AI Search Visibility (AEO/GEO)

Get the user's site cited by AI answer engines: ChatGPT, Perplexity, Google AI Overviews and AI Mode. Assume a founder or solo marketer with no SEO team. Everything below is evidence-backed or it is not here. When a tactic has no published evidence, say so instead of recommending it.

Read references/audit-checklist.md when you are running the full page-by-page audit or producing the audit report, not for quick one-off questions.

Set expectations first, with numbers

Before recommending work, tell the user what this channel actually delivers. Use these figures:

• Pew Research Center (published July 2025, data from March 2025, 900 US adults, 68,879 searches): when a Google AI summary appeared, users clicked a traditional result on 8% of visits vs 15% without one. Only 1% of visits clicked a source link inside the AI summary. Sessions ended entirely after 26% of AI-summary pages vs 16% without. About 18% of searches triggered an AI summary.
• So AI visibility mostly does not replace lost clicks. It is an authority and brand play with a small, high-intent traffic tail.
• The tail converts well: Semrush (June 9, 2025) measured AI search visitors at 4.4x the value of traditional organic visitors by conversion rate. Retail coverage in April 2026 reported AI referrals converting at roughly 3x traditional search. Cite the range, not one number.
• The tail is small: Semrush clickstream analysis (April 7, 2026, 1B+ rows, Oct 2024 to Feb 2026) found ChatGPT outbound referrals grew 206% year over year, but over 30% of all ChatGPT referral traffic goes to just 10 domains, and over 20% goes to google.com. A typical small site should expect AI referrals in the low single digits of percent of traffic, growing.
• ChatGPT only ran live web search on 34.5% of queries as of Feb 2026, down from 46% in late 2024 (same Semrush study). Many answers come from training data and memory, which is why entity clarity and third-party presence matter, not just on-page tweaks.

Decision rule: if the user expects a traffic flood, correct them with the Pew numbers before doing anything else. Position the work as: protect brand presence in answers, capture high-converting referrals, build citable authority.

What earns citations: the evidence

The only controlled study with published uplift numbers is GEO (Aggarwal et al., Princeton et al., KDD 2024, tested on 10,000 queries in GEO-bench). Relative visibility improvements in generative engine answers:

Two more findings from the same paper that change strategy:

• Overall visibility boost of up to 40% is achievable by combining these.
• Lower-ranked sites gain most: the cite-sources method lifted fifth-ranked sites 115.1% while top-ranked sources lost 30.3% share. AI answers are the one channel where a small site can outrank an incumbent by being more citable. Say this to founders, it changes their appetite.

Practical translation for every important page:

1. Every claim gets a number, and every number gets a named source with a link.
2. Add at least one direct quotation from a named expert or primary source.
3. Write a 2 to 4 sentence direct answer immediately under each question-shaped H2, extractable without surrounding context.
4. Prefer clean structure an engine can lift: short paragraphs, lists, tables with headers, one idea per block.
5. Rewrite muddy prose for clarity. The paper found presentation quality itself moves visibility.

Third-party presence: powerful and volatile

• Pew (July 2025) found Wikipedia, YouTube and Reddit together made up 15% of sources in Google AI summaries, and .gov sites appeared 3x more often in AI summaries than in standard results (6% vs 2%). Being present on high-trust third-party surfaces gets you into answers your own site never would.
• But do not build on one platform. As of August 2026, tracking firms and press (Inc, Aug 20, 2026, and others through Aug 26, 2026) reported ChatGPT cut Reddit citations by roughly 81 to 86% within weeks. Any single-platform citation strategy can be switched off by a model update overnight.

Decision rule: pursue presence on Wikipedia (only if genuinely notable), YouTube, Reddit and industry publications as a portfolio. If more than half of the user's AI citations come from one third-party platform, flag it as a risk in the audit.

