The AI visibility audit: 40 checks to find why AI won’t cite you

TL;DR: Run the 40 checks below across five layers, crawl access, content structure, fact density, entity signals, and measurement, to find the specific reason ChatGPT, Perplexity, and Google AI Overviews skip your pages. Most sites fail on 3-6 checks, not all 40, and fixing those specific gaps moves citations faster than a general rewrite.

---

Table of contents

1. Why a checklist beats a general rewrite
2. Layer 1: Crawl and access (checks 1-8)
3. Layer 2: Content structure and extractability (checks 9-18)
4. Layer 3: Fact density and sourcing (checks 19-26)
5. Layer 4: Entity and trust signals (checks 27-34)
6. Layer 5: Measurement and tracking (checks 35-40)
7. Where structured data actually helps and where it does not
8. A visual audit flow
9. How to score your audit
10. Frequently asked questions
11. Key takeaways

---

Why a checklist beats a general rewrite

You rank on page one for your target keyword and ChatGPT still does not mention you. If that sounds familiar, the seven named causes behind ranking on Google but staying invisible in ChatGPT are worth reading alongside this checklist, since each cause maps to a specific layer below. That gap has a checkable cause, not a vague “AI doesn’t like your content” problem. Google’s own May 2026 guidance is explicit: “there are no additional requirements to appear in AI Overviews or AI Mode, nor special optimizations necessary,” according to Google Search Central’s generative AI fundamentals guide. That statement rules out one popular theory (you need a magic markup tag) and points at the real answer: fundamentals, applied unevenly, that a checklist can isolate. For the parts of SEO that carry over versus what only moves AI answers, see the delta between good SEO and good GEO.

This audit splits into five layers: crawl access, content structure, fact density, entity and trust signals, and measurement. Run all 40 checks against one page you expected to get cited and did not. Most sites fail 3-6 checks, not all 40. Fix those specific gaps before you touch anything else.

Layer 1: Crawl and access (checks 1-8)

If an AI system cannot fetch your page, nothing else in this audit matters. Start here.

1. Is GPTBot allowed or deliberately blocked? OpenAI’s GPTBot feeds ChatGPT’s training data, not live retrieval. Per Anagram’s 2026 AI crawler guide, a common 2026 policy blocks GPTBot (training) while allowing OAI-SearchBot (live ChatGPT Search retrieval). Check which one you actually blocked, they are different user agents with different jobs.

2. Is OAI-SearchBot allowed? This is the bot that lets ChatGPT cite you in real-time answers. If your robots.txt blocks it, you are structurally invisible to ChatGPT Search regardless of content quality.

3. Is PerplexityBot allowed? Perplexity indexes continuously to answer live queries. Digital Applied’s 2026 decision matrix treats PerplexityBot as a retrieval bot worth allowing on public commercial content, distinct from bulk training crawlers.

4. Is ClaudeBot or Claude-SearchBot allowed? Same split applies: ClaudeBot trains Anthropic’s models, Claude-SearchBot and Claude-User retrieve live. Block one, allow the other, based on whether you want training inclusion.

Before you touch robots.txt, decide whether you actually need llms.txt at all. The engine-by-engine truth table on llms.txt covers which of these bots the file actually influences, since Google ignores it entirely while Perplexity and Claude read it.

5. Does your CDN or WAF silently block AI user agents? Cloudflare, Akamai, and some WAF rulesets ship default bot-blocking that catches AI crawlers even when robots.txt allows them. Check your CDN dashboard, not just robots.txt.

6. Does the page return a 200 status to an unauthenticated crawler? Paywalls, login walls, and cookie-consent gates that block content until interaction will return empty or truncated HTML to a crawler that does not execute your consent-manager JavaScript.

7. Does the page render its core content without JavaScript? AI crawlers vary in JS execution support and often time out before hydration completes. View your page with JavaScript disabled. If the answer to your target question is missing, so is your citation.

