AI Visibility Spans Many Surfaces. Google’s Guide Covers Just One

What the data shows, and where Google’s guide fits inside it

AI visibility is being measured across hundreds of published studies, experiments, vendor analyses, and patents. ChatGPT. Claude. Perplexity. Copilot. Gemini outside Search. The mobile apps where the Graphite analysis put 83% of AI usage worldwide. Plus Google’s AI Overviews and AI Mode. The territory is large, the surfaces behave differently, and the evidence has been accumulating fast enough that anyone honest about the field knows the picture is still forming.

On May 15, Google published a guide for optimizing toward generative AI features on Google Search. The headline position: optimizing for generative AI search is optimizing for the search experience, and thus still SEO. The guide names AEO and GEO and dismisses them as separate disciplines. It tells you what to do, what not to do, and names two retrieval mechanisms by name. It’s clear, specific, and authoritative for the surfaces Google operates, per usual. SEO bros cheered.

It’s also one position inside a much larger empirical record and so the question has to be asked (and answered): how meaningful is that guidance for overall “AI Visibility” optimization and strategy work?

This piece walks the field’s evidence layer by layer. Where Google’s guide aligns with what’s measured, the piece says so. Where it diverges from what’s measured on non-Google surfaces, the piece says so. Where it’s silent on tactics or surfaces the data has things to say about, the piece names that too. The frame is the evidence, not the guide. Google has a vote on Google AI. The evidence has a vote on everything else, and most of where AI-mediated search is happening is everything else.

This is an empirical comparison, not a position in the SEO-vs-GEO discourse on X. The data doesn’t take that position either.

What Google’s guide actually claims (11 items)

What it tells you to do (4):

1. Create valuable non-commodity content for your audience
2. Build and maintain a clear technical structure (semantic HTML, JS SEO best practices, page experience, Search Console verification)
3. Optimize local business and ecommerce details (Merchant Center feeds, Business Profiles, Business Agent)
4. Explore agentic experiences (browser agents, Universal Commerce Protocol)

What it tells you NOT to do (5):

1. Create llms.txt files or other “special” AI markup
2. Chunk your content into tiny pieces for AI
3. Rewrite content specifically for AI systems
4. Seek inauthentic mentions across the web
5. Overfocus on structured data as an AI-citation lever (continue using it for rich results)

The two mechanisms it names (2):

1. Retrieval-augmented generation (RAG / grounding)
2. Query fan-out (concurrent related queries the model generates)

The shape of the field

The evidence on AI visibility sorts into four layers, each with its own mechanics.

1. Training layer. What a base model has absorbed during training. Frozen at the model’s cutoff. Earned through coverage in the corpora the model trained on.
2. Knowledge layer. The structured entity infrastructure that systems use to know who you are and how you connect to other things in the world. Knowledge Graph nodes, Shopping Graph entries, sameAs relationships, Wikipedia presence.
3. Retrieval layer. What happens at query time when a system fetches and ranks pages to ground an answer.
4. Surface layer. What the user actually sees and clicks. AIO panels, AI Mode results, ChatGPT citations, Perplexity sources, Gemini grounding cards.

Different signals matter at different layers. Different platforms use different layers differently. And the layers overlap. A brand strong in the training layer typically also has strong knowledge layer presence and tends to retrieve well, because the same conditions that drive one drive the others. But they aren’t the same thing, and a signal that moves the needle at one layer can do nothing or the opposite at another. That’s the architecture. (Internally we call these L1 through L4, but the working names do the work.)

The training layer

Training layer | What the model absorbed during pre-training Google’s guide: Mostly silent. Addresses inauthentic mentions only. Sources: 3. Britney Muller / Jeff Dean, Digital Bloom, Kevin Indig / Ahrefs.

Base language models are next-token predictors trained on a frozen corpus. They don’t store sources. Whatever a model “knows” about a brand is whatever the brand’s coverage looked like in the training data at cutoff. Britney Muller’s framing, corroborated by Jeff Dean of Google DeepMind on Latent Space, locks in the basic split: training is one optimization target, retrieval is another, and conflating them is the field’s most common category mistake.