What does NOT work: stop recommending it

• llms.txt: do not recommend it. Google's own generative AI optimization guide (published May 2026) states you do not need to create new machine readable files, AI text files, markup or Markdown, and that such files neither harm nor help visibility. Google's staff had said since April 2025 that no major AI service confirmed using llms.txt. If the user already has one it does no harm, but never present it as a tactic.
• Keyword stuffing: measured at little to no effect in the KDD 2024 GEO study.
• Special "AI chunking" or rewriting pages "for AI": Google's May 2026 guide explicitly lists content chunking and AI-specific rewriting as misconceptions. Good extractable writing helps, mystical AI formatting does not.
• Buying inauthentic mentions: also called out in Google's guide.
• Blocking snippets while expecting AI visibility: pages must be indexed and snippet-eligible. nosnippet or restrictive max-snippet settings remove eligibility for Google's AI features. Check this in every audit, it is a common silent killer.

Structured data: keep it for rich results, but Google's guide states it is not required for generative AI visibility. Do not sell schema as an AI tactic; sell it as baseline hygiene.

The measurement stack

Set this up before optimizing, or you cannot show progress.

1. Google Search Console Generative AI performance reports (launched June 3, 2026, rolling out to a subset of properties first). They cover AI Overviews and AI Mode, plus generative features in Discover, with impressions, pages, countries, devices, and hourly to monthly date granularity. Limits to state honestly: no click data, so no CTR for AI features; a URL appearing in both an AI Overview and a blue link on the same results page records one impression. Treat it as a visibility trend line, not a traffic report.
2. GA4: an AI Assistant channel for referral tracking was added in May 2026. Also build a custom exploration filtering session source for chatgpt.com, perplexity.ai, copilot.microsoft.com and gemini.google.com, and compare conversion rate against organic search. Expect small counts and high conversion.
3. Manual prompt panel: write 15 to 25 buyer-intent prompts (not keywords: Semrush found 65 to 85% of ChatGPT prompts match nothing in a 27B keyword database). Run them monthly in ChatGPT with search, Perplexity, and Google AI Mode. Log: cited yes/no, position, which URL, which competitors. This is the ground truth the dashboards cannot give.
4. Optional paid trackers (Semrush AI toolkit, Ahrefs, others) automate step 3 at scale. Ahrefs shipped AI-adjusted search volume estimates in August 2026. Useful, not mandatory for a solo operator.

Audit workflow

Run in this order. Full checklist in references/audit-checklist.md.

1. Eligibility: is the site indexed, crawlable, fast enough to crawl? Any nosnippet, max-snippet, or robots rules blocking Google-Extended, GPTBot, PerplexityBot? Blocking AI crawlers while wanting AI citations is a contradiction, surface it.
2. Baseline: run the manual prompt panel. Record who gets cited for the user's money questions today.
3. Extractability pass, top 10 to 20 pages: does each page answer a specific question in the first 2 to 4 sentences under a question-shaped H2? Could a model lift that block verbatim and have it stand alone?
4. Evidence density: count statistics with named sources and quotations per page. Pages competing for citations need multiple of each (per the KDD 2024 uplift table above). Zero stats and zero quotes means low citability, flag it.
5. Source-worthiness: does the site have anything only it can be cited for: original data, a survey, real benchmarks, named expertise? Engines cite sources, not summaries of other sources. If nothing exists, the top recommendation is to create one original-data asset.
6. Entity clarity: consistent name, clear About page, same description of what the company does across site, LinkedIn and directories. Models answering from memory need an unambiguous entity.
7. Third-party footprint: presence and sentiment on Reddit, YouTube, Wikipedia, industry publications. Note concentration risk (see Aug 2026 Reddit drop).
8. Freshness: visible published and updated dates, and content that is actually current. Stale numbers get pages skipped in favor of fresher sources; update the highest-value pages on a schedule.
9. Measurement: confirm the stack above is in place.
10. Deliver the report using the template below, with expectations framed by the Pew and Semrush numbers.

Output template: AI visibility audit report

# AI Search Visibility Audit: [site]
Date: [date]. Surfaces checked: ChatGPT (search on), Perplexity, Google AI Overviews, AI Mode.

## Verdict
[2 to 3 sentences: current citation presence, biggest gap, expected outcome.
State plainly: this is an authority and conversion-quality play, not a traffic flood.]