8. Is the canonical URL correct and self-referencing? A wrong or missing canonical splits citation signals across URL variants (with/without trailing slash, http/https, tracking parameters), diluting which version gets cited.

Layer 2: Content structure and extractability (checks 9-18)

Once a crawler can reach your page, it needs to lift a clean, self-contained answer out of it. A Princeton-affiliated GEO study found content with clear headings, bullets, and tables is 28-40% more likely to be cited, because the model can extract a bounded block without stitching context from elsewhere on the page.

9. Does the page answer its own title question in the first 100 words? If a reader has to scroll past three paragraphs of scene-setting to find the answer, so does the model, and it will often pull a competitor’s more direct answer instead.

10. Is the first 500 words dense with sourced, extractable facts? Front-load your strongest, most citable claims. This is the section most likely to get lifted into a synthesized answer.

11. Does each H2 section stand alone without needing prior context? Models frequently extract a single section, not the whole page. A section that says “as covered above” fails when extracted in isolation. Restate the fact.

12. Are your key numbers in a table, not buried in prose? Tables are directly extractable as structured rows. A number inside a long sentence requires the model to parse it out correctly, which it does not always do.

13. Do you use plain declarative sentences instead of hedged ones? The GEO research above found that removing subjective phrases like “I think” and “we believe” increases the odds your text gets selected, because objective, declarative sentences carry lower extraction risk for the model.

14. Are your headings descriptive questions or clear statements, not clever wordplay? “The three-tier pricing trap” tells a model less than “SaaS pricing tiers: what each tier actually includes.” Extractability beats cleverness here.

15. Does the page avoid walls of unbroken paragraph text? Break every claim-heavy paragraph at 3-4 sentences. Long unbroken blocks are harder to lift cleanly.

16. Do your lists use real list markup, not dashes inside a paragraph? Semantic <ul>/<ol> markup parses more reliably than a paragraph with inline dashes pretending to be a list.

17. Is there a comparison table with named, measurable criteria? Comparative content accounts for roughly a third of AI citations by category, per the GEO research cited above. A table with concrete criteria, not adjectives, is the most citable single asset on a page. Google’s own guidance that structured data is not required for AI search does not mean structure does not matter, see the reconciliation of Google’s statement with measured citation lift for where the line actually falls.

18. Does the page avoid keyword-stuffed headings that read unnaturally? A heading written for an algorithm instead of a reader reads as lower-trust text to a model trained to recognize natural language patterns.

Layer 3: Fact density and sourcing (checks 19-26)

This is the layer with the strongest, most specific research behind it. The Princeton GEO study tested nine optimization methods; the three strongest, citing sources, adding quotations, and adding statistics, each lifted visibility by roughly 30-40% against baseline, according to AI Thinker Lab’s summary of the Princeton-backed GEO playbook.

19. Does every major claim carry a named, linked source? “Studies show” is not a source. “Ahrefs found X in a May 2026 study” is. Models weight the second far more heavily as a citable, corroborated fact.

20. Do you cite at least one statistic every 150-200 words? This density keeps fact-checkable claims flowing through the whole page, not just the intro, which matters because AI systems extract from anywhere on the page, not only the top.

21. Do you quote real named people, not anonymous “experts”? A quote needs a name, role, org, and ideally a link. An invented or unattributed quote is both a citation-quality failure and, per most style guides, a credibility risk if discovered.

22. Do you cite primary sources instead of secondary roundups? A page that cites Ahrefs directly outranks, in citation terms, a page that cites “a marketing blog that cited Ahrefs.” Corroboration strength drops with each layer of indirection.

23. Does your page itself cite credible outside sources? The Princeton research found that a page which cites its own sources functions as a trust signal at the claim level, the model treats your sourced claim as lower risk to repeat than an unsourced one.

24. Are your statistics date-stamped? “As of June 2026” tells both readers and models the claim is current, not a number that has quietly gone stale since 2023.

25. Do you avoid vague attribution phrases like “research suggests” or “experts agree”? These phrases are unverifiable by a model doing fact-checking during generation. Name the source or cut the sentence.

