Executive summary
In 2026 the goal is no longer to rank but to be one of the handful of sources an AI engine chooses to cite, and the evidence says that choice rewards substance over markup. Roughly 30 domains take two thirds of ChatGPT's citations in a topic, AI summaries halve clicks on Google, and the tactics most heavily sold as GEO (schema for citations, llms.txt, chunking) are the ones 2026 data contradicts.
Seven principles carry the rest of this guide:
- Win the sub-questions, not the keyword. Engines fan one prompt out into many hidden queries; cover a topic from every angle a buyer would ask.
- Make specific, attributed claims. Statistics, quotations and cited sources lifted visibility by up to 40% in the only controlled experiment.
- Put the answer first. A short answer under each question-shaped heading, in the top third of the page, is where citations cluster.
- Publish what only you know. Google's 2026 guidance centers on first-hand, non-commodity content; that requires real expert input, not a persona.
- Be talked about elsewhere. Brand mentions across the web are the strongest measured correlate of AI visibility.
- Be readable without JavaScript. The crawlers behind ChatGPT, Claude and Perplexity read raw HTML only.
- Open a door for agents. MCP and MCP Apps are production standards; discovery and browser-agent standards are still drafts.
Measure share of answers per engine, not just clicks, and grade every tactic by its evidence before you scale it.
How AI search works in 2026
Every major answer engine now runs the same loop: split the question into sub-queries, retrieve candidate pages from a search index, read them, and cite only the few passages that best support each claim. Google documents this openly: AI Overviews and AI Mode ground answers with retrieval-augmented generation on its core ranking systems, and use query fan-out, a set of concurrent related queries the model writes itself (Google, Optimizing for generative AI).
One prompt becomes several hidden sub-queries; the engine reads many pages per sub-query and cites only the passages that answer one cleanly.
The scale is no longer marginal. Google's AI Overviews reach roughly 2.5 billion monthly users and ChatGPT about 1 billion, per Evertune's mid-2026 aggregation (Evertune).
Four consequences shape everything else in this guide.
- You compete for questions you never see. Google's own example: a query about fixing a weedy lawn fans out into sub-queries about herbicides, chemical-free removal and prevention. In ChatGPT, 89.6% of searches triggered two or more fan-out queries (AirOps). A third of cited pages were found through fan-out queries, and 95% of those had zero search volume (Kevin Indig via Search Engine Land).
- Retrieval is not citation. ChatGPT retrieves about six times as many pages as it cites, and 85% of retrieved pages are never cited. The bottleneck is selection: the model keeps the passage that answers one sub-question cleanly and drops the rest.
- The index is still the gate. A page must be indexed and snippet-eligible to appear in Google's AI features, with no additional technical requirements (Google, AI features). Pages ranking first on Google were cited by ChatGPT 43.2% of the time, 3.5 times more often than pages beyond the top 20.
- Not every answer searches. Evertune measured ChatGPT answering from training knowledge 62% of the time and Gemini 54% (September 2025). Brand presence across the wider web shapes what models already know about you before any retrieval happens.
Citation seats are scarce and concentrated. Roughly 30 domains capture 67% of ChatGPT citations within a topic, and 58% of cited URLs are cited only once. Engines also disagree: the same questions produced 838 unique URLs on Perplexity versus 6,687 on Google AI Mode, and Reddit supplies about one in five Perplexity citations. Treating AI search as one channel is a mistake; each engine is its own market.
What the research says drives AI citations
The evidence converges on six drivers: verifiable specifics, off-site brand presence, topical coverage, answer-first placement, title-to-question fit and freshness. None of them is a markup trick, and all of them reward the same thing Google now calls non-commodity content.