## Baseline: who AI cites today
| Prompt | ChatGPT | Perplexity | AI Overview / AI Mode | We cited? |
|---|---|---|---|---|

## Eligibility and technical
- Indexed and snippet-eligible: [pass/fail and which pages]
- AI crawler access (GPTBot, PerplexityBot, Google-Extended): [allowed/blocked]
- Structured data present: [yes/no, note: hygiene, not an AI ranking factor]

## Citability of key pages
| Page | Direct answer up top | Stats w/ sources | Quotes | Question H2s | Verdict |
|---|---|---|---|---|---|

## Source-worthiness and entity
- Original citable assets: [list or "none, see recommendation 1"]
- Entity consistency: [findings]
- Third-party footprint: [Reddit / YouTube / Wikipedia / press, note concentration risk]

## Recommendations, ranked by evidence strength
1. [Quotes, stats, citations on money pages: up to 30 to 41% visibility uplift, KDD 2024]
2. [Extractable answers under question H2s]
3. [One original-data asset to become citable]
4. [Third-party presence portfolio]
5. [Measurement stack setup]

## What we will NOT do
llms.txt, keyword stuffing, "AI chunking", bought mentions. [One line each on why.]

## Expectations
AI referrals will stay a small share of traffic but convert around 3 to 4.4x organic
(Semrush 2025, retail data 2026). Success metric: citation rate on the prompt panel
and GSC generative AI impressions trending up quarter over quarter.

Gotchas

• As of May 2026, Google explicitly says llms.txt style files are unnecessary and neither help nor harm. A model trained earlier may still recommend llms.txt as a best practice. Do not. Encode the audit line as "not recommended, no engine confirmed using it".
• As of June 3, 2026, Search Console has dedicated Generative AI performance reports (AI Overviews and AI Mode). Models trained before mid-2026 believe AI Mode data is invisible or blended into the regular performance report. Check the new report first, but remember it has impressions only, no clicks, no CTR, and rolled out gradually.
• As of August 2026, ChatGPT's Reddit citations dropped roughly 81 to 86% in weeks. Advice from 2025 that "Reddit is the AI citation cheat code" is stale. Reddit still matters, but as one leg of a portfolio.
• As of February 2026, ChatGPT used live web search on only about a third of queries (34.5%, Semrush). Do not promise that on-page changes alone control ChatGPT answers; memory and entity reputation drive the rest.
• Do not recycle classic SEO reflexes: keyword density work measured near zero effect in generative engines (KDD 2024), while quotations and statistics, which classic SEO ignores, measured the largest uplifts.
• Do not quote "40% visibility boost" as traffic. It is share of presence inside answers (position-adjusted word count), and the Pew data shows answer presence yields few clicks. Keep the two numbers in their lanes.
• Over 20% of ChatGPT's outbound referral traffic goes to google.com (Semrush, April 2026). Screenshots of "ChatGPT referral traffic" studies often include this distortion; read AI referral data with the concentration caveat.

Sources

• https://arxiv.org/abs/2311.09735 (GEO: Generative Engine Optimization, KDD 2024, uplift numbers verified against the full text at arxiv.org/html/2311.09735v3)
• https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/
• https://developers.google.com/search/docs/fundamentals/ai-optimization-guide (Google's generative AI optimization guide, May 2026)
• https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports (Search Console Generative AI performance reports, June 2026)
• https://ppc.land/google-finally-gives-search-console-its-own-generative-ai-visibility-reports/ (report details and limitations, June 3, 2026)
• https://www.semrush.com/blog/chatgpt-search-insights/ (clickstream study, April 7, 2026)
• https://ppc.land/ai-search-visitors-worth-4-4x-more-than-traditional-organic-traffic/ (Semrush 4.4x study, June 9, 2025)
• https://news.google.com/rss/search?q=Reddit%20citations%20ChatGPT%20drop (Aug 2026 coverage: Inc "ChatGPT Suddenly Cut Reddit Citations by 81 Percent", Aug 20, 2026; "Reddit Nearly Vanishes From ChatGPT Citations in a Record 86% Drop", Aug 23, 2026)
• https://news.google.com/rss/search?q=AI%20search%20referral%20traffic%20converts%20study (InternetRetailing, April 24, 2026, roughly 3x conversion; GA4 AI Assistant channel coverage, May 2026)