Table of Contents
Why We Instrumented Our Own Site First
Generative engine optimization is easy to sell and hard to prove, because most GEO advice is published without measurement behind it. We decided our own site would be the test bench: every technique we recommend runs in production at cloudnsite.com, and we watch the results in Search Console and analytics like any other channel.
This post documents the architecture as it actually runs, including the parts that produce measurable results and the parts that do not. Positions and query figures below come from our own Search Console and analytics data, pulled the first week of August 2026.
The Stack, Layer by Layer
Layer 1: crawler access. robots.txt explicitly allows the major AI crawlers (GPTBot, ClaudeBot, PerplexityBot, Google-Extended, and peers). This is the zero-cost prerequisite: if the crawlers cannot read you, nothing downstream matters. Anthropic documents that its bots honor robots.txt directives, and Google publishes Google-Extended as the token controlling training use, so the access decision is genuinely yours to make.
Layer 2: llms.txt. The proposed standard is a markdown file at the site root: an H1, a summary blockquote, and H2-delimited sections of curated links with one-line descriptions. Ours indexes the core pages, the pricing surfaces, and the briefs described below, and is regenerated on every build so it can never go stale. We wrote a full implementation walkthrough in our llms.txt guide.
Layer 3: machine-readable briefs. A directory of self-contained markdown files, one per offering, each carrying positioning, deliverables, pricing posture, and a canonical URL. These exist because assistants synthesizing an answer do better with a 600-word structured document than with a marketing page's DOM. The briefs are hand-written, updated when the offer changes, and listed in llms.txt so crawlers find them in one hop.
Layer 4: structured data and extractable prose. Every page ships schema (Organization, Service, FAQPage, BlogPosting, BreadcrumbList as appropriate), and long-form content is written for extraction: question-form headings with a direct answer in the first sentence beneath them. When an assistant needs a quotable answer, the page hands it one.
Layer 5: a JSON index. ai-search.json aggregates the site's offerings and content into one machine-readable file for anything that prefers structured over prose.
The Build Pipeline That Keeps It Honest
The stack would rot in a month if it were maintained by hand. It is not. The site is statically prerendered, and the build pipeline regenerates the blog manifest, the sitemap, llms.txt, and ai-search.json on every deploy, sourcing pricing language for the site's pages and structured data from a single canonical data file. The hand-maintained surfaces, the briefs above all, are covered differently: a rule-based guard in the verification suite sweeps every surface, hand-written files included, for retired vocabulary, stale prices, and unconditioned claims before a deploy ships, because an AI assistant will happily quote your stale pricing forever if you leave it lying around in a forgotten file.
That single-source-of-truth discipline is, in our experience, the highest-value and least-discussed part of GEO. Assistants cache and cross-check; inconsistency reads as unreliability.
What the Data Shows
Three observations from our own measurement, stated plainly.
First, sentence-form queries now dominate our impression volume. In the most recent month, 1,860 of our roughly 4,000 Search Console queries were long, sentence-form questions, the shape AI Overviews and answer engines generate, carrying about 18,600 impressions. Our largest single query is a full sentence comparing the total cost of ownership of a private AI server against ChatGPT Enterprise, at roughly 2,900 impressions with our comparison article at an average position around 5. A cluster of ambient-clinical-documentation questions averages positions roughly 3 through 14 against our healthcare vendor guide.
Second, assistants send small but unusually qualified traffic. In a recent 30-day window we measured roughly 110 sessions referred directly from AI assistants: ChatGPT referrals landed most often on our case studies page, Claude referrals on agency and pricing content, and Perplexity referrals on the llms.txt guide and our GEO service page. The volume is modest. Where those sessions land is the interesting part: proof and pricing surfaces, not the homepage.
Third, citations do not equal clicks. Our highest-impression AI-shaped queries show strong positions and near-zero click-through, the pattern consistent with answer-engine surfaces: strong average positions on question-form queries with few of those impressions converting to visits. We treat that visibility as brand distribution and design the click-bearing capture separately, with dedicated pages targeting the classic query forms of the same intent.
The Extraction Pattern in Practice
The writing pattern that shows up in our best-positioned pages is mechanical enough to teach in one paragraph. Every substantive section gets a heading phrased the way a person would ask the question, and the first sentence under it answers the question completely, with the qualifications and context following rather than leading. "Is there such a thing as a HIPAA certified AI tool?" is answered "No." in the first word of its section on our healthcare content, and that page now holds top-five positions on a family of sentence-form queries asking variations of exactly that.
The reason this works is unglamorous: answer engines quote. A section that spends three sentences building context before committing to an answer gives the extractor nothing quotable; a direct first sentence gives it the whole citation. Writing this way also improves the page for human skimmers, which is why we stopped thinking of it as an AI trick and started treating it as house style.
The same logic drove a second pattern: when Search Console shows a sentence-form query where a page ranks between positions 4 and 15, we add a section to that page whose heading matches the question and whose first sentence answers it from facts already on the page. That is targeted work, a few sections at a time on pages that have already demonstrated relevance, not a mass rewrite.
What Did Not Work
Honesty section. Publishing the machine files produced no measurable step change on its own; the impression growth tracked content quality and structure, not file presence. Schema alone moved nothing without extractable prose to go with it. And no amount of GEO plumbing compensated for pages that lacked verifiable sources: the pattern in what gets cited strongly favors content that cites its own evidence, which is why every post on this site now carries a verified source section.
The moat is not any single file. It is the compounding system: clean crawler access, current machine-readable surfaces, extractable answers, and claims an assistant can check.
What We Would Tell a Client
Run the stack in this order: fix crawler access today, add llms.txt this week, restructure your highest-value pages for extraction this month, and put your offer facts in one canonical source before any of it. Measure sentence-form queries separately from classic ones, because they are a different channel with different economics. Expect visibility before clicks, and build the click capture deliberately.
Or have it built: this architecture is what our agent-ready websites service installs, and the generative engine optimization page covers the engagement shape. Either way, the free AI Readiness Self-Check takes minutes if you want to locate the bigger picture first.
FAQs
What is a GEO stack? The set of site infrastructure that makes content usable by AI assistants and answer engines: crawler access rules, llms.txt, machine-readable content files, structured data, and extraction-friendly prose, kept current by the build pipeline rather than by hand.
What does llms.txt do on our own site? On our own site, Perplexity referrals land on our llms.txt guide and AI crawlers fetch the file, but the file alone produced no measurable step change. It works as one layer of a system whose value shows up in sentence-form query impressions and assistant referrals.
How do you measure GEO results? Separate sentence-form queries from classic ones in Search Console, track positions on the question-shaped queries, and segment analytics referrals from assistant domains. Expect high impressions with low clicks on AI-shaped queries, and judge that visibility as distribution rather than failure.
How long did this take to build? The individual layers are small; the discipline is the work. Generation hooks in the build pipeline, a canonical pricing source, and verification guards were built across normal development cycles; the generated surfaces maintain themselves, and the hand-written ones are swept by the guards.
---
Sources
- llmstxt.org, "The /llms.txt file". The proposal this stack implements: an H1 name, optional summary blockquote, and H2-delimited sections of markdown links.
- Google, "Google's common crawlers". Documents Google-Extended as the robots.txt product token controlling use of crawled content for AI model training and grounding.
- Anthropic, "Does Anthropic crawl data from the web?". Documents ClaudeBot and related crawlers and states Anthropic's bots honor robots.txt directives.