| Driver | What the data shows | Source | Evidence type |
|---|---|---|---|
| Statistics, quotations, cited sources | Adding them lifted source visibility in generative answers by up to 40%. Lower-ranked pages gained most; keyword stuffing did not help. | Aggarwal et al., GEO, KDD 2024 | Controlled experiment |
| Brand mentions across the web | Branded web mentions had the strongest correlation (0.664) with AI Overview brand visibility across 75,000 brands. The top quartile averaged 169 AI Overview mentions, the next quartile 14. 26% of brands had none. | Ahrefs | Large correlation study |
| Topical coverage and depth | About 30 domains take 67% of citations per topic. Pages over 20,000 characters averaged 10.18 citations versus 2.39 under 500; Finance was the exception, where dense short pages won. | Indig via Search Engine Land | Large correlation study |
| Answer-first placement | ChatGPT cites most from the 10 to 20% band of a page. The bottom 10% earned 2.4 to 4.4% of citations; conclusions were largely ignored. | Indig via Search Engine Land | Large correlation study |
| Title matches the question | Pages with 50% or more title-query overlap had a 20.1% citation rate versus 15% overall for retrieved pages. | AirOps | Vendor study |
| Freshness | AI-cited URLs averaged 1,064 days old versus 1,432 in organic results, 25.7% fresher. AI Overviews were the exception. | Ahrefs | Large correlation study (17M citations) |
| Earned media and list formats | 32% of AI-cited domains are earned media, from 7% to 87% by topic. Half of ChatGPT citations in product categories are listicles. | Evertune | Vendor study |
Read this table with two cautions. First, only the GEO paper is a controlled experiment; everything else is correlation, mostly measured by companies that sell visibility tools. Second, the drivers overlap: a brand that publishes original data earns mentions, links and freshness at the same time, so no single factor should be optimized in isolation.
The practical synthesis is simple. Engines cite the page that makes a specific, checkable claim near the top, from a source the rest of the web already talks about, on a topic the site covers from every angle.
E-E-A-T for machines
E-E-A-T earns AI citations only when it is real and checkable: a named person with a footprint beyond your own site, contributing experience a model could not have produced on its own. Google's July 2026 guidance makes first-hand experience the main differentiator, contrasting a first-hand review that offers a unique perspective with a summary that merely restates what already exists (Google).
The same guidance names the enemy: commodity content. Its example of commodity is a generic tips list for first-time homebuyers; its example of non-commodity is a specific, experienced account of one decision and what it cost. Information gain is not a score you add afterwards. It is the part of the article only your people could have written.
| Dimension | What a machine can check | How to build it |
|---|---|---|
| Experience | Specific details, own photos and numbers, outcomes, mistakes, dates | Capture the expert's real input first: interview, voice note, case data. Generate around it, never instead of it |
| Expertise | Author entity with role, credentials and history across the web | Author page per expert, Person markup linking to profiles, talks and publications elsewhere |
| Authoritativeness | Independent mentions and citations of the brand and the person | Original research, data others quote, PR and community presence (see the brand-mention evidence above) |
| Trust | Accuracy, cited sources, visible dates, disclosed method | Fact-check every claim, cite primary sources, show updated dates, say how the content was made |
The ambassador rule
Writing in the first-person voice of a real expert who did not supply the experience is fabricated E-E-A-T. It fails the trust test the moment a claim is wrong, and it is exactly what Google warns about: generative models predict likely words rather than retrieve facts, so every AI-generated draft needs manual fact-checking before publishing (Google, generative AI content). Google also suggests telling readers how automated content was produced.
The working model is expert-in-the-loop: the ambassador contributes raw experience, the system structures and drafts, and the ambassador signs off. That sign-off is the asset. Without it, an ambassador vault is a liability rather than an authority signal.
Dual-layer content: one page, two readers
The machine layer that earns citations is the visible one. When ChatGPT, Claude, Perplexity, Gemini and AI Mode fetched pages live in a searchVIU test, every system extracted only visible HTML and ignored JSON-LD (searchVIU, reported by Ahrefs). So the dual layer is not prose for humans plus hidden data for bots. It is one page whose structure serves both.
The strongest structural signal measured so far is the answer capsule: a short, self-contained answer placed right under a question-shaped heading. In a 15-domain audit, 72.4% of ChatGPT-cited blog posts had one, and the best-performing posts (34.3%) paired a capsule with original or owned insight. Only 13.2% of cited posts had neither (Search Engine Land). The sample is small, but it agrees with the GEO experiment and the page-position data.
Anatomy of a dual-layer article, top to bottom:
- A title that matches how the question is asked. Title-query overlap of 50% or more lifted citation rates in the AirOps data.
- A real byline and a visible date. Named expert, role, link to an author page, last-updated date, and one line on how the piece was made.
- An answer capsule under every H2. One or two sentences, specific, link-free, written to survive being quoted out of context.
- The owned insight. The expert's own example, number, decision or mistake. This is the information gain; place it in the top third, where citations cluster.
