GEO Audit Checklist: 27 Points to Optimize for AI Search

A 27-point GEO audit identifies every high-impact fix before you invest in a GEO campaign. The audit covers four dimensions: technical access (8 points), content quality (9 points), authority signals (5 points), and measurement (5 points). Run this checklist against any site before publishing new content or measuring visibility.

The Princeton GEO study quantified the lift from strategies like expert quotations (+41%), statistics (+33%), fluency (+29%), and citations (+28%). None of those lifts are reachable if the foundational audit points are failing. In 2026, 32% of digital marketing leaders named GEO their top priority, but only 23% are currently investing in GEO measurement¹. The gap means most sites have fixable issues that competitors are not addressing. Audit first, optimize second, measure third.

Updated September 2026: add AI-attribution to the measurement section

The audit's measurement dimension (points 23–27) gained a native tool on May 13, 2026, when GA4's Default Channel Group shipped an "AI Assistant" channel that classifies traffic from ChatGPT, Gemini, Claude, and other AI surfaces by referrer header (Adamarant, 2026). It is now a checklist item: enable the native channel and add a custom channel-group regex, because 35–70% of AI referral sessions arrive without a referrer and still land in Direct¹⁸. Separately, instrument a baseline mention-rate check as a pre-audit step: AthenaHQ's State of AI Search 2026 found the average brand appears in only 17.24% of relevant prompts while leaders reach 56.71%¹⁹ — that gap is your GEO headline KPI.

"The gap between intention and measurement in GEO is stark: 32% of leaders say it's their top priority, but only 23% have the measurement infrastructure to know whether their optimizations work. A GEO audit closes that gap by identifying exactly where to invest."

Step 0: baseline your citation rate before you audit anything

Most audits fail because they measure the wrong starting point. Teams assume a #1 ranking should produce a citation, then conclude the audit found a content problem when nothing changed. The largest AI Overview citation study published so far says otherwise: Conductor's 2026 analysis of 167,867,680 citations found the position-1 result was cited 24.9% of the time and the position-10 result 10.2%, with 57% of all citations landing outside the top 10²⁰. Holding the top spot still means going uncited three times out of four.

Record your current per-engine citation rate for a fixed query set before touching anything. Without it, the 40-60% monthly citation churn⁷ will be indistinguishable from the effect of your own fixes.

Section A: Technical access (8 points)

Technical access is the prerequisite for every other GEO strategy. If AI search crawlers cannot fetch your pages, no amount of content quality or structured data matters. With 12 AI crawlers now scanning the web³, misconfigured access can silently block a growing share of your potential visibility.

1. 1.
   Site is crawlable by AI search engines

   Verify via server logs that OAI-SearchBot, PerplexityBot, Claude-SearchBot, and Gemini-Crawler fetch your pages. If logs show no AI bot traffic after 30 days, content is invisible to AI search. wislr.com's 48-day log study found 12,099 AI bot requests across multiple sites⁴ — use this as a baseline.

2. 2.
   Robots.txt explicitly allows AI search crawlers

   Include User-agent blocks for OAI-SearchBot, ChatGPT-User, PerplexityBot, Claude-SearchBot, Claude-User, Google-Extended, Gemini-Crawler, and Applebot-Extended with Allow: / rules. Place AI-specific blocks before the generic * block. SERanking's audit of 300,000 domains found 12% still block GPTBot⁵ — likely also blocking OAI-SearchBot. Note the two OpenAI crawlers are governed differently: OAI-SearchBot follows robots.txt and decides whether you appear in ChatGPT search citations, while ChatGPT-User fetches pages for a live user task and is governed by CDN and firewall bot rules instead²¹.

3. 3.
   No accidental blocks on AI crawlers

   Test robots.txt in Google Search Console's robots.txt Tester with each AI user-agent. A single misplaced Disallow can silently block an entire crawler. The three-layer visibility system (robots.txt + llms.txt + ai.txt) is now the 2026 standard³. Also check the CDN layer separately: from 15 September 2026 Cloudflare defaults new domains to block Agent and Training bot traffic on ad-displaying pages while leaving Search traffic allowed²¹. A clean robots.txt does not prove the crawler reached you.

