You wrote a case study: “We helped a client — they were thrilled, thanked the whole team, the result exceeded expectations.” Beautiful, heartfelt, convincing to a human. A neural network will walk right past it. Another case study: “In 6 months, we increased the client’s sales from 2 to 3.2 million rubles (+60%). Link to report.” Dry, short, boring for a human. The model will pick this fragment for citation. Models don’t need emotions. They need numbers that can be verified and used for comparison.

In this article — how to write case studies for generative optimization, which numbers work, what structure the model loves, and how to turn a case study into a source of citable data.

What the model looks for in a case study

When a model selects a case study to answer a user, it doesn’t evaluate the “beauty” of the text. It looks for data that can be used as evidence or an example. The model looks for specific numbers — measurable results, percentages, amounts, timelines. It looks for verifiable facts — company names, dates, source links. It looks for a clear structure — initial data → actions → result. It expects clear headings, sections, lists. The model values objectivity — if there are both positive and negative results (or method limitations), trust grows. Emotional assessments (“the client was thrilled”) are ignored by the model — they are subjective and cannot be verified. A professional website audit for search engines will help assess the readiness of your case studies.

Numbers the model loves most

Not all numbers are equally useful to the model. Here are the numbers that work best.

Growth percentages. Example: “Increased traffic by 150%,” “Conversion grew by 40%,” “Reduced cost per lead by 25%.” The model can use a percentage as a quick answer to the query “how much can traffic be increased.”

Absolute numbers. Example: “Sales grew from 2 to 5 million rubles per month,” “Number of leads increased from 30 to 80 per month,” “Support response time dropped from 4 hours to 20 minutes.” Absolute numbers plus percentages — that’s maximum information for the model. Percentages — for quick comparison. Absolute numbers — for context.

Timelines. Example: “Result achieved in 4 months,” “First growth noticed after 2 weeks,” “Full implementation cycle — 30 days.” The model can use timelines to answer queries like “how fast will results come,” “how long does implementation take.” Timelines tied to the calendar — “by December 2025” — work better than abstract “in six months.”

Ratios and metrics. Example: “ROI — 320%,” “CPA dropped from 500 to 300 rubles,” “Average check increased from 1,500 to 2,200 rubles.” The model can use these metrics to compare effectiveness. ROI is one of the most cited metrics. Specify the period: “ROI over 6 months,” not just “ROI.”

Quantities. Example: “Processed 1,000 leads,” “Conducted 50 audits,” “Set up advertising in 10 channels.” Numbers that show scale also work. They answer queries like “do you have experience with large projects,” “what’s your volume of work.” Effective geo-optimization of a website also requires numbers in local case studies.

Case study structure for a neural network

Headline — formula: “Client + main result in numbers.” Example: “Increased sales of an online shoe store by 60% in 4 months.” The model sees the client, result, and timeline in the headline. Bad headline: “Case study on promoting an online store” — no numbers, no timeline, no specifics.

Brief summary — 2–3 sentences with key numbers. Place it at the beginning of the case study, before the main text. The model can take this summary as a ready answer, even if it doesn’t read to the end. Example: “Client: online shoe store. In 4 months, increased traffic by 150%, sales by 60% to 5 million rubles per month. ROI — 320% over six months.”

Section “Initial data” — what was before. “Before” numbers are as important as “after.” Without them, the model cannot calculate growth. Example: “Traffic: 5,000 visitors per month → 12,500. Sales: 2 million rubles per month → 3.2 million. Conversion: 1.2% → 1.9%.”

Section “What we did” — only facts, no emotions. Not “conducted a deep analysis,” but “analyzed 1,500 keywords.” Not “optimized all pages,” but “optimized 200 product cards.” The model can use this section to answer “what work was done.” Specific numbers show the scale of work.

Section “Result in numbers” — what the model came for. The main section of the case study. Several metrics are better than one. Percentages and absolute numbers together. Specify the period: “growth over 6 months.” Example: “Traffic: +150% (from 5,000 to 12,500). Sales: +60% (from 2 to 3.2 million). Conversion: +58% (from 1.2% to 1.9%). CPA: -40% (from 500 to 300 rubles). ROI: 320% over 6 months.”

Section “Evidence” — links, screenshots, testimonials with numbers. The model may not click the link, but the very fact of evidence is a trust signal. Example: “Screenshot from analytics: traffic growth,” “Link to case study on client’s website,” “Client testimonial: “Sales grew by 60%, we plan to continue cooperation.”” Quality content creation helps properly format such sections.

Example: case study before and after

Before optimization (emotional, for humans). “Our client, an online shoe store, approached us with a problem of falling sales. We conducted a deep market analysis, identified weak points, and developed a strategy. Thanks to our individual approach and modern technologies, we managed to turn the situation around. The client was thrilled with the results and thanked the whole team. Now their business is thriving.” The model sees: lots of emotions, no numbers, no structure. Chance of citation — 0%.

