How AI Search Works
1
Attention Is All You Need: The Google Paper That Accidentally Ended Google's Search Monopoly.
On June 12, 2017, eight researchers at Google Brain uploaded a 15-page paper to arXiv. It accidentally made Google's own search business obsolete.
2
How Large Language Models Actually Work: Tokens, Context Windows, and Why Your Content Gets Ignored.
The answer to "how do I get cited by AI search?" is technical, not strategic — rooted in how large language models actually process information.
3
What Is RAG: the Only Thing That Matters for AI Search.
Once you see how Retrieval-Augmented Generation works, the rest of AI search visibility clicks into place. It's the mechanism that connects your content to every answer these systems produce.
4
The AI Search Race: ChatGPT, Gemini, Claude, Copilot, and Perplexity. A Technical Timeline.
The models, the milestones, and the architectural decisions that shaped every AI search product in market — a reference for anyone trying to make sense of where this came from.
5
How Google AI Overviews and AI Mode Actually Work: Why Your SEO Rankings Don't Guarantee AI Visibility.
AI Overviews aren't a ranking bonus layered on top of existing positions. In May 2024, Google restructured how most people receive information — the blue links became supplementary.
6
How to Be Surfaced in AI Search: What the Platforms' Own Documentation Actually Says.
Most GEO content produces a list of tactics and implies they work across all platforms. This article goes directly to what each platform has actually published about how its own system works.
7
Why the Bing AI Performance Report Is the Most Underused Tool in GEO.
Microsoft's Bing AI Performance report is the only first-party citation dashboard from any AI search platform. Most practitioners aren't using it.
8
9
Constitutional AI vs RLHF: Why the Alignment Method Affects What Gets Cited.
ChatGPT is aligned with RLHF. Claude is aligned with Constitutional AI. These aren't interchangeable — and the difference affects what each model treats as trustworthy and citable.
Writing  ·  Series: How AI Search Works  ·  Follow-on 2

The GEO Documentation Gap: Why Three Platforms Haven't Told You How to Rank in Their Systems.

Google published a guide. Microsoft published measurement tooling. Anthropic, OpenAI, and Perplexity have published nothing equivalent. Understanding why is more useful than waiting.

Google published its first explicit AI search optimisation guide in May 2026. Microsoft launched citation measurement tooling in February 2026. Both documents are notable not just for what they say, but for the fact that they exist at all.

Anthropic has published nothing equivalent. OpenAI has published nothing equivalent. Perplexity has published nothing equivalent. This is not an accident, and it is not a temporary gap that will be filled soon. Understanding why these platforms haven't published publisher-facing optimisation guidance — and what that absence implies for how to approach them — is more useful than waiting for guidance that may never arrive.

Why Google published a guide.

Google published an AI search optimisation guide because it had to.

Google's relationship with publishers is structurally adversarial in a way that OpenAI's, Anthropic's, and Perplexity's are not — yet. Google's core business is built on indexing and surfacing publisher content. Publishers have always had the ability to de-index their content from Google if they chose. The rise of AI Overviews intensified this tension: if Google's AI generates a synthesised answer that satisfies the user's query without a click to the publisher, the publisher's traffic decreases even as their content contributes to the AI's response.

Publishers need a reason to remain indexed and crawlable by Google. A guide that explains how to be cited in AI responses — turning AI Overviews from a traffic threat into a visibility opportunity — is a strategic response to that tension. The guide also serves Google's legal and regulatory interests: publishing explicit guidance creates a record of transparency about how AI features use web content.

Microsoft published the Bing AI Performance report for analogous reasons. Bing has always had a smaller publisher relationship to manage than Google, but the Copilot integration into Microsoft 365 — used by hundreds of millions of enterprise users — creates a new scale of publisher dependency. The measurement tool is a gesture of transparency that supports the publisher relationship.

Why OpenAI has not published a guide.

OpenAI's relationship with publishers is different. ChatGPT's training data — the source of its parametric knowledge — has been the subject of copyright litigation from publishers including The New York Times. OpenAI is in active legal and commercial negotiation with publishers over training data licensing, not optimisation guidance.

Publishing an explicit guide to getting content cited in ChatGPT Search would implicitly acknowledge the degree to which ChatGPT depends on publisher content and could be used in ongoing legal proceedings. It would also invite regulatory scrutiny of the boundary between licensed and unlicensed use of publisher content in AI training and retrieval.

The documentation OpenAI has published is developer-focused: API documentation explaining how the web search tool works technically. This is useful for building on top of ChatGPT, not for publishers trying to improve citation frequency. The asymmetry is intentional.

Why Anthropic has not published a guide.

Anthropic's situation is different again. Anthropic is primarily a research and API company — its customer base is developers and enterprises building on Claude, not publishers optimising for consumer AI search citation. The product surface where citation visibility matters (claude.ai with web search enabled) is newer and smaller than Google's AI Overviews or ChatGPT.

More importantly: Anthropic's Constitutional AI framework positions Claude's outputs as a function of explicit values rather than publisher optimisation incentives. Publishing guidance on how to get content cited by Claude would be in mild tension with the Constitutional AI premise — that Claude cites content because it is the best answer to the user's query, not because the publisher has optimised for citation.