4. 4.
   Server-side rendering of article content

   AI crawlers do not always execute JavaScript. Verify with curl or a headless browser that article HTML is in the initial response, not loaded client-side. Use Nuxt SSR, Next.js SSR, or static generation. Pages with sequential headings + rich schema earn 2.8x higher citation rates⁶.

5. 5.
   Valid XML sitemap

   Sitemap at /sitemap.xml listing all article URLs. Submit to Google Search Console. Reference sitemap URL in robots.txt. AI engines use sitemaps as a discovery signal.

6. 6.
   Fast server response times

   TTFB under 600ms. AI crawlers have budget limits — slow sites get fewer pages crawled. Monitor with WebPageTest or Pingdom.

7. 7.
   Mobile-responsive layout

   AI engines crawl mobile versions. Verify with Google Mobile-Friendly Test. Layout breaks on mobile hurt extraction.

8. 8.
   Canonical URLs configured

   Each article has a canonical link tag pointing to its preferred URL. Prevents duplicate-content dilution across query parameters and mirrors.

Section B: Content quality (9 points)

Content quality is where the Princeton GEO study measured the largest lifts. The +41% from quotations, +33% from statistics, +29% from fluency, and +28% from citations all live in this section. In 2026, pages updated within the past two months earn 28% more AI citations than pages older than six months⁷ — freshness is a direct citation signal. Structure matters as much: 68.7% of AI-cited pages follow a strict H1→H2→H3 hierarchy (ConvertMate 2026 benchmark), and an Ahrefs study of 863,000 keywords found only 38% of pages cited in AI Overviews also rank in Google's top 10 for the same query — down from 76% seven months earlier after the Gemini 3 upgrade. Traditional rankings no longer guarantee AI visibility, which makes the audit's content and authority sections the real differentiator.

9. 9.
   Specific statistics on every page

   Each article contains 4–8 well-placed statistics with sources. Target one statistic per 200–300 words. Content with statistics every 150–200 words sees 30–40% higher AI visibility⁸.

10. 10.
    Inline citations to authoritative sources

    Every factual claim has an inline citation: claim + source name + year. Hyperlink the source name. Aim for Tier 1–3 sources (academic, official, recognized industry). Citations boost AI visibility by +28%⁸.

11. 11.
    Expert quotations on analysis-heavy pages

    Include 1–3 named expert quotations per article where relevant. Format: quote + named source + role + venue + date. Use blockquote HTML element. Quotations boost AI visibility by +41%⁸.

12. 12.
    Fluency passes the read-aloud test

    Read each paragraph aloud. If you stumble, fix it. Common killers: passive voice chains, sentences over 30 words, parenthetical asides over 8 words. Fluency optimization lifts AI visibility by +29%⁸.

13. 13.
    One idea per paragraph

    Paragraphs are 2–4 sentences, 40–80 words. Multiple ideas per paragraph force re-rankers to split text, losing context. Test: can you summarize the paragraph in 6 words?

14. 14.
    Descriptive H2 and H3 headings

    Headings describe their section content. "How to Add Citations for AI Search" beats "Diving Deeper." 4–12 words per heading. No skipped heading levels. 44.2% of LLM citations come from the first 30% of content⁹ — headings near the top carry extra weight. In ConvertMate's 2026 benchmark, 68.7% of AI-cited pages follow a strict H1→H2→H3 structure — the single strongest on-page predictor of citation after statistics density.

15. 15.
    Numbered lists for procedural content

    Any procedural or sequential content uses <ol>. Numbered lists outperform inline prose for procedural queries by 40–60% in citation rate.

16. 16.
    Tables for comparative data

    Any comparative data uses <table> with captioned source. Tables are 3–5× more extractable than inline prose for comparative data.

17. 17.
    FAQ section with 3–5 questions

    Every article ends with an FAQ section. Each answer is 30–60 words, self-contained. Pairs with FAQPage JSON-LD schema (point 21). Note: Google retired FAQ rich results on May 7, 2026¹⁰, but the visible FAQ section + JSON-LD still supports AI citation extraction.

