The AI case studies that get shared are enterprise monoliths. The more useful ones are small sites with small budgets doing specific things well. Seven real implementations — what they built, what it cost, what moved.

1. The recipe blog's unit converter. A custom AI-built converter widget (cups↔grams per ingredient) embedded on every recipe. One evening of work; time-on-page rose 40% and the widget pages became the site's most-linked asset. 2. The plumber's quote bot. A guided form that asks five questions and emails a price range. Built on a form tool plus AI logic for $0/month extra. Quote requests tripled; the owner reports the qualified ones "arrive pre-sold".

3. The newsletter's archive engine. All 140 back issues embedded and searchable conversationally — "what did you write about pricing in March?" Support emails about old content dropped near zero; archive pageviews doubled. 4. The review site's comparison tables. AI-generated structured comparisons from product specs, human-verified, schema-marked. Rich results appeared within weeks; CTR on those listings rose visibly.

5. The local directory's auto-summaries. Each business page got a generated "what to know" summary from its existing data, edited in batches monthly. Task that wouldn't have happened manually: 400 pages, one weekend. Dwell time up, bounce down. 6. The course creator's FAQ assistant. A retrieval chatbot grounded in course materials, answering student questions at 2am. Refund requests fell noticeably in the following quarter — the instructor attributes it to unblocking students before frustration compounds.

7. The niche site's translation pilot. Machine translation with native review on the top 30 posts, opening a Spanish version. Six months later: 15% of total traffic from a market the site had never touched.

The pattern across all seven: small, specific, embedded in an existing workflow, human-reviewed. None required a data scientist; all required an owner who picked one problem and finished it. That's the actual small-site AI playbook — one tool, one use case, shipped.