26. Does your content avoid contradicting itself across pages on the same topic? If your pricing page says one number and your comparison page says another, a model that cross-references both loses confidence in citing either. For the broader list of sourcing and structure mistakes that keep pages out of AI answers, see 11 AEO mistakes paired with their fixes.

Layer 4: Entity and trust signals (checks 27-34)

27. Does your brand have consistent NAP (name, address, phone) or entity data across your site, GBP, and third-party listings? Inconsistent entity data makes it harder for a model to confirm you are a real, single, coherent brand worth citing.

28. Do unlinked brand mentions appear on third-party sites? A BrandMentions study tracking roughly 410,000 public mentions across 240 brands over a 90-day window from April 19 to July 17, 2026, found text-only unlinked mentions made up the majority of the third-party footprint AI systems draw on, per BrandMentions’ 2026 report. A backlink is not required for a mention to count as a corroborating signal, see why brand mentions are becoming the new backlinks for AI search for how to build this signal deliberately.

29. Does your brand appear in relevant Reddit threads without being promotional? Reddit was the single most-cited domain by Google AI Overviews and Perplexity from August 2024 through June 2025, according to Sitebulb’s analysis, and ChatGPT cites Reddit in roughly 12% of US answers. If your brand never comes up in genuine community discussion, you are absent from a major citation source, see the compliant Reddit playbook for AI citations for how to participate without getting banned.

30. Does an About page or dedicated author bio establish who is behind the content? Google’s generative AI guidance points at “content that reflects genuine expertise” as a core, non-special-case requirement, per Semrush’s summary of Google’s May 2026 guide. An anonymous page with no visible author or organization signal is weaker on this axis, structured data cannot substitute for it.

31. Do you have a Wikipedia, Wikidata, or well-established third-party entity profile? Entity resolution, confirming your brand is a distinct, known thing rather than a generic phrase, draws on knowledge graph sources external to your own site.

32. Is your site fast and does it load reliably on mobile? Google’s guidance names “fast, accessible websites” directly as a foundational fundamental for AI feature eligibility, not a nice-to-have. If your traffic dropped and you suspect a core update rather than an audit gap, the core update recovery playbook separates update damage from continuous re-evaluation drift.

33. Does your page avoid intrusive interstitials that block content on first load? An interstitial that a crawler cannot dismiss functions the same as a hard content block, layer 1’s access checks apply here too.

34. Do you have a documented content review or update cadence? A visible “last updated” date paired with actual content changes signals the page is maintained, not abandoned, which matters for freshness-sensitive queries.

Layer 5: Measurement and tracking (checks 35-40)

35. Have you set up GA4 to isolate AI referral traffic? Without a referrer regex for chatgpt.com, perplexity.ai, and similar domains, AI-driven visits get lumped into “direct” traffic and you cannot measure whether your fixes worked. The full referrer regex and GA4 exploration setup walks through this check step by step.

36. Do you track brand-mention rate across at least one AI platform? Visionary’s 8,400-prompt AI search visibility tracker, run across ChatGPT, Claude, Perplexity Sonar, and Gemini in Q1 2026, found ChatGPT cited at least one named brand in 71.4% of commercial responses, Perplexity in 84.2%. Run a comparable set of prompts for your own category monthly and log whether you appear. If you need a single number to report instead of raw prompt logs, share of answer as a replacement for share of voice walks through how to define and track it.

37. Do you know your organic CTR impact from AI Overviews on your top queries? Ahrefs measured a 58% CTR reduction on the top-ranking page for AI-Overview-triggering keywords in a February 2026 analysis, per seo-kreativ’s summary. Pull Search Console data filtered to queries with an AI Overview present and compare CTR against non-AIO queries in the same position range. For a 30/60/90 plan keyed to your measured CTR-loss band, see the AI Overviews traffic-loss recovery plan.