Section C: Authority signals (5 points)

Authority signals tell AI engines who you are and why your content is trustworthy. These signals support entity disambiguation and source-weighting in re-ranking. Pages with schema markup are 3x more likely to earn AI citations than pages without¹¹.

18. 18.
    Organization schema on homepage

    JSON-LD Organization block with name, url, logo, description, foundingDate, and sameAs array linking to your profiles on Twitter, GitHub, LinkedIn, etc.

19. 19.
    Article schema on every article

    JSON-LD Article block with headline, description, url, datePublished, dateModified, author, publisher, inLanguage. Update dateModified on every revision. Google I/O 2026 confirmed schema.org v30.0 and emphasized the importance of accurate date metadata for AI citations¹⁰.

20. 20.
    WebSite schema for sitelinks and entity

    JSON-LD WebSite block with name, url, and potentialAction for sitelinks search box. Supports brand entity disambiguation.

21. 21.
    FAQPage schema matching visible FAQ text

    JSON-LD FAQPage block with mainEntity of Question/Answer pairs. Question name and acceptedAnswer text must match visible FAQ text verbatim. Mismatches trigger spam flags. FAQ rich results were retired May 2026, but the structured data still supports AI extraction¹⁰. Score this point on specificity, not presence. A 2026 controlled test of 1,006 pages found attribute-rich markup was cited 61.7% of the time, no markup at all 59.8%, and generic markup just 41.6% — roughly 18 points below having no schema, significant at p = .012²². A plugin default is worse than nothing.

22. 22.
    HowTo schema on procedural content

    JSON-LD HowTo block on any step-by-step guide. Each step has position, name, and text. Do not apply HowTo to non-procedural content — triggers spam classification.

Section D: Measurement (5 points)

Without measurement, GEO optimization is invisible. This section ensures you can detect visibility changes and attribute them to specific optimizations. In 2026, only 16% of Fortune 500 companies track AI search performance⁶ — measurement is a competitive advantage.

23. 23.
    Defined query set of 50–200 queries

    Representative set covering branded, category, comparison, and informational intents. Fixed across weeks for trend reliability.

24. 24.
    Multi-run query execution (10+ runs per query)

    Same query returns different results on 99 of 100 runs. AI recommendation lists repeat less than 1% of the time. Single-run tracking is noise. Minimum 10 runs per query per engine per week, and 15–25 on high-churn surfaces — Pineprompt's October 2026 seven-engine benchmark measured day-to-day churn up to ~69% on Perplexity and a ~36–56% band across engines²⁴. The scale is now measured: SE Ranking studied 10,000 queries in Google AI Mode and found only 9.2% average URL overlap across three repeated searches of the same query, with 21.2% of queries returning zero overlapping results²³. At that overlap rate, a week-over-week change on fewer than 10 runs is almost always the model or the noise, not your content.

25. 25.
    Tracking across 3+ AI engines

    At minimum ChatGPT Search, Perplexity, and Google AI Overviews. Add Claude and Gemini if budget allows. The engine-count floor is now empirical: in Pineprompt's 18,792-answer / 7-engine / 14-day benchmark only 12.9% of named brands appeared on every engine, while 46.2% appeared on a single engine only²⁴; LLM Scout separately found the four major assistants disagreed on whether to name a tracked brand in 27.2% of 749 identical buying questions²⁵. A one-engine panel is a sample of one market, not of your category. URL overlap between AI Overviews and AI Mode is only 10.7%⁵ — tracking one engine gives a fragmented picture.

26. 26.
    Weekly cadence with 4-week trend analysis

    Run weekly. Compare week-over-week. React to 4-week trends, not single-week swings. Daily is too noisy; monthly misses trends. With 40–60% of citations churning monthly⁷, monthly checks miss most of the signal.

27. 27.
    Mention rate, citation frequency, sentiment, share of voice tracked

    All four metrics tracked per engine per week. Share of voice requires competitor tracking. Sentiment classified via LLM-based classifier for consistency.

October 2026 addendum: the 5 retro-checks that reclassify a passing audit

The 27 points above audit your site as it exists. They do not catch three failure modes that October 2026 research made measurable: compliance that hides a silent loss, single-engine evidence mistaken for market evidence, and retrieval failure misread as page failure. Run these five checks after the 27-point pass; any failure downgrades the relevant section even if every original point scored.

