Most “SEO audit” prompts hand back a generic checklist anyone could Google. These 12 target the failure modes that actually sink content sites — duplicate titles, orphan pages, hreflang mismatches, thin pages, keyword cannibalization — each with a measurable threshold and one concrete fix per finding, not vague advice.
TL;DR
- Paste real data (slugs, word counts, headings, JSON-LD, Lighthouse numbers) into each prompt. An audit with no data returns a checklist, not findings.
- Every prompt carries an exact threshold (
title >60 chars,LCP >2.5s,word count <600) and asks for one action per finding: expand, merge, noindex, or 301. - Run them on a 1M-token model so a whole site fits in one paste: Claude Opus 4.7 / Sonnet 4.6 or Gemini 3.1 Pro (both 1M tokens as of June 2026). ChatGPT Plus tops out around 320 pages in-app.
- Replace each
[paste]placeholder with your data before sending.
Best for
- Pre-launch site QA
- Quarterly content refresh
- Bilingual / multi-locale sites
- Post-migration sanity check
- Programmatic SEO QA
Which model to run these on
These prompts work on any frontier model, but they live or die on context length, because you want to paste a whole site’s worth of slugs, headings, or JSON-LD in one shot. As of June 2026:
| Model | Context | Notes |
|---|---|---|
| Claude Opus 4.7 / Sonnet 4.6 | 1M tokens | Strongest at structured “output a fix table” work; 1M standard, no surcharge |
| Gemini 3.1 Pro | 1M tokens | 1M on Google AI Pro ($19.99/mo); good for very long URL lists |
| ChatGPT Plus (GPT-5.5) | ~320 pages in-app | Fine for one section at a time; full 1M only on the $200 Pro tier |
For a 1,000-article site, dump the slug + metadata table into one prompt rather than paginating — the model catches duplicate titles across the whole set that page-by-page audits miss. If your dataset is genuinely huge, run the audit per section (one subfolder at a time) so the model can still reason coherently across what it sees.
1. Metadata audit
Below are 50 article slugs with title and meta description. Identify, with the exact slug for each: (a) titles >60 chars (truncate near Google's ~600px desktop limit), (b) titles <30 chars (probably weak), (c) descriptions >160 or <120 chars, (d) duplicate or near-duplicate titles, (e) titles missing the primary keyword, (f) clickbait or all-caps. Output a fix-needed table with current then suggested.
Data:
[paste]2. Internal-link audit
Below is an article list with slug, H2 headings, and tags. For each article: (a) suggest 3 outbound internal links to other articles in the list with the anchor text to use, (b) flag the article as an orphan if it has zero inbound links from the list, (c) flag any article with >40 inbound links as a hub that needs pruning. Output as a table.
Articles:
[paste]3. Hreflang & translation-pair audit
Below is a list of en and zh slugs with their translationKey. Identify: (a) translationKeys that exist in only one language, (b) pairs where the slugs are very different (suggest the correct match), (c) typoed or near-duplicate translationKeys that probably should be unified, (d) hreflang tags that point to a 404 or a redirect. Output an action list grouped by severity.
Data:
[paste]4. Thin-content detector with thresholds
Below are slugs with word count, number of H2s, number of internal links, and last-updated date. Flag thin content using this rule: word count <600 AND H2 count is 2 or fewer AND internal links is 1 or fewer AND not updated in 6 months. For each flagged slug recommend one action: expand (worth saving), merge into a named target slug (cite the best merge target), noindex, or delete with 301. Justify each call.
Data:
[paste]5. Title-uniqueness & cannibalization check
Below is a list of article titles with their primary target keyword. Identify: (a) near-duplicate titles (>70% token overlap), (b) different titles targeting the same primary keyword (cannibalization), (c) titles where the primary keyword is not in the title. For each conflict pair, recommend: which to keep, which to merge or de-target, and the new title for the survivor.
Titles + keywords:
[paste]6. Canonical & duplication audit
Below are URL patterns with their canonical tags and any redirects. Identify: (a) pages missing a self-referencing canonical, (b) multiple pages canonicalizing to the same target (consolidation candidate), (c) trailing-slash inconsistency between canonical and rendered URL, (d) cross-language canonical pointing to the wrong language, (e) canonical then redirect then final URL chains. Fix per finding.
Data:
[paste]7. Structured-data / JSON-LD audit
Below is the JSON-LD for an article. Audit: (a) required fields for the chosen schema.org type are present and non-empty, (b) datatypes match spec (Date in ISO 8601, Person/Organization for author), (c) recommended fields that would unlock rich results are missing (e.g., FAQPage, BreadcrumbList, image dimensions), (d) the chosen type fits the page (Article vs BlogPosting vs HowTo). Output fix-needed JSON.