38. Have you checked Search Console for indexing errors on the pages you expect AI systems to cite? A page that fails standard indexing is unlikely to surface in AI systems that lean on the same underlying index for Google-connected features.

39. Do you re-run this audit after each Google core update? Google confirmed two core updates in the first half of 2026, March 27 to April 8 and May 21 to June 2, according to ALM Corp’s tracking, with third-party volatility trackers also recording elevated movement outside those announced windows through June and July. Re-audit after each confirmed update and after unexplained traffic swings, and if a drop happens outside an announced window, the diagnostic playbook for traffic drops with no announced update covers what to check before assuming it is this audit’s fault.

40. Have you logged a baseline before making changes? Without a documented before-state (citation rate, AI referral traffic, ranked queries with AIO present), you cannot attribute a later improvement to a specific fix versus normal fluctuation.

Where structured data actually helps and where it does not

This is the check most audits get wrong, so it earns its own section instead of a single checklist line.

Ahrefs published a study on May 11, 2026 by Louise Linehan and Xibeijia Guan, “We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved,” tracking pages that added JSON-LD schema between August 2025 and March 2026, matched against 4,000 control pages, per seroundtable’s coverage. The result: Google AI Mode showed a +2.4% change, ChatGPT +2.2%, both statistically indistinguishable from zero, while Google AI Overviews showed -4.6%, a result significant at roughly 1-in-2,500 odds of chance. Every page in the dataset already had 100+ AI Overview citations before schema was added.

That study measured pages already winning citations. It did not test whether schema helps a page get crawled, parsed, and indexed in the first place, a distinction the study’s own authors flag. Google’s official position, from its May 2026 generative AI guide, is consistent with this: no special markup is required to appear in AI features. Structured data still has a job (rich results, product data, disambiguation for genuinely ambiguous entities), it is just not the citation lever many AEO vendors sell it as.

The practical rule for check 12 and check 17 above: use tables and clear headings because they help extraction, not because you believe schema alone will win you a citation. Treat schema as hygiene, not strategy.

Signal	Citation impact (2026 evidence)	Verdict
JSON-LD schema on already-cited pages	+2.4% AI Mode, +2.2% ChatGPT, -4.6% AI Overviews (Ahrefs, n=1,885)	✗ Not a citation lever
Cited sources + quotations + statistics	+30-40% visibility lift each (Princeton GEO study)	✓ Strongest documented lever
Clear headings, bullets, tables	28-40% more likely to be cited (Princeton GEO study)	✓ Strong
Blocking OAI-SearchBot / PerplexityBot / Claude-SearchBot	Structural exclusion from that engine’s citations	✗ Hard blocker, fix first
Unlinked brand mentions on third-party sites	Majority of tracked AI-citation footprint (BrandMentions, n≈410k mentions)	✓ Underused lever
llms.txt file	No effect on Google Search; consumed by Perplexity and Claude	~ Engine-dependent, not universal

A visual audit flow

flowchart TD
    A[Page not cited by AI] --> B{Can AI crawlers fetch it?}
    B -- No: robots.txt or WAF blocks retrieval bots --> C[Fix Layer 1: crawl access]
    B -- Yes --> D{Does the page answer the question in the first 100 words?}
    D -- No --> E[Fix Layer 2: structure and extractability]
    D -- Yes --> F{Are claims sourced with named links and stats?}
    F -- No --> G[Fix Layer 3: fact density and sourcing]
    F -- Yes --> H{Does the brand have entity and mention signals off-site?}
    H -- No --> I[Fix Layer 4: entity and trust signals]
    H -- Yes --> J[Re-audit monthly and after core updates]
    C --> K[Re-run audit]
    E --> K
    G --> K
    I --> K
    K --> A

How to score your audit

Run all 40 checks against one target page. Score each as pass or fail, no partial credit, a check either meets its criteria or it does not.

