What this covers
Most indie sites have the same five SEO problems (long titles, vague descriptions, missing h1, weak internal links, no structured data) - and an LLM can catch all five in one pass if you prompt it right. This guide is the exact prompt chain plus the input format that turns a sitemap into a per-page fix list, including the verification step that stops you from trusting AI claims about meta tags it can’t actually see.
Who this is for
Indie devs, content creators, and small marketing teams running content sites (Astro, Next, WordPress, Hugo). Useful for anyone with 50-500 pages who can’t afford a $5k SEO consultant but does have an LLM subscription.
When to reach for it
Before launching a content site (audit while it’s still cheap to fix). Quarterly after launch (because titles drift, descriptions get stale, internal links die). Also after any major content migration or theme change.
Before you start
- Have your sitemap accessible (
/sitemap.xml) or a list of top URLs. Without input, the LLM will invent pages. - Decide on your title length target (50-60 chars for primary, up to 70 if needed) and description target (140-160 chars). Different targets = different audits.
- Choose your LLM: GPT-5.5 / Claude with web browsing, or feed it the HTML directly via a script. Don’t rely on a model’s training data for current page state.
- Set aside time for the fix pass - the audit is fast, fixing 50 pages takes a few hours.
Step by step
- Feed AI a sample of your sitemap or top pages. Start with 20 highest-traffic URLs (from Search Console) before scaling to all pages. Sample first, scale second.
- Ask it to check structured fields per page: title length, description specificity, h1 presence, h1 uniqueness, internal links present, schema.org markup, canonical tag. List them explicitly.
- Generate a fix list per page in a single table with URL, problem, severity (P0/P1/P2), proposed fix. Don’t accept “your titles could be better” - you want page-by-page rows.
- Apply fixes - either manually for sites without a CMS API, or via a script if the content is in markdown / MDX with frontmatter.
- Re-validate with the same prompt after a deploy. The model should now report a clean audit; if it doesn’t, the gap is the next iteration.
A prompt that actually works
You are an SEO auditor. For each URL below, fetch the page (or use the HTML I provide)
and report:
- Title: text, char count, primary keyword present (Y/N), under 60 chars (Y/N)
- Description: text, char count, specific to page (Y/N), under 160 chars (Y/N)
- H1: present, unique vs title, includes primary keyword
- Schema.org JSON-LD: present, type (Article/Product/FAQ), valid
- Canonical: present and self-referential
- Internal links out: count, anchor diversity (good/repetitive)
- Issues found: P0 (broken), P1 (weak), P2 (nice-to-have)
Output as a markdown table. Be specific - "title is too long (74 chars)" not "title could be shorter."
URLs:
{paste 20 URLs here}
The “be specific” sentence does heavy lifting. Without it, you get vague advice.
What AI is good and bad at in SEO
- Good: Per-page audit against a fixed checklist, title/description rewriting, identifying internal link gaps, generating FAQ schema, spotting cannibalization between similar titles.
- OK: Keyword research (better when paired with a real keyword tool like Ahrefs/Semrush), competitor analysis (limited by what’s in training data).
- Bad: Estimating search volume, predicting ranking impact, current SERP layout, anything requiring real-time SERP data without browsing.
- Dangerous: Inventing structured data fields that don’t exist in spec, claiming a page “ranks #3” without checking, fabricating backlink counts.
Always verify AI SEO claims against Search Console, Ahrefs, or a curl of the actual page.
Recommended workflow
sitemap sample (20 URLs) -> audit prompt -> per-page fix list (P0/P1/P2) -> apply fixes -> deploy -> re-audit same URLs -> if clean, scale to remaining pages -> schedule quarterly re-run. The first audit takes 30-60 minutes; quarterly re-runs after that take 10-15 minutes.
FAQ
- Can AI estimate keyword difficulty? - No, not reliably. Pair with a real SEO tool for KD; use AI for prioritization and rewriting.
- What about Core Web Vitals? - AI can read your PageSpeed report and prioritize fixes, but it can’t measure the live numbers. Pull data from PageSpeed Insights or CrUX.
- Should I let AI write the new titles? - Yes for first draft; always review. AI-generated titles tend to be slightly bland; humanize them.
- Can I automate the audit on every PR? - Yes - hook AI into a CI step that audits changed pages and posts findings as a PR comment. Useful for content-heavy sites.
- How often should I re-audit? - Quarterly minimum, monthly if content velocity is high, after any redesign or migration.
- What about LLM-specific SEO (GEO)? - Same fundamentals apply, plus: clear h1, FAQ blocks, factual tone, and citations. AI engines reward the same hygiene Google does.
Common mistakes
- Trusting AI claims without testing - the model may describe meta tags it cannot actually see.
- Auditing once and never re-checking - titles drift, descriptions get stale, internal links die.
- Asking “is this SEO-friendly?” - too vague; produces vague output.
- Bulk-applying AI rewrites without spot-checking - one page will have a weird translation or factual error.
- Ignoring P2 items entirely - they compound; quarterly attention pays off.
- Using only AI - keyword tools (Ahrefs/Semrush) and Search Console catch things AI can’t.
Related
Tags: #AI coding #Tutorial