AI SEO Testing · Evidence → Action
An evidence library across four buckets: official guidance, large-scale studies, controlled experiments, and technical labs. Each is mapped to a test plan your team can scope, defend, and measure.
Gen AI: Growth in Usage & Referrals (US Web, Jan 25 – Jan 26)
Bars: visits · Line: referrals · Source: Similarweb, March 2026 · Illustrative adaptation
AI Tools vs Search Engines across the Purchase Journey
Purple: AI tools · Orange: search engines · Source: Similarweb, March 2026 · Illustrative adaptation
The Problem
Teams are reading studies from platforms, tool vendors, publishers, and independent researchers, alongside official guidance from Google and OpenAI. The findings rarely arrive separated by evidence type, scored for confidence, or mapped to a test you can actually run.
"We separate official guidance from vendor studies, controlled experiments, risk analyses, and technical labs, then map each to a test you can execute."
The System
Every published finding is normalized into a Learning Card with source, evidence type, trust role, finding, business implication, test idea, confidence, primary metric, risks, and an activation path.
Click Customize this test on any card. A guided refinement flow uses the card's recommended template, your goal, timeframe, and resources to generate a structured, source-cited test plan.
Evidence Library · Launch Set
The goal is not to chase every AI SEO trend. It is to separate guidance, studies, experiments, risk analyses, and technical labs, then identify which findings are worth testing in your business.
How AI engines pick, cluster, and re-cite sources.
Citation mechanics & co-citation behavior
ChatGPT web citations concentrate early in conversations, use multiple sources per cited turn, and often cluster sources by topic or domain category.
Secondary metrics. Co-citation frequency · Citation position · Citation neighbor overlap · Brand mention sentiment
Best fit
Watchouts & limits
Why this confidence: Large dataset and clear citation mechanics, but platform behavior may shift as ChatGPT search changes.
SERP overlap & query fan-out evidence
AI Overview citations don't simply mirror the top organic results. Ahrefs' updated study found a lower top-10 overlap than its earlier work, suggesting more fan-out influence.
Secondary metrics. Organic rank overlap · Fan-out topic coverage · Supporting-page diversity · Citation source type
Best fit
Watchouts & limits
Why this confidence: Large-scale SERP and citation dataset, but Google AI Overview behavior changes quickly.
Paired experiments isolating one variable at a time.
Clean format test (anti-shortcut)
In Otterly's 14-day test, paired Markdown pages received no AI crawler visits or citations, while the HTML versions did receive AI crawler activity and citations.
Secondary metrics. Citation pickup · Log-file AI bot requests · Prompt-level answer inclusion
Best fit
Watchouts & limits
Why this confidence: Clean controlled setup with same content and discovery path, but limited to one brand/site context.
Structured-data quality test
A small controlled test found the well-implemented schema page was the only one to appear in an AI Overview.
Secondary metrics. Indexation · Organic rank movement · Rich-result eligibility · Crawl frequency
Best fit
Watchouts & limits
Why this confidence: Very clean design, but only three pages, promising rather than conclusive.
Page-level credibility treatment test
The GEO benchmark found that Cite Sources, Quotation Addition, and Statistics Addition improved generative engine visibility. AutoGEO later surfaced similar preference rules around source citation, factual accuracy, and specific evidence. The practical hypothesis: pages that support claims with named sources, concrete data, and attributed quotes may be easier to cite than pages that rely on generic assertions.
Secondary metrics. Citation position · Answer influence score · Citation quality · AI referral signals
Best fit
Watchouts & limits
Why this confidence: Consistent benchmark signal across GEO and AutoGEO, but not direct revenue proof. Visibility lift still needs to be tied to traffic, assisted conversions, or lead quality in production.
Extractability and opening-structure test
AutoGEO surfaced preference rules around conclusion-first, comprehensive, and in-depth content. The practical hypothesis: pages that answer the core question in the opening block may be easier for generative engines to extract, paraphrase, and cite than pages that begin with delayed narrative intros.
Secondary metrics. Citation inclusion · Citation share · Answer position · Scroll depth
Best fit
Watchouts & limits
Why this confidence: Strong directional signal from AutoGEO preference-rule extraction, but it still needs page-type validation. The effect may vary by content format, query intent, and engine.
Official guidance and eligibility validation. Necessary, not sufficient.
Platform guardrail & crawl-access baseline
Google says SEO fundamentals remain relevant for AI features and that AI Overviews / AI Mode may use query fan-out. OpenAI documents separate crawlers for search, user-triggered browsing, and model training.
Secondary metrics. Robots allow/deny status · Snippet eligibility · OAI-SearchBot requests · GPTBot requests
Best fit
Watchouts & limits
Why this confidence: Official platform documentation. Defines eligibility and access controls, not ranking causality.
Anti-hype technical validation
SE Ranking found low adoption of llms.txt and no observed correlation between the file and AI citation frequency.
Secondary metrics. AI crawler requests to /llms.txt · Bot access status · Indexed priority pages · Server log evidence
Best fit
Watchouts & limits
Why this confidence: Large domain-level dataset, but correlation studies cannot prove non-effect across every platform or future implementation.
Competitive risk analyses and edge / infrastructure experiments.
Comparison-page & reputation-risk evidence
Self-promotional listicles still appear in AI citations, especially in software/review contexts, but Peec does not recommend the tactic because of reputational and algorithmic risk.
Secondary metrics. Self-promotional source rate · Competitor mentions · Sentiment · Third-party source mix
Best fit
Watchouts & limits
Why this confidence: Large citation dataset with fixed prompts over time, but vertical-specific to software and review-style queries.
Crawler discovery & structural metadata experiment
A site owner can expose internal links and headings through HTTP response headers, enabling fast crawl/audit workflows without parsing full HTML.
Secondary metrics. Header size · 5xx rate · AI crawler requests · Internal-link graph coverage
Best fit
Watchouts & limits
Why this confidence: Medium for crawl/audit utility, low for AI visibility impact, not a proven AI citation lever.
All cards link to source material. Some sources are studies, some are official guidance, some are experiments, and some are technical labs. Confidence ratings reflect study design, dataset size, and replication risk.
Interactive · Test Customization
Pick any Learning Card, confirm your context, and generate a plan tied to that study's recommended template: source, hypothesis, setup, measurement, success, and watchouts.
Who it's for
Get Started
Turn the evidence library into a prioritized testing roadmap your team can scope, defend, and measure.
Start with one evidence-backed test plan. Use it to decide what to test, how to measure it, and what not to overclaim.