• 0-5 fails: You are close. Fix the specific fails, re-audit in 30 days.
• 6-15 fails: Pick one layer at a time, starting with Layer 1 (crawl access), since nothing else matters until a crawler can reach the page.
• 16+ fails: Treat this as a rebuild, not a patch, starting from Layer 1 through Layer 3 in order.

Layer 1 always comes first regardless of your score distribution. A page that fails crawl access checks gets zero benefit from perfect fact density.

If you are running this audit across dozens of pages at once, the bottleneck is usually not diagnosis, it is generating the fixed content fast enough to keep up with a growing site. Vrid.ai’s keyword research module surfaces the primary keyword and gap analysis you need before you rewrite a page for Layer 2 and Layer 3 fixes, so the audit and the content fix pull from the same keyword data instead of two disconnected tools.

Frequently asked questions

What is an AI visibility audit?

An AI visibility audit is a structured check of why a page does or does not get cited by ChatGPT, Perplexity, Google AI Overviews, or similar systems. It covers crawl access, content extractability, fact density, and entity signals, distinct from a traditional SEO audit’s focus on rankings and backlinks.

How is an AI visibility audit different from a technical SEO audit?

A technical SEO audit focuses on crawlability, indexation, and ranking factors for the traditional search index. An AI visibility audit adds checks specific to generative systems: AI crawler-specific robots.txt rules, extractable content structure, fact density per section, and off-site brand-mention corroboration, layers a standard SEO audit does not cover.

Do I need to block AI crawlers to protect my content?

Not if you want AI-search citations. A 2026 best practice, per Anagram’s crawler guide, blocks bulk training crawlers like GPTBot and ClaudeBot while allowing live retrieval bots like OAI-SearchBot, PerplexityBot, and Claude-SearchBot, opting out of training while staying eligible for citations.

Does adding schema markup improve my AI citation rate?

Ahrefs tracked 1,885 pages that added JSON-LD schema against 4,000 control pages and found no meaningful citation lift on already-cited pages, with results statistically indistinguishable from zero on AI Mode and ChatGPT. Schema still helps with rich results and entity disambiguation, it is not a proven citation lever.

Do I need llms.txt?

Google Search does not use llms.txt for ranking or AI Overviews, confirmed by Google’s John Mueller and Gary Illyes. Perplexity and Claude do consume it. Add it if you specifically target Perplexity or Claude citations; skip it if Google is your only concern.

How often should I re-run this audit?

Re-run it after every confirmed Google core update and after any unexplained traffic swing. Google confirmed two core updates in the first half of 2026, and third-party volatility trackers recorded additional movement outside those windows, so a quarterly baseline audit plus event-triggered re-checks covers both scheduled and unscheduled shifts.

Which layer should I fix first if I fail checks across all five?

Layer 1, crawl and access, always first. A page that AI crawlers cannot fetch gets no benefit from perfect content structure or fact density in Layers 2 and 3. Fix access, confirm the page is fetchable, then move to structure.

What is fact density and why does it matter?

Fact density is the concentration of sourced, statistic-backed, citable claims per section of a page. The Princeton GEO study found citing sources, adding quotations, and adding statistics each lifted AI visibility by roughly 30-40% against baseline, making it the single strongest documented lever in this audit.

Do unlinked brand mentions actually help AI visibility?

A BrandMentions study tracking about 410,000 public mentions across 240 brands over 90 days in 2026 found unlinked, text-only mentions made up the majority of the third-party footprint AI systems draw on. The study reports this as an observed correlation, not a proven causal mechanism, but it argues for treating unlinked mentions as a real signal worth earning, not ignoring.

Should I optimize for Reddit specifically?

Reddit was the single most-cited domain by Google AI Overviews and Perplexity from August 2024 through June 2025, and ChatGPT cites it in roughly 12% of US answers, per Sitebulb’s analysis. Genuine, non-promotional participation in relevant threads is a real citation lever, not a growth hack you can fake at scale without getting flagged.

How do I measure whether my audit fixes worked?