The empirical picture on this layer is correlational and modest. The Digital Bloom analysis of 680 million citations found a 0.664 Spearman correlation between brand web mentions and AI citation rate, with brand search volume at 0.334. Ahrefs’ independent work landed near the same range; Kevin Indig at Growth Memo cites similar numbers. The convergent finding: mentions consistently outrank backlinks as a citation predictor. Causal evidence isn’t available yet.

Google’s one line on this layer (“seeking inauthentic mentions across the web isn’t as helpful as it might seem”) addresses manipulation, not the operational question. The operational question is what authentic coverage actually moves the needle on training-layer presence, and Google doesn’t own ChatGPT’s training corpus, doesn’t own Claude’s, can’t speak to the layer for non-Google surfaces. The layer is one of the largest in operational scope and one of the smallest in the guide’s coverage.

The knowledge layer

Knowledge layer | Structured entity infrastructure Google’s guide: Mixed. Aligned with the field on schema-as-AI-lever (it isn’t). Diverges from the cross-platform evidence on schema behavior (same intervention, opposite outcomes). Sources: 5. Ahrefs DiD, Fiorelli, Pedro Dias, Peham N=1, Natzir leak.

The knowledge layer is where the field has done the most recent thinking and where Google’s guide makes its most testable specific claim.

Four pieces of evidence converge on the same architectural reading. Ahrefs’ difference-in-differences natural experiment (Linehan, Guan, May 11) tracked 1,885 pages adding JSON-LD against matched controls and found no causal lift on already-cited pages. +2.4% on AI Mode, +2.2% on ChatGPT, −4.6% on AIO, all within statistical noise. Gianluca Fiorelli published the scope refinement the same day: schema operates at index time, during entity parsing and Knowledge Graph construction, well upstream of retrieval. Content schemas and entity-identity schemas do different work; pooling them hides the function. Fiorelli’s frame is schema as registration, not advertising. You don’t measure VAT registration ROI by 30-day revenue lift. It’s the prerequisite for everything downstream. Pedro Dias at The Inference made the architectural case a week earlier: transformers parse tokens, not schema; the vendor playbook promising structured data “ensures AI engines can parse and connect your content” is selling appearance-of-control because the parsing layer being imagined isn’t there. Pedro names specific vendor frameworks.

Then Peham’s finding broke the convergence open. Sitewide schema deployment in his N=1 experiment at OtterlyAI produced +1,500% citations on Google AIO, +377% on Google AI Mode, decreases on ChatGPT, Gemini, and Copilot, and zero on Perplexity. Same intervention, opposite outcomes across platforms. The architectural difference between Google AI’s retrieval pipelines (which use schema-fed entity data) and ChatGPT’s, Gemini’s, Copilot’s (which don’t, in the same way) shows up directly in the experiment.

Google’s guide arrives and aligns with the first four pieces. Structured data isn’t required for generative AI search; continue using it for rich results. That preserves Fiorelli’s index-time function, confirms the Ahrefs null result, and validates Pedro’s architectural critique. All clean.

Where it diverges is the Peham finding. The guide’s “don’t overfocus on schema” reads coherent on Google AI, where Peham measured +1,500% lift. Read by an operator whose buyers are in ChatGPT, Gemini, or Copilot, the same guidance points the wrong direction, because the same word (“schema”) describes a tactic that diverges across surfaces. Not because Google is wrong about Google AI. Because the guide is silent on the divergence.

Natzir surfaced a directional leak on May 13: a Google AI Mode response on a commercial query rendered an internal variable, marcas_identificadas: [Shopping]. If the reading holds, AI Mode routes commercial-intent queries through Shopping Graph entity extraction before normalizing against the Web Graph. One observation, no replication, but it suggests the knowledge layer for commercial queries decomposes further into general Knowledge Graph and a separate Shopping Graph pipeline. Google’s guide hints at this in its Merchant Center and Business Profile lines, but doesn’t spell out the architecture.