JSON-LD:
[paste]8. Sitemap & crawl-budget audit
Below is the sitemap.xml (or a list of URLs). Identify: (a) URLs in sitemap that return non-200, (b) URLs in sitemap that are noindex'd or canonicalized elsewhere (waste), (c) parameterized URLs that should be excluded, (d) missing high-priority pages, (e) lastmod dates that look stale or wrong. Output a sitemap cleanup diff.
Sitemap:
[paste]9. Heading hierarchy & content structure audit
Below is the heading outline (H1-H4) for 10 articles. For each article, flag: (a) multiple H1s, (b) H2 immediately followed by H4 (skipped level), (c) more than 8 H2s without H3s under them (probably should restructure), (d) headings that are sentences rather than scannable phrases. Recommend a corrected outline per flagged article.
Outlines:
[paste]10. Search-intent vs page-format mismatch
Below are pages with their target keyword and current page format (list / how-to / comparison / definition / commercial). Determine the dominant intent for each keyword (informational, commercial, transactional, navigational) and flag mismatches — e.g., a commercial keyword pointed at a thin definition page. Recommend the right format for each mismatch.
Pages + keywords:
[paste]11. Core Web Vitals & content-page perf audit
Below are Lighthouse / CrUX numbers for 10 article pages: LCP, INP, CLS. Identify: (a) which fail Google's thresholds (LCP >2.5s, INP >200ms, CLS >0.1), (b) the likely root cause given that these are content pages (heavy hero image, JS-injected ads, layout-shifting embeds, render-blocking fonts), (c) the cheapest fix per page. Sort by traffic so the high-impact fixes come first.
Data:
[paste]The thresholds to feed the model (Google’s “good” bar, unchanged through June 2026):
| Metric | Good | Needs work | Common content-page cause |
|---|---|---|---|
| LCP | < 2.5s | 2.5-4.0s | Unoptimized hero image, render-blocking font |
| INP | < 200ms | 200-500ms | Heavy ad / analytics JS on the main thread |
| CLS | < 0.1 | 0.1-0.25 | Ads or embeds inserted without reserved space |
A URL group only shows “Good” in Search Console when at least 75% of real visits clear all three at once. INP is the one most content sites fail.
12. Keyword-cluster overlap & consolidation plan
Below are 20 articles with their primary keyword and top 5 ranking keywords from Search Console. Identify clusters where 2+ articles rank for the same query (cannibalization), and clusters where one article ranks for what should be split into two. For each cluster, recommend: merge (cite which slug wins), split (give the new outlines), or keep separate with a clearer differentiator.
Data:
[paste]Common mistakes
- “Do an SEO audit” with no data and no measurable threshold
- Findings without an action — “your meta description is too short” with no fix
- Auditing meta tags but ignoring the actual content quality
- Treating every thin page as “expand” — sometimes the right move is merge or delete
- Skipping the bilingual hreflang and translationKey integrity check
- Pasting a screenshot of a report instead of the raw table — the model can only audit data it can read
FAQ
Can AI replace a tool like Ahrefs or Screaming Frog? No. Crawlers gather the data (status codes, canonicals, word counts, CrUX numbers); the model interprets it and prioritizes fixes. Export from your crawler, paste into prompt 2, 6, or 8, and let the model turn a 5,000-row CSV into a ranked action list.
Which model handles the biggest site in one paste? Claude Opus 4.7, Sonnet 4.6, and Gemini 3.1 Pro all ship a 1M-token context window as of June 2026, enough for several thousand slugs with metadata. ChatGPT Plus tops out around 320 pages in-app, so audit it section by section.
How do I stop the model from inventing slugs in its fixes? Tell it explicitly: “only reference slugs that appear in the data above; if no good merge target exists, say so.” The thin-content (4) and cannibalization (5, 12) prompts are where hallucinated targets sneak in, so verify every suggested merge before acting.
Are these thresholds still current? Yes as of June 2026. Title truncation near 600px (~60 chars), meta descriptions ~120-160 chars, and the Core Web Vitals bars (LCP 2.5s, INP 200ms, CLS 0.1) are unchanged. Re-check the CWV numbers on web.dev if you read this later.
Do I run all 12 every time? No. Pre-launch: run 1, 3, 6, 7. Quarterly refresh: 4, 5, 12. After a migration: 3, 6, 8. Performance complaint: 11. Pick by the symptom, not the calendar.
Related
- Meta description prompts
- SEO title prompts
- Code review prompts
- Deployment check prompts
- Article rewrite prompts
Tags: #Prompt #AI coding #SEO audit #SEO