Set up GA4 with a referrer regex isolating AI-platform domains, log a baseline before changes, then track branded search volume, AI referral sessions, and citation rate across a fixed set of monitored prompts monthly. Without a documented baseline, you cannot separate your fix’s effect from normal fluctuation.

Does site speed matter for AI citations?

Google’s May 2026 generative AI guidance names fast, accessible websites as a foundational requirement for AI feature eligibility, not a separate AI-specific optimization. A slow site that already underperforms in traditional search will underperform in AI features for the same underlying reason.

Can I pass this audit with AI-generated content?

The audit does not test whether content is human or AI-written, it tests whether content is accessible, extractable, sourced, and corroborated. Google’s guidance focuses on quality and genuine expertise regardless of production method, so the same 40 checks apply either way.

What percentage of pages typically fail this audit on the first pass?

There is no universal published benchmark, run the audit yourself and track your own pass rate over time. Most sites in practice fail a small cluster of related checks (commonly one full layer) rather than failing uniformly across all 40, which is why layer-by-layer scoring is more useful than a single aggregate number.

Is AI Overviews CTR loss the same thing as an AI visibility problem?

They are related but distinct. CTR loss happens even when you are cited, since a synthesized answer can satisfy the searcher without a click. Ahrefs measured a 58% CTR reduction on the top-ranking page for AI-Overview-triggering keywords in February 2026. A visibility audit fixes whether you are cited at all; CTR recovery is a separate, harder problem tied to answer satisfaction.

Do comparison tables really get cited more than prose?

Comparative, listicle-style content accounts for roughly a third of AI citations by category, per the Princeton-affiliated GEO research. A table with named, measurable criteria is one of the highest-value single assets you can add to a page failing Layer 2 checks.

What tools do I need to run this audit?

Google Search Console (indexing and CTR data), GA4 with a custom referrer regex (AI traffic), your robots.txt and CDN bot-management dashboard (crawl access), and a manual content read-through with JavaScript disabled (extractability). No proprietary AEO platform is required to complete all 40 checks.

Should I run this audit on every page or a sample?

Start with your 10-20 highest-value pages, the ones you most need AI systems to cite. Fix layer-by-layer failures there, confirm the pattern, then apply the same fixes at template level across similar page types rather than auditing every URL individually.

Does my industry (YMYL, local, e-commerce) change which checks matter most?

The core 40 checks apply broadly, but weighting shifts. YMYL sites should weight Layer 4 (entity and trust) more heavily given Google’s expertise emphasis for sensitive topics. Local businesses should add GBP-specific entity consistency to check 27. E-commerce sites gain more from structured product data even though the Ahrefs schema study found limited citation lift for already-cited editorial content specifically.

Can one page fix everything, or do I need to fix templates?

A single page fix confirms the pattern works; it does not scale on its own. Once you identify which checks a page type fails (product pages, blog posts, comparison pages), apply the fix at the template level so every page of that type inherits it, rather than manually re-auditing each URL one at a time.

Key takeaways

• Run all 40 checks against one underperforming page first. Most sites fail a small, related cluster of checks, not all 40, and that cluster tells you exactly what to fix.
• Layer 1 (crawl access) always comes first. A blocked retrieval bot makes every other fix irrelevant for that engine.
• Fact density, cited sources, quotations, and statistics, is the strongest documented lever in this audit, with a 30-40% visibility lift in the Princeton-affiliated GEO study. Schema markup is not: Ahrefs found no meaningful citation lift on 1,885 already-cited pages that added JSON-LD.
• Unlinked brand mentions and Reddit participation are real, underused signals, not backlink substitutes you can ignore.
• Re-audit after every confirmed Google core update and after unexplained traffic swings, using a logged baseline so you can tell a real fix from normal fluctuation.

Run this audit on your next 10 target pages before you write a single new article. If keyword research and a clean gap analysis are the bottleneck once you know which pages to fix, Vrid.ai handles that research step alongside the content rewrite, so fixing Layer 2 and Layer 3 failures does not mean juggling a separate keyword tool.