The retrieval layer

Retrieval layer | Query-time fetching and grounding Google’s guide: Aligned on Google AI surfaces. Diverges from cross-platform evidence on AI-specific rewriting tactics. Silent on non-Google retrieval architectures. Sources: 8. Seer Interactive, SISTRIX, AirOps + Indig, Yext, Bernard Huang / Clearscope, Dan Petrovic / Dejan, Cyrus Shepard / Zyppy Signal, OppAlerts.

This is where Google’s guide does its most specific work and where the field has the densest empirical evidence outside Google’s control. The divergence beat of the piece lives largely here.

Google’s guide names RAG and query fan-out as the two mechanisms. The description is accurate and well-described for Google AI surfaces, and the evidence on those surfaces confirms it. Seer Interactive’s longitudinal CTR study (53 brands, 5.47M queries, 2.43B impressions) found brand-cited AIO pages get +120% more clicks than uncited ones on the same SERP. The 2025 AIO CTR decline narrative reversed in Jan-Feb 2026 with an 85% rebound from the December floor. SISTRIX’s 100-million-keyword German market analysis confirmed the AIO presence effect at position 1, where CTR drops from 27% to 11% when an AI Overview is present. On Google AI surfaces specifically, the guide and the data tell the same story.

On non-Google surfaces the story changes in three ways the guide doesn’t address.

ChatGPT runs on retrieval rank, not domain authority. AirOps + Kevin Indig’s fan-out study (353,799 pages, 16,851 queries) found retrieval rank dominance is the single strongest predictor: 58.4% citation rate at rank 1, 14.2% at rank 10, holding content quality constant. Heading-to-query semantic match adds 19 percentage points on top of rank. Moderate fan-out coverage outperforms exhaustive. Domain authority shows no positive correlation with ChatGPT citation; it’s slightly inverse. Same-DA pages span from 2.4% to 59.2% citation rate depending on content fit. None of this looks like the AIO playbook.

Platforms have fundamentally different citation logic. Yext’s multi-million-citation dataset (the company sells listings management, worth noting alongside the category split) shows Gemini favors websites (52.1%), ChatGPT leans on listings (48.7%), Perplexity diversifies across third-party aggregators. Bernard Huang of Clearscope sharpened the underlying distinction: ChatGPT seeks to verify and update training data and cites official sites; Gemini seeks objective third-party sources and actively ignores first-party content. Both can be true at once. Superlines measured up to 615x citation volume difference between Grok and Claude for the same brand. Averi’s 680-million-citation analysis found only 11% of cited domains overlap between ChatGPT and Perplexity. A brand that wins on one platform can be invisible on another. Google’s guide is silent on this cross-platform variance because Google’s guide is about Google AI.

The “don’t rewrite for AI” line runs into documented retrieval tactics that are exactly that. Dan Petrovic’s work at Dejan on Google’s Gemini found the system uses a strict retrieval cap per URL, and content near the top of the page is significantly more likely to make the cut. That isn’t general “good writing.” It’s a structural restructuring move specific to how AI retrieval mechanics work. Cyrus Shepard’s synthesis (more on this below) identifies Answer Near the Top (8.8), Self-Contained Passages (8.0), Explicit Phrasing (8.1), and Factually Specific (8.3) as rewriting tactics with documented empirical support across studies. A practitioner who reads Google’s guide and concludes don’t restructure for AI is acting against the cross-platform evidence on what actually moves citations. The guide’s wording invites a reading the evidence doesn’t support. Not a contradiction on Google AI specifically. A divergence between broadly-worded guidance and measured cross-platform behavior.

Cyrus Shepard published a parallel synthesis on May 7, eight days before Google’s guide landed. He read 54 published studies, experiments, and patents on AI citations and normalized the findings into 23 named factors, scored by repeatability, evidence strength, and official support. His landing was convergent with Google’s: most of the strongest signals align with traditional SEO. The factors clustered into Relevance, Trust, Topical Authority, and Extractability.