After optimization (factual, for models). Headline: “Increased sales of an online shoe store by 60% to 3.2 million rubles/month in 4 months.” Brief summary: “Client: online shoe store. In 4 months, increased traffic by 150%, sales by 60% to 3.2 million rubles/month. ROI 320% over six months.” Initial data: “Traffic: 5,000 → 12,500 (+150%). Sales: 2 million → 3.2 million (+60%). Conversion: 1.2% → 1.9% (+58%).” What we did: “Analyzed 1,500 keywords. Optimized 200 product cards. Created 30 review articles. Set up an advertising campaign in Yandex.Direct.” Result: table with metrics. Evidence: “Screenshots from analytics. Client testimonial: “Sales grew by 60% in 4 months. ROI 320%.”” The model sees: all key numbers highlighted, structure, evidence. Chance of citation — high.

What the model ignores in a case study

Emotional assessments. “The client was thrilled,” “We are very proud of the result,” “The team did a colossal job.” The model skips such phrases. They don’t affect trust, but they don’t help either. Better to remove them, freeing up space for numbers.

General words without specifics. “Deep market analysis,” “Comprehensive approach,” “Individual strategy,” “Modern technologies.” Without explanation, these phrases are useless. If you wrote “modern technologies,” clarify: “used a neural network to cluster 1,500 keywords.”

Subjective characteristics. “User-friendly interface,” “Attractive design,” “High speed.” The model cannot verify “user-friendliness” or “attractiveness.” Replace with objective: “page load time dropped from 3 to 1.2 seconds,” “Google PageSpeed score increased from 45 to 92.”

Long testimonials without numbers. “Thanks a lot, guys, you’re the best! I’ll definitely recommend you to friends.” The model cannot extract specifics from this. If a testimonial is the only evidence but has no numbers, its value to the model is near zero. Ask clients to include numbers in testimonials: “Sales grew by 60% in 4 months.”

How to check if a case study is ready for a neural network

Method 1. Measurability test. Go through the case study and find all numbers. If there are fewer than 5 numbers per 1,000 characters, the case study is not ready. A good case study: 8–12 numbers per 1,000–1,500 characters.

Method 2. Verifiability test. Can the model verify your claims? Are there links to sources? Are company names and dates specified? If the case study is anonymous (“client N”), chances are lower. But an anonymous case study with numbers is better than a case study without numbers.

Method 3. Structure test. Does the case study have sections “initial data,” “actions,” “result”? Are there headings, lists, tables? If the case study is solid text, the model may skip it.

Method 4. Headline test. Can you tell from the headline what the case study is about and what the result is? If the headline is “Case study: online store promotion” — bad. If “Increased sales by 60% to 3.2 million rubles/month in 4 months” — good.

Read about new search trends in the article new search trends. To comprehensively improve site visibility, use the service comprehensive website promotion.

Conclusion

Artificial intelligence cites numbers, not emotions, because numbers can be verified, compared, and used to justify recommendations. Emotions are subjective. One user is “thrilled,” another may be dissatisfied. The model cannot rely on subjective assessments. A case study for a neural network is not a success story, but a set of structured data. Five rules for a good case study: headline with the main number. Brief summary with key metrics. Sections “initial data,” “actions,” “result.” At least 8–12 numbers per case study. Links to evidence (screenshots, reports, testimonials with numbers).

Rewrite your case studies. Remove emotions. Add numbers: percentages, timelines, absolute values. Structure: initial data → actions → result. Check with the four tests. And then the neural network will start citing your case studies in answers to queries like “examples of successful promotion,” “SEO case studies,” “real optimization results.”

Frequently asked questions

What to do if there are no significant numbers (modest result)?
Be honest. “Increased traffic by 10% in 3 months” — that’s also a number. Don’t exaggerate. The model can verify (e.g., via Wayback Machine or analytics). Better a modest honest case study than a loud fake one. A modest case study can also be useful to the model as an example of “realistic expectations.”

Can the model use a case study without a client link?
It can, but trust will be lower. An anonymous case study is harder to verify. If it’s impossible to disclose the client (NDA), at least specify the industry and region. Example: “Client — a large furniture manufacturer in the Moscow region. Name not disclosed due to NDA.” This is better than just “anonymous client.”

Should I add photos and screenshots to a case study?
Yes, screenshots from analytics are excellent evidence. ALT tags for screenshots should contain numbers. Example: “Screenshot from Google Analytics: traffic growth from 5,000 to 12,500 in 4 months.” The model doesn’t “see” the image but reads the ALT tag. Textual description of the screenshot in the ALT tag is your main tool.

How often should I update old case studies?
If a case study has dates, it remains relevant as an example of past experience. But the model prefers fresh data. Add new case studies regularly (at least 1–2 per quarter). Old case studies can be kept, but not necessarily updated — an archive of past work is also valuable. Update only if new data appears for the same project.