- A visible fact table. Specifications, comparisons and figures in an HTML table or definition list, with units and sources.
- Inline sources. Statistics and quotations attributed to primary sources, the tactic with the strongest experimental support.
- Structured data that mirrors the page. Article, Person and Organization markup that matches the visible text. Treat it as hygiene for rich results and entity clarity, not as a citation lever.
Two cautions keep this honest. Google states that chunking content into tiny pieces is not required and that there is no ideal page length (Google). Capsules work because they serve skimming humans too; once they become a template stamped onto every paragraph, the page reads as machine-made and loses the trust it was meant to earn.
Machine interfaces: what is proven and what is speculative
Server-rendered HTML and open crawler access matter far more than any AI-specific file. The major AI crawlers do not execute JavaScript, and 97% of published llms.txt files received no requests at all in May 2026.
| Interface | What the 2026 evidence says | Verdict |
|---|---|---|
| Server-rendered HTML | Across hundreds of millions of fetches, GPTBot, OAI-SearchBot, ClaudeBot and PerplexityBot downloaded JavaScript files but never ran them (Vercel and MERJ). Client-rendered pages arrive as empty shells. | Mandatory |
| robots.txt and CDN access | Google treats Googlebot rules as the control for AI Overviews and AI Mode; Google-Extended covers other Google AI systems (Google). OpenAI, Anthropic and Perplexity run separate search and training bots. | Mandatory: allow search bots, decide on training bots separately |
| Schema.org JSON-LD | A matched study of 1,885 pages found no meaningful citation lift in ChatGPT (+2.2%) or AI Mode (+2.4%), and a small decline in AI Overviews (Ahrefs). Google says no special markup is needed. | Hygiene: keep for rich results and entity clarity |
| llms.txt | 28% of 137,210 domains publish one; 97% got zero requests. Coding agents and training crawlers were the main readers, retrieval bots almost absent (Ahrefs). Google Search ignores it. | Optional: worth it for developer docs, APIs and MCP products |
| Markdown via content negotiation | Cloudflare launched Markdown for Agents in February 2026, converting HTML to Markdown when a client sends an Accept: text/markdown header (Cloudflare). No citation data yet. | Optional: cheap, useful for agents, unproven for search |
One finding deserves emphasis because it is easy to miss. Many AI app builders generate client-side React single-page apps by default. Google renders JavaScript in a second pass; the crawlers behind ChatGPT, Claude and Perplexity do not. A GEO strategy published on a client-rendered site is invisible to the engines it targets until the pages are pre-rendered or server-rendered.
The pattern across this table is consistent: AI-specific files are low-cost bets on future agent behavior, while the proven levers are the oldest ones in technical SEO.
The agentic web: agents as a second audience
Agents do not read your page and cite it; they discover your capabilities, call them and act on the result. For a brand, that means a second front door next to the website, and in 2026 its standards are maturing at very different speeds.
| Layer | Standard | Status, October 2026 | What it does |
|---|---|---|---|
| Connect | Model Context Protocol (MCP) server | Production across Claude, ChatGPT, Cursor and others | Exposes your tools and data to any MCP host |
| Render | MCP Apps (@modelcontextprotocol/ext-apps) | Production: the first official MCP extension, live since 26 January 2026 in Claude and ChatGPT (MCP blog) | Tools return interactive UI that renders in a sandboxed iframe inside the chat |
| Discover | Agentic Resource Discovery (/.well-known/ard.json) | Draft: announced by Google in June 2026, spec v0.91 marked Proposal; Lighthouse now audits it (WebFX summary) | One manifest listing your MCP servers, APIs and agents for registries and agents |
| Act in the browser | WebMCP (document.modelContext) | Experimental: Chrome origin trial, versions 149 to 156 (Chromium intent) | A page registers tools that browser agents can call directly |
| Transact | Universal Commerce Protocol (UCP) and similar | Emerging; Google names UCP as what will let Search agents do more (Google) | Agent-completed purchases and bookings |
Commerce is where this moves fastest. Evertune found shopping experiences in 87% of ChatGPT answers to product questions by early 2026, up from 20% or less in October 2025. When the answer itself is the storefront, a product catalog an agent can query beats a product page it has to parse.