Our piece draws on a wider source set, but the medium is the same. Published studies, experiments, vendor analyses, and practitioner work from researchers, PhD candidates, and marketing agencies, all blended together as study data. Cyrus did the taxonomy work the field needed someone to do. The visual below uses his 23 factors as the input axis. What it adds is sorting them against the four-layer architecture and the surfaces where the evidence actually applies. That sorting surfaces a pattern his ranked-list view doesn’t show: the field is over-indexed on retrieval-layer measurement, several of his factors diverge sharply across platforms, and Google’s guide aligns with him on what doesn’t work while diverging from his rewriting-tactic factors on the AI-specific work that does.

The largest cross-industry correlation study on ChatGPT recommendations sits behind both syntheses. OppAlerts measured 13 signals across 145 industries, 34,092 domains, and 1,595 buyer personas. The headline finding is the variance ceiling: no single signal explains more than 5.8% of variance pooled across industries. The strongest signals, ranked:

• Search Engine Appearances: Spearman ρ +0.241
• Best Search Engine Rank: ρ +0.238
• Backlinks: ρ +0.20
• PageRank: ρ +0.194
• Reddit Comments: ρ +0.111
• Wikipedia Citations: ρ +0.077

Per-industry weights diverge sharply. Wikidata is the dominant predictor for furniture stores (ρ +0.706) and collision-repair centers (ρ +0.723), and shows strong negative correlations in car-wash chains (Wikipedia Citations ρ −0.700) and garage-door services (Wikidata ρ −0.632). Same signal wins in one industry, hurts in another. There is no universal AI visibility lever, and the vendor frameworks promising single-pillar wins are selling something the empirical data doesn’t support.

The surface layer

Surface layer | What the user sees and clicks Google’s guide: Aligned on AIO query coverage. Silent on non-Google surface behavior, non-US dynamics, and platform localization. Sources: 4. Aleyda Solis, Peham localization, Seer CTR-vs-impression, Conductor.

Surface is where the contested takes about “AI search” actually live, because surface is what the user sees and what brands measure when they say their visibility moved.

Local infrastructure beats global defaults at the head of non-US AI search. Aleyda Solis’ 10-market analysis (87.6M visits, Similarweb) carries the strongest surface-level finding available right now: in most non-US markets, locally trusted domains, public sector entities, and infrastructure operators beat or compete with global defaults at the head of AI search. Bol.com beats Amazon.nl. MercadoLivre beats Amazon.com.br. Lefrecce beats Booking.com in Italian travel. Bahn.de leads German travel overall. The concentration gradient varies by an order of magnitude (in Italian ecommerce 2 domains take 50% of clicks; in UK travel it takes 129). The translate-and-hreflang international SEO model isn’t enough because the problem isn’t translation. It’s structured-inventory ownership and local entity presence.

Platforms localize at radically different rates. Perplexity essentially doesn’t. Peham’s localization experiment ran in this same territory. 100+ prompts across Finance, Insurance, and Pharma in 6 markets, English and local language. Localization Index: Copilot 77.4% local sources, Google AI Mode 68.5%, Google AIO 65.7%, ChatGPT 58.2%, Perplexity 9.3%. That single finding has more operational implication for a non-US brand than most of what’s been published this year.

CTR can compress because impressions doubled, not because clicks dropped. The Seer Interactive data surfaced this methodologically. When brand-cited informational CTR dropped 52% in October 2025, the raw numbers showed clicks were flat and impressions had doubled. The CTR compression was denominator-driven. More queries earned brand citations; the math of dividing by a bigger number made the rate look like it collapsed. CTR alone misleads. Decompose clicks and impressions before declaring anything has moved.