1. E1.
   Index health — did any unindexed window erase your AI surface?

   A page must be indexed and eligible to appear in AI features; an unindexed page cannot be cited however good it is. One publisher recorded a 28-hour site-wide noindex and saw daily AI impressions fall roughly 55% at the low point (Value Add VC, September 2026)²⁶. Audit action: diff "Indexed pages" against "Generative AI impressions by page" — a page that keeps ranking but loses AI impressions is an index or canonical problem, not a content problem.

2. E2.
   Retrieval test — are you retrieved at all, or only badly composed?

   A site-wide analysis of 158,099 answers and 468,037 citations (Chen & Koudas, arXiv:2610.15033, October 2026) attributes roughly 32% of citation variance to whether a page is retrieved, versus about 2.8% to on-page attributes, found 100% of cited pages were retrieved before being cited, and observed that only about 3% of retrieved pages rank in Google's top 50²⁷. Audit action: before rewriting content on a page with zero AI impressions, confirm the URL is physically retrievable by the engine's search crawler with curl -A <bot> — a robots.txt or edge-rule error looks identical to a content score of zero.

3. E3.
   Measurement honesty — are you reporting impressions as if they were demand?

   A randomised field experiment from the University of Pennsylvania and Northeastern University (1,100 US Google users; reported 23 September 2026) found forcing searches into AI Mode cut clicks to external sites by 18.8 percentage points, while hiding AI Overviews lifted external clicks by an estimated +8.8 points²⁸. Audit action: any KPI deck that shows AI impressions without clicks and sessions fails this point. Report AI share (AI impressions ÷ web impressions, same page, same month) alongside clicks.

4. E4.
   Off-domain share — is your citation footprint actually yours?

   Only 2.9% of brand citations point to the brand's own domain (Pratyush Kumar, 102,025 prompt responses / 149,912 citations), and a 2026 corpus study across 1,000 prompts found 45.7% of the top-10 cited sources were off-site²⁸. Audit action: if every remediation item in your 90-day plan is an on-site fix, the audit is structurally incomplete — add at least two items that change third-party pages (roundups, review sites, directories, data partnerships).

5. E5.
   Content-type mix — do you own the format engines actually pull?

   In one audited property, a comparison article earned 39 AI impressions per 100 web impressions, while glossary/definition pages earned 5 — roughly an 8× gap in how often Google pulls each format into an answer²⁹. Audit action: score your sitemap by format. If more than half your pages are definitions or news-style posts and fewer than a tenth are comparison or "best-of" formats, you are structurally underweight in the format that gets cited.

How to score the audit

Score 1 point for each fully-implemented item. Half-credit for partial implementation. Add the four section scores for a total out of 27.

"Pages ranked 5th in traditional search achieved up to 115% visibility improvement through GEO optimization. The 1st-ranked page lost 30%. GEO audit fixes have outsized impact for pages struggling in traditional SEO — the audit is the highest-ROI activity in a GEO program."

The 90-day remediation plan

1. Days 1–14:
   Fix Section A (technical access). Update robots.txt, verify SSR, fix sitemap. Test each AI crawler user-agent. Ensure OAI-SearchBot, PerplexityBot, Claude-SearchBot, and Gemini-Crawler can crawl freely.
2. Days 15–45:
   Fix Section C (authority signals). Deploy Organization, Article, FAQPage, HowTo, WebSite schema. Validate with Google Rich Results Test. Schema v30.0 support confirmed at Google I/O 2026.
3. Days 46–75:
   Fix Section B (content quality) on top 10 trafficked articles. Add statistics, citations, quotations, FAQ sections. Apply 6 formatting rules. Freshness bonus: pages updated within 2 months earn 28% more citations⁷.
4. Days 76–90:
   Implement Section D (measurement). Build query set, choose tracking tool (Semrush AI Visibility, Profound, or Peec AI), run baseline measurement. Only 23% of marketers currently measure GEO¹ — early measurement creates months of competitive data.

Frequently asked questions