Tool descriptions are the new meta descriptions
In an MCP host, the model decides whether to call your tool from its name and description, just as a searcher once decided from a title and snippet. Write descriptions that state precisely when the tool helps and what it returns. Resist descriptions that order the model around, such as declaring a tool mandatory before any writing task. Hosts increasingly treat instruction-like text in tool metadata as a prompt-injection risk, and users who connect a server can read every word. The durable play is the same as on the web: be the most useful option, and say so accurately.
The agentic layer does not replace GEO; it compounds it. An agent that discovers your MCP server still needs a reason to trust your brand, and that reason is built by everything in the earlier sections.
Measuring AI visibility
Measure share of answers first and clicks second, because the click is shrinking. In Pew's panel of 68,879 Google searches, users clicked a traditional result in 8% of visits when an AI summary appeared versus 15% without one, and clicked a link inside the summary in just 1% (Pew Research Center). Google counters that clicks from pages with AI Overviews are higher quality, with visitors spending more time on site.
| Metric | What it tells you | How to measure |
|---|---|---|
| AI impressions on Google | How often your URLs appear in AI Overviews and AI Mode, by page, country and device | Search Console's Generative AI performance report, launched 3 June 2026 and rolling out in phases; impressions only, no clicks yet (Google) |
| Citation share per engine | Your share of cited sources across a fixed prompt panel, per engine | Third-party trackers sampling ChatGPT, Perplexity, Gemini, Copilot and AI Mode weekly |
| Brand mention rate and framing | Whether answers name you, recommend you, and describe you accurately | Same prompt panel, scored for mention, position and accuracy |
| AI referral sessions and conversions | The traffic and revenue that still arrives | Analytics segmented by AI referrers such as chatgpt.com, perplexity.ai, gemini.google.com and claude.ai |
| Retrieval bot activity | Whether search bots fetch your pages at all, and which ones | Server logs for OAI-SearchBot, ChatGPT-User, Claude-SearchBot and PerplexityBot |
| Agent actions | Whether agents use your capabilities | MCP tool calls, connector installs and completed agent tasks |
Build the prompt panel from real buyer language, not keywords: questions from sales calls, support tickets and community threads, plus the sub-questions a fan-out would generate. Run each prompt several times, because answers vary between runs and users; since May 2026, a searcher's own Preferred Sources selections can also shape which sources AI Mode shows.
Treat vendor dashboards as estimates. Google warns that no third-party tool has access to its internal ranking or AI systems (Google). Trends within one tool over time are more reliable than absolute numbers compared across tools.
Myths and weak evidence: an honest grade
Of eleven popular GEO claims, five are contradicted by 2026 evidence, and the two best supported are the least technical: cite your sources, and publish what only you know.
| Claim | Evidence | Why | What to do instead |
|---|---|---|---|
| Adding schema markup gets pages cited by AI | Contradicted | Matched study of 1,885 pages: no meaningful lift; live AI fetchers ignore JSON-LD | Keep schema for rich results and entities, not citations |
| llms.txt boosts AI visibility | Contradicted | 97% of files never requested; Google Search ignores the file | Publish it for developer docs and agent tooling only |
| A first-place Google ranking is enough | Contradicted | 56.8% of first-ranked pages were not cited by ChatGPT | Win sub-questions and off-site mentions as well |
| Blocking Google-Extended keeps you out of AI Overviews | Contradicted | Googlebot rules control AI features in Search; Google-Extended covers other systems | Use nosnippet, max-snippet or noindex if you need control |
| A page for every fan-out query wins | Contradicted | Google names this scaled content abuse when done to manipulate AI answers | Cover sub-questions inside strong pages and clusters |
| Content must be chunked for AI | Weak | Google says chunking is not required; answer-first placement still correlates with citation | Answer first under question headings, written for people |
| Proprietary E-E-A-T or citation-likelihood scores predict citations | Weak | No public validation; no tool sees inside the engines | Validate any score against observed citations before trusting it |
| Fresh content gets cited more | Moderate | 17M citations: AI-cited URLs 25.7% fresher, except in AI Overviews | Update with real changes and show the date; do not fake freshness |
| Brand mentions across the web drive AI visibility | Moderate | Strongest correlation (0.664) in a 75,000-brand study, but correlational | Earn authentic coverage, reviews and community presence |
| Statistics, quotations and cited sources lift visibility | Strong | Controlled experiment: up to 40% visibility lift | Attribute every number and quote to a primary source |
| First-hand, non-commodity expert content wins | Strong | Google's central 2026 recommendation; capsule plus owned insight is the top cited pattern | Capture real expert experience before drafting |
The grades are deliberately conservative. Strong means a controlled experiment or explicit platform guidance; moderate means large correlation studies that agree; weak means little or no data; contradicted means the best available test points the other way.