For Google’s surfaces specifically, the guide is the most authoritative document available. The Conductor 21.9-million-query analysis found AIOs now appear on 25% of Google searches, doubled in twelve months from 13%. A brand whose buyers search on Google should read the guide carefully and follow it. The surface layer is also where the scope mismatch between Google’s guide and where AI usage actually happens becomes hardest to ignore.

Where the evidence sits and Google’s guide doesn’t

Scope question | Surfaces and markets outside Google’s reach Google’s guide: Silent. By design. The guide covers Google AI surfaces accessible via web crawling. Sources: 3. Graphite mobile/app data, Mike King + Vercel/MERJ, G2 Answer Economy.

Three pieces of evidence make the scope question concrete.

83% of AI usage happens in mobile apps. That’s most of the territory. The Graphite analysis from Ethan Smith combined Similarweb web and mobile app data across the top 5 LLMs and top 6 search engines. AI usage is roughly 56% the size of search worldwide and 34% the size in the US. The mobile share is the headline: 83% worldwide, 75% in the US. Web-only estimates miss four to five times the actual usage.

AI crawlers don’t render JavaScript. That’s not positioning, that’s infrastructure. Mike King’s piece on cloaking for LLMs (May 14) anchors on this architectural fact. Vercel + MERJ analyzed 500 million GPTBot fetches and found zero JS execution. ClaudeBot fetches JS 24% of the time but doesn’t run it. 69% of major AI crawlers can’t render JS at all. Googlebot and Bingbot render. Everyone else gets initial HTML and moves on. The retrieval pipelines that feed ChatGPT, Claude, and the others are architecturally different from Google’s.

51% of B2B buyers now start research with AI chatbots, not Google. The G2 Answer Economy report (1,076 B2B buyers, March 2026) found this is up from a self-reported 29% a year prior. 69% chose a different vendor than initially planned based on AI guidance. 33% bought from a vendor they’d never heard of before AI surfaced them. ChatGPT holds 73% share at the discovery stage. None of this happens on a surface Google operates.

A guide that’s silent on 83% of AI usage and 51% of B2B buyer discovery is a guide about a slice of the territory. Useful for that slice. Not a guide about AI visibility.

The full evidence map

The piece walked roughly 23 distinct sources of evidence across the four layers and the scope question.

23 sources. 6 aligned. 3 partial / hints at. 3 diverges from. 11 silent.

What it adds up to

The evidence record on AI visibility is large, fast-moving, and measured across surfaces that mostly aren’t Google. ChatGPT, Claude, Perplexity, Copilot, Gemini outside Search, and the mobile apps where 83% of AI usage actually happens. The field has been measuring these surfaces and publishing what it finds for two years. Google’s guide, published May 15, is one position inside that record.

The position is authoritative on Google AI surfaces. The piece walked aligned findings on six points, mostly knowledge-layer schema critique and retrieval-layer behavior on AIO. Read inside its scope, the guide is the single most useful document on how Google AI surfaces work, according to them.

The position diverges from the field’s evidence in three named places.

1. Schema cross-platform. Google’s “don’t overfocus on schema as AI lever” is true on Google AI surfaces, where Peham measured +1,500% lift. On ChatGPT, Gemini, and Copilot the same intervention produces decreases. On Perplexity it produces zero. A non-Google operator reading the guide’s schema position as universal is reading against the cross-platform evidence.
2. AI-specific rewriting. Google’s “don’t rewrite for AI” runs into Dan Petrovic’s Gemini retrieval-cap finding, Cyrus Shepard’s documented factors (Answer Near the Top, Self-Contained Passages, Explicit Phrasing, Factually Specific), and the broader empirical work on what gets cited. These tactics are AI-specific structural choices the engines reward. The guide’s wording invites the wrong reading.
3. Cloaking for LLMs. Google’s general anti-cloaking position, extended to LLM crawlers, would forbid serving optimized content to AI bots. Mike King’s piece, anchored on Vercel + MERJ’s 500-million-GPTBot-fetch architectural data, argues the practice is a defensible response to crawler infrastructure that doesn’t render JavaScript. The guide doesn’t address LLM-specific cloaking directly, but the broader Google position is in tension with the architectural reality.