Anthropic has published its retrieval architecture (web search documentation, Citations API documentation) and its alignment principles (Claude's Constitution). These are the available signals. Publisher-facing optimisation guidance is not currently part of Anthropic's product philosophy.

Why Perplexity has not published a guide.

Perplexity sits in an interesting position. It has no legacy search index to protect, no established publisher relationship to manage, and no enterprise software business creating complex data governance obligations. Its relationship with publishers is the most openly acknowledged of the three: the publisher revenue-sharing programme is a direct attempt to create economic alignment between citation and compensation.

The absence of explicit optimisation guidance from Perplexity likely reflects the company's size and prioritisation rather than a strategic decision to withhold it. Perplexity is a small company relative to Google, Microsoft, and OpenAI — building a revenue-sharing programme and growing to 1 billion monthly queries in under three years has consumed resources that might otherwise go toward publisher-facing documentation.

The Sonar documentation's explicit framing around factuality and readability as optimisation targets is the closest thing Perplexity has published to content guidance. It is directionally useful even without being a dedicated optimisation guide.

What the documentation gap means in practice.

The absence of explicit guidance from three of the five major AI search platforms has a practical consequence: practitioners who understand the underlying architecture — the RAG pipeline, the passage-level retrieval mechanism, the semantic similarity basis for retrieval — are not disadvantaged by the absence of platform guidance. They have derived the requirements from first principles.

The requirements that emerge from the architecture are consistent across all five platforms: content must be crawlable and indexable; content must be factually precise and clearly attributed; content must be structured for passage-level extraction; content must provide non-commodity value that the model cannot generate without consulting the source.

These requirements do not change because a platform has or has not published a guide confirming them. They are properties of the retrieval mechanism. The practitioner who understands the mechanism does not need the guide.

The documentation gap also creates a measurement gap. Without platform-provided citation data, the available proxies — traffic from AI-referring domains, brand mention tracking, manual query-by-query checking — are indirect and labour-intensive. The Bing AI Performance report is the only exception. Until Google, OpenAI, Anthropic, and Perplexity provide equivalent tooling, measurement will remain harder than optimisation.

The GEO documentation gap will narrow. Google has shown what explicit guidance looks like. Microsoft has shown what citation measurement looks like. The competitive pressure between platforms will push the others toward greater transparency. The practitioners who have invested in understanding the architecture rather than waiting for platform guides will not be caught flat-footed when the guides arrive. They will already know what the guides say.

/ Next in the series

Article 9 is the most technically ambitious piece in this series: why ChatGPT's RLHF alignment and Claude's Constitutional AI produce different citation preferences — and what that means at the margin. Read Article 9 →

SOURCES
  1. Google Search Central, "Optimizing for Generative AI Features on Google Search," 2026 — developers.google.com/search/docs/fundamentals/ai-optimization-guide
  2. Microsoft Bing, "Introducing AI Performance in Bing Webmaster Tools," February 2026 — blogs.bing.com/webmaster/February-2026/Introducing-AI-Performance-in-Bing-Webmaster-Tools-Public-Preview
  3. Anthropic, "Introducing Web Search on the Anthropic API," 2025 — anthropic.com/news/web-search-api
  4. Anthropic, "Introducing Citations on the Anthropic API," 2025 — anthropic.com/news/introducing-citations-api
  5. Anthropic, "Claude's Constitution," 2022 — anthropic.com/news/claudes-constitution
  6. OpenAI, "Introducing ChatGPT Search," 2024 — openai.com/index/introducing-chatgpt-search
  7. OpenAI API Documentation, "Web Search" — platform.openai.com/docs/guides/tools-web-search
  8. Perplexity, "Sonar Models Documentation" — docs.perplexity.ai/docs/sonar/models
  9. Perplexity, "Introducing the Perplexity Publishers' Program," 2024 — perplexity.ai/hub/blog/introducing-the-perplexity-publishers-program
How AI Search Works

The full series.

1
Available now
Attention Is All You Need: The Google Paper That Accidentally Ended Google's Search Monopoly
2
Available now
How Large Language Models Actually Work: Tokens, Context Windows, and Why Your Content Gets Ignored
3
Available now
What Is RAG: the Only Thing That Matters for AI Search
4
Available now
The AI Search Race: ChatGPT, Gemini, Claude, Copilot, and Perplexity. A Technical Timeline
5
Available now
How Google AI Overviews and AI Mode Actually Work: Why Your SEO Rankings Don't Guarantee AI Visibility
6
Available now
How to Be Surfaced in AI Search: What the Providers' Own Documentation Actually Says
7
Available now
Why the Bing AI Performance Report Is the Most Underused Tool in GEO
8
You are here
The GEO Documentation Gap: Why Anthropic, OpenAI, and Perplexity Haven't Told You How to Rank in Their Systems
9
Available now
Constitutional AI vs RLHF: Why the Alignment Method Affects What Gets Cited
tc
/ Written by

Thomas Cox

Twelve years in B2B SEO, most recently at VP level. Now independent — helping companies stay discoverable as buyer search moves into ChatGPT, Perplexity, and Google AI Overviews. Remote · UK.

Want this in your inbox?

Roughly one essay a month. No drip sequences, no upsells.

Subscribe