The operating playbook, and what it means for Geowrite
The research supports Geowrite's core thesis: AI engines cite brands that pair real expertise with specific, structured, verifiable content, and agents increasingly reach those brands through MCP. It also says where the master prompt should change, mostly where it leans on hidden markup, unvalidated scores or volume.
Experience capture is the step the master prompt lacks today, and it is the one that makes every downstream step credible; each measurement cycle restarts at the question map.
| Geowrite component | Research verdict | Change to the master prompt |
|---|---|---|
| Lovable front end (client-side React) | Critical gap: ChatGPT, Claude and Perplexity crawlers do not run JavaScript | Pre-render or server-render the landing page, ICP guides and every /content/:id page; test with a non-JS fetch |
| E-E-A-T Ambassador Vault | Strongly supported, but only with real expert input | Add an experience-capture step (interview prompts, voice notes, case data) and an expert sign-off status that gates publishing |
| Canvas 2.0 volume (1 to 50 articles) | Risky: mass pages for fan-out queries can count as scaled content abuse | Gate each path on a distinct owned insight, de-duplicate sub-questions, and route drafts through a review queue |
| E-E-A-T, Information Gain and AI Citation Likelihood scores | Weak: no public validation | Make each score a transparent checklist, then calibrate it against measured citations from the prompt panel |
| Dual-layer generator | Supported for visible structure; contradicted for hidden JSON-LD as a citation lever | Keep answer capsules and visible fact tables; treat Article, Person and FAQ markup as hygiene that mirrors the page |
| Landing page comparison matrix | Partly contradicted | Replace JSON-LD FAQ schemas as a GEO pillar with answer capsules, owned data and named experts |
| MCP tools and MCP Apps | Supported: production standard in Claude and ChatGPT | Keep the three mini-apps; rewrite tool descriptions to state when they help rather than declaring them mandatory |
| Measurement | Missing from the master prompt | Add a prompt-panel citation tracker per engine and import Search Console's Generative AI report |
| Brand DNA and persona extraction | Supported | Also extract proof assets (data, case studies, reviews, third-party mentions) and turn each persona into a question set |
| llms.txt, Markdown endpoints, ard.json, WebMCP | Optional forward bets; drafts or experiments | Ship them for Geowrite's own developer and agent surface, and label them as agent readiness rather than ranking factors |
The positioning that follows from the evidence is stronger than the one in the brief. Geowrite should not promise that markup or files make brands citable. It can promise something rarer and defensible: it turns a company's real expertise into the specific, attributed, well-structured content that every engine already rewards, at a pace a content team can actually review.
Sources
Platform documentation
- Google: Optimizing your website for generative AI features, updated July 2026
- Google: AI features and your website
- Google: Guidance on using generative AI content, updated October 2026
- Google: Generative AI performance reports in Search Console, June 2026
- Cloudflare: Markdown for Agents, February 2026
Research and data studies
- Aggarwal et al.: GEO, Generative Engine Optimization, KDD 2024
- Kevin Indig, ChatGPT citation concentration, via Search Engine Land, 2026
- AirOps: retrieval versus citation in ChatGPT, 2026
- Search Engine Land: content traits LLMs quote most
- Ahrefs: AI Overview brand visibility factors, 75,000 brands
- Ahrefs: AI assistants prefer fresher content, 17M citations
- Ahrefs: 1,885 pages adding schema, May 2026
- Ahrefs: 97% of llms.txt files never read, June 2026
- searchVIU: what AI systems see of schema markup
- Vercel and MERJ: the rise of the AI crawler
- Evertune: AI search statistics, updated August 2026
- Pew Research Center: clicks when AI summaries appear, July 2025
Standards
- MCP Apps: first official MCP extension, January 2026
- Chromium: Intent to Experiment, WebMCP, May 2026
- WebFX: Agentic Resource Discovery explained
Most figures come from vendor studies with their own methods and incentives; treat them as directional and re-test on your own prompt panel.