The position is silent on eleven findings the field has measured. ChatGPT retrieval rank dominance. Cross-platform citation source divergence. The 5.8% variance ceiling. Localization gaps across platforms. Non-US market dynamics. The 83% mobile-app usage share. AI crawler JS-rendering. B2B buyer behavior shifts. None of these are surfaces Google operates. The guide makes no claim to cover them. A reader who extends the guide to cover them is doing reader work the document didn’t authorize. Dun dun dun.

So here’s the placement. Google’s guide is a useful document about Google AI Overviews and AI Mode. The field has measured a much larger territory and continues to measure it weekly. On the territory Google doesn’t operate, the evidence has more to say than the guide does, sometimes the evidence says different things than the guide says, and most of where AI-mediated search is happening sits on that territory.

The conflation of Google’s guide with AI visibility as a whole is what the SEO discourse has been doing in the days since publication….and for….years. The data doesn’t support the conflation. Google has a vote on Google AI surfaces. It doesn’t have a vote on Gemini outside Search, ChatGPT, Claude, Perplexity, Copilot, or mobile apps. The field’s evidence has a vote on those, and the piece walked it.

The guide is a part. Treating any part as the whole is what produces strategy that works in one place and fails in five others. The piece walked the parts. The placement is the takeaway. AI Visibility is everywhere.

Sources

Google official gGoogle official guidance. Google Search Central, “Optimizing your website for generative AI features on Google Search,” last updated May 15 2026. https://developers.google.com/search/docs/fundamentals/ai-optimization-guide

Two-layer architecture (training vs retrieval). Britney Muller (@BritneyMuller), “How LLMs Actually Work,” Feb 2026. Corroborated: Jeff Dean (@JeffDean) / Google DeepMind, Latent Space podcast, Feb 2026. https://www.latent.space/p/jeff-dean

Brand mentions correlation. The Digital Bloom, 2025 AI Citation Report (680M+ citations). https://thedigitalbloom.com/learn/2025-ai-citation-llm-visibility-report/ Corroborated: Kevin Indig (@Kevin_Indig) / Growth Memo; Ahrefs (@ahrefs) Evolve 2025.

Schema natural experiment. Louise Linehan, Xibeijia Guan, Ryan Law (reviewer) / Ahrefs (@ahrefs), “Does Schema Markup Drive AI Citations?” May 11 2026. https://ahrefs.com/blog/schema-ai-citations/

searchVIU schema retrieval experiment. https://www.searchviu.com/en/schema-markup-and-ai-in-2025-what-chatgpt-claude-perplexity-gemini-really-see/

Schema scope refinement. Gianluca Fiorelli (@gfiorelli1), “The Ahrefs Schema Study is Right. And It’s Testing the Wrong Thing,” I Love SEO, May 11 2026. https://www.iloveseo.net/the-ahrefs-schema-study-is-right-and-its-testing-the-wrong-thing/

Vendor playbook architectural critique. Pedro Dias (@pedrodias) / The Inference, “The Whole Point Was the Mess,” May 5 2026. https://theinference.io/p/the-whole-point-was-the-mess

“GEO is just evolved SEO” position. Lily Ray (@lilyraynyc) / Amsive, Search Engine Land. https://searchengineland.com/how-to-get-cited-by-chatgpt-the-content-traits-llms-quote-most-464868

OtterlyAI BrightonSEO experiments (schema cross-platform, llms.txt, Markdown, localization, YouTube). Thomas Peham (@thomaspeham) / OtterlyAI, BrightonSEO 2026, April 29 2026. https://speakerdeck.com/thomaspeham/geo-experiments-2026-what-we-tested-what-failed-and-what-actually-works

Google AI Mode Shopping Graph leak. Natzir (@natzir9), X post May 13 2026. https://x.com/natzir9

