June 2026, 100 marketers) are practitioners' subjective assessment; the 85% in AirOps' "State of AI Search" 2026 is a vendor-published observational statistic. All four before-and-after pairs are illustrations of structure and involve no real customer. None of the above represents the results of any specific company. The methods described raise the probability of being retrieved and cited; they do not guarantee that a company will appear in any AI answer, and they promise no timeline or business result.

Sources

1. Google Search Central: "AI features and your website" (query fan-out; no special schema or markup required for AI features)
2. arXiv:2410.22349 (citation completeness of AI answers against source pages, 20–25%; preprint, not peer-reviewed)
3. Semrush: "ghost citations" study across more than 540,000 AI answers (53.1% citation rate when the brand is mentioned, 10.6% when absent; vendor-published)
4. Trustpilot: review profiles and AI citation figures, 12 May 2026 (1% / 53.5% / 75.3%; group differences, not causation)
5. NP Digital: assessment of AI search signals by marketers, June 2026, 100 interviews (six-platform signal scores 4.5–4.8)
6. AirOps: "State of AI Search" 2026 (85% of brand mentions from third-party pages; vendor-published)