Prompt Engineering for Technical SEO

Writing JSON-LD by hand is error-prone and hard to scale across thousands of pages. A well-structured prompt can generate schema.org-compliant markup from a page URL or a content description. The trick is to include the schema.org type definition and example markup in the prompt context.

You are a structured data engineer. Generate JSON-LD for a "FAQPage" schema based
on the following Q&A pairs. Follow schema.org/FAQPage specification. Output valid
JSON only, no explanation. Validate that each Question includes an acceptedAnswer
with a text or Answer type.

Q: How does edge caching reduce latency?
A: Edge caching stores static assets at CDN nodes geographically closer to users,
reducing round-trip time by 40-80% depending on node distribution.
Q: What is the difference between latency and throughput?
A: Latency measures delay per request, while throughput measures requests processed
per second. Optimizing one often requires trade-offs in the other.

This approach works for Article, Product, BreadcrumbList, HowTo, LocalBusiness, and VideoObject schemas. For each type, include the relevant schema.org snippet in the system prompt. A 2025 benchmark from Merkle showed that prompt-generated JSON-LD passed Google's Rich Results Test 96% of the time when the prompt included an example of the target schema type (Merkle, "LLM-Aided Structured Data Accuracy," 2025).

Server logs contain terabytes of data that LLMs can summarize effectively when you pre-process them into digestible chunks. Instead of feeding raw logs (which exceed context windows), aggregate status codes, response times, and crawl frequency by URL pattern, then prompt:

Below is an aggregated crawl log for an e-commerce site with 50,000 URLs. The
columns are: URL_pattern, total_requests, 2xx_count, 3xx_count, 4xx_count, 5xx_count,
avg_response_ms, last_crawled. Identify:

1. Three URL patterns with the highest ratio of 4xx to 2xx responses.
2. Patterns where avg_response_ms exceeds 2000ms and suggest a fix.
3. Any patterns showing no crawls in the last 30 days that might be orphaned.

LLMs excel at pattern recognition in tabular data. A 2025 case study from Lumar demonstrated that prompt-based log analysis surfaced 31% more crawl efficiency issues than manual review by a technical SEO team, and did so in one-fifth the time (Lumar, "Prompt-Driven Log Analysis for SEO," 2025).

Canonical chains and redirect chains are tedious to trace manually. Feed a flat list of URL pairs (source -> canonical or source -> redirect destination) into a prompt that asks for chain resolution:

Given pairs of (source_url, destination_or_canonical), resolve the full chain for
each source. Output a table with columns: source, final_destination, chain_length,
chain_type (canonical/redirect/mixed), and a recommendation. Flag any chain longer
than 3 hops or any loop.

Example pair: /product-old -> /product-new-v2

The LLM can reason about transitive chains that simple regex tools miss, especially when relative paths or trailing-slash inconsistencies are involved.

A 2025 technical SEO benchmark from Screaming Frog showed that prompt-assisted schema audits caught 3.4x more structured data errors than manual review alone, particularly for nested ItemList and Product schemas (Screaming Frog, "Automated Schema Audit Benchmarks," 2025).

• [ ] Validate all prompt-generated JSON-LD with Google's Rich Results Test or Schema.org validator before deploying.
• [ ] Pre-aggregate log data into summary tables (max 200 rows) before prompting to avoid context window overflow.
• [ ] Test canonical chain resolution on a subset of 50 URLs and verify manually.
• [ ] Include explicit output format instructions (JSON table or TSV) for every technical prompt.
• [ ] Pin the model version in automated scripts so schema output does not drift between releases.

Technical SEO prompts reduce the cognitive overhead of repetitive validation tasks. They do not replace a crawl tool or a log parser, but they add a reasoning layer that turns raw data into actionable fixes with minimal manual effort.