AIO CTR longitudinal study. Tracy McDonald, Hannah Cooley, Marketa Williams / Seer Interactive (@SeerInteractive), “AIO Impact on Google CTR — 2026 Update,” April 24 2026. https://www.seerinteractive.com/insights/aio-impact-on-google-ctr-2026-update

SISTRIX German market AIO analysis. Johannes Beus / SISTRIX (@sistrix), 100M+ keywords, Feb-Mar 2026. https://sistrix.com/blog/

Fan-out effect study. AirOps + Kevin Indig (@Kevin_Indig), “The Fan-out Effect,” April 2026 (353,799 pages, 16,851 queries). https://airops.com/research/the-fanout-effect

Cross-industry ChatGPT recommendation correlation study. OppAlerts, March 2026 (145 industries, 34,092 domains, 1,595 buyer personas). https://oppalerts.com/LLM-Ranking-Factors/

AI citation factor synthesis (54 studies, 23 factors). Cyrus Shepard (@CyrusShepard) / Zyppy Signal, “AI Citation Ranking Factors Analysis,” May 7 2026. https://signal.zyppy.com/p/ai-citation-ranking-factors

Gemini retrieval cap per URL. Dan Petrovic (@dejanseo) / Dejan, “SRO Grounding Snippets.” https://dejan.ai/blog/sro-grounding-snippets/

Citation source type distribution. Yext, “AI Citations Release,” October 2025 (6.8M citations). https://www.yext.com/about/news-media/ai-citations-release

ChatGPT vs Gemini citation logic. Bernard Huang (@bernardjhuang) / Clearscope, Siege Media podcast, January 2026.

Cross-platform variance. Superlines, March 2026; Averi 680M citations, March 2026.

10-market AI search analysis. Aleyda Solis (@aleyda), “Global AI Search Strategy,” April 29 2026 (87.6M visits, Similarweb). https://www.aleydasolis.com/en/ai-search/global-ai-search-strategy/

Market sizing including mobile apps. Ethan Smith (@ethan_graphite) / Graphite, “AI Is Much Bigger Than You Think,” March 2026. https://graphite.io/five-percent/ai-is-much-bigger-than-you-think

AIO query coverage. Conductor, 21.9M queries, March 2026. https://www.conductor.com/academy/information-technology-aeo-geo-benchmarks/

Cloaking for LLMs / AI crawler JS rendering. Mike King (@iPullRank) / iPullRank, “The Case for Cloaking for Large Language Models,” May 14 2026. https://ipullrank.com/cloaking-for-llms Underlying: Vercel + MERJ analysis (https://vercel.com/blog/the-rise-of-the-ai-crawler), Cloudflare Markdown for Agents (https://blog.cloudflare.com/markdown-for-agents/).

B2B buyer behavior. Tim Sanders / G2 (@G2dotcom), “The Answer Economy: How AI Search Is Rewiring B2B Software Buying,” April 2026 (1,076 buyers). https://learn.g2.com/g2-2026-ai-search-insight-reportuidance. Google Search Central, “Optimizing your website for generative AI features on Google Search,” last updated May 15 2026. https://developers.google.com/search/docs/fundamentals/ai-optimization-guide

Two-layer architecture (training vs retrieval). Britney Muller, “How LLMs Actually Work,” Feb 2026. Corroborated: Jeff Dean / Google DeepMind, Latent Space podcast, Feb 2026. https://www.latent.space/p/jeff-dean

Brand mentions correlation. The Digital Bloom, 2025 AI Citation Report (680M+ citations). https://thedigitalbloom.com/learn/2025-ai-citation-llm-visibility-report/ Corroborated: Kevin Indig / Growth Memo; Ahrefs Evolve 2025.

Schema natural experiment. Louise Linehan, Xibeijia Guan, Ryan Law (reviewer) / Ahrefs, “Does Schema Markup Drive AI Citations?” May 11 2026. https://ahrefs.com/blog/schema-ai-citations/

searchVIU schema retrieval experiment. https://www.searchviu.com/en/schema-markup-and-ai-in-2025-what-chatgpt-claude-perplexity-gemini-really-see/

Schema scope refinement. Gianluca Fiorelli, “The Ahrefs Schema Study is Right. And It’s Testing the Wrong Thing,” I Love SEO, May 11 2026. https://www.iloveseo.net/the-ahrefs-schema-study-is-right-and-its-testing-the-wrong-thing/

Vendor playbook architectural critique. Pedro Dias / The Inference, “The Whole Point Was the Mess,” May 5 2026. https://theinference.io/p/the-whole-point-was-the-mess

“GEO is just evolved SEO” position. Lily Ray / Amsive, Search Engine Land. https://searchengineland.com/how-to-get-cited-by-chatgpt-the-content-traits-llms-quote-most-464868

OtterlyAI BrightonSEO experiments (schema cross-platform, llms.txt, Markdown, localization, YouTube). Thomas Peham / OtterlyAI, BrightonSEO 2026, April 29 2026. https://speakerdeck.com/thomaspeham/geo-experiments-2026-what-we-tested-what-failed-and-what-actually-works

Google AI Mode Shopping Graph leak. Natzir (@natzir9), X post May 13 2026. https://x.com/natzir9

AIO CTR longitudinal study. Tracy McDonald, Hannah Cooley, Marketa Williams / Seer Interactive, “AIO Impact on Google CTR — 2026 Update,” April 24 2026. https://www.seerinteractive.com/insights/aio-impact-on-google-ctr-2026-update

SISTRIX German market AIO analysis. Johannes Beus / SISTRIX, 100M+ keywords, Feb-Mar 2026. https://sistrix.com/blog/

Fan-out effect study. AirOps + Kevin Indig, “The Fan-out Effect,” April 2026 (353,799 pages, 16,851 queries). https://airops.com/research/the-fanout-effect

Cross-industry ChatGPT recommendation correlation study. OppAlerts, March 2026 (145 industries, 34,092 domains, 1,595 buyer personas). https://oppalerts.com/LLM-Ranking-Factors/

AI citation factor synthesis (54 studies, 23 factors). Cyrus Shepard / Zyppy Signal, “AI Citation Ranking Factors Analysis,” May 7 2026. https://signal.zyppy.com/p/ai-citation-ranking-factors

Gemini retrieval cap per URL. Dan Petrovic / Dejan, “SRO Grounding Snippets.” https://dejan.ai/blog/sro-grounding-snippets/

Citation source type distribution. Yext, “AI Citations Release,” October 2025 (6.8M citations). https://www.yext.com/about/news-media/ai-citations-release

ChatGPT vs Gemini citation logic. Bernard Huang / Clearscope, Siege Media podcast, January 2026.

Cross-platform variance. Superlines, March 2026; Averi 680M citations, March 2026.

10-market AI search analysis. Aleyda Solis, “Global AI Search Strategy,” April 29 2026 (87.6M visits, Similarweb). https://www.aleydasolis.com/en/ai-search/global-ai-search-strategy/

Market sizing including mobile apps. Ethan Smith / Graphite, “AI Is Much Bigger Than You Think,” March 2026. https://graphite.io/five-percent/ai-is-much-bigger-than-you-think

AIO query coverage. Conductor, 21.9M queries, March 2026. https://www.conductor.com/academy/information-technology-aeo-geo-benchmarks/

Cloaking for LLMs / AI crawler JS rendering. Mike King / @iPullRank , “The Case for Cloaking for Large Language Models,” May 14 2026. https://ipullrank.com/cloaking-for-llms Underlying: Vercel + MERJ analysis (https://vercel.com/blog/the-rise-of-the-ai-crawler), Cloudflare Markdown for Agents (https://blog.cloudflare.com/markdown-for-agents/).

B2B buyer behavior. Tim Sanders / G2, “The Answer Economy: How AI Search Is Rewiring B2B Software Buying,” April 2026 (1,076 buyers). https://learn.g2.com/g2-2026-ai-search-insight-report