How to Run a GEO Audit: The Complete 2026 Checklist

A GEO audit answers one question in five parts: can an answer engine reach your pages, read them, understand them, find something quotable on them, and act on them? Each part is a category, and they are not equally weighted — the evidence says some matter far more than others. The five categories, in the order the rest of this guide walks through them: crawler access (are the bots that feed live AI answers allowed in), quotable content (is there anything on the page worth lifting into an answer), structured data (can the page be parsed without guessing), technical readability (does the page render without JavaScript, with a valid llms.txt and clean sitemap), and agent readiness (can an AI agent, not just a crawler, discover and act on the site). The trap this guide exists to prevent: treating the audit as a pass/fail technical checklist. Crawler access makes you eligible to be cited. Quotable material makes you cited. A site can clear every technical box and still give an engine nothing worth quoting — we know, because ours did.

Start with robots.txt, because nothing downstream matters if the crawlers can't read the page. But "allow crawlers" is too vague to audit. There are two distinct populations, and they need different decisions. Training crawlers collect data to train future models — GPTBot (OpenAI), ClaudeBot (Anthropic), Google-Extended, CCBot (Common Crawl), Applebot-Extended. Blocking these keeps your content out of the next model's weights; it does not remove you from today's answers, and it is a legitimate rights decision. Live-answer crawlers fetch your page in real time to build an answer a user is asking right now — OAI-SearchBot (ChatGPT search), PerplexityBot and Perplexity-User, ClaudeBot when browsing, Google's crawlers behind AI Overviews and AI Mode, and Bingbot behind Copilot. Block one of these and you disappear from that engine's live citations entirely. The single most common — and most expensive — mistake is confusing the two: blocking GPTBot to "keep AI out," and unknowingly also blocking OAI-SearchBot, so you vanish from ChatGPT's cited answers while your competitors stay in. Audit each user-agent against your robots.txt individually. Then do the thing most audits skip: check your competitors' robots.txt too. "You block PerplexityBot while three of five competitors allow it" is the finding that actually changes a decision, and it's why CrunchJunkie crawls the competitor set alongside your own.

This is the step almost every GEO checklist underweights, and the research says it's the one that matters most. Aggarwal et al. tested roughly 10,000 queries and found that adding statistics, direct quotations and authoritative source citations lifted a page's visibility in generative answers by up to 40% — while classical keyword density barely moved it at all ("GEO: Generative Engine Optimization", KDD 2024). So the audit here isn't about word count or keywords. It asks: does the page contain cited statistics (a number with a source), direct quotations an engine can lift verbatim, and authoritative outbound links that signal the page sits in a real evidence chain? Those three are the strongest on-page levers we have evidence for. Here is the worked example, shown below. On 8 August we ran this exact audit on our own Spanish agency, pmax.online. Technically it was strong: composite 93/100, band AI-ready, AI-crawler access 100/100, agent readiness 100/100, valid llms.txt, content server-rendered (1,398 words visible without JavaScript). On every checklist circulating in this industry, that's a finished job. The audit disagreed and flagged three high-priority content gaps on the same pages: no authoritative outbound citations (two external links, none to recognised authoritative domains), direct quotations on only one of six pages, and statistics present but left unattributed to a source. Its own summary was blunt — citing credible sources is the single strongest signal for AI citation, and we were missing it entirely. A technically strong site was giving answer engines nothing worth quoting. That is the gap this step catches and the others can't.

Structured data (Schema.org markup) tells an engine what a page is instead of making it infer. The audit checks for the types that describe your actual content — Organization and its sameAs profiles, Article or BlogPosting with a named author, Product with price and availability, FAQPage for question-and-answer blocks, BreadcrumbList for site structure. Two things separate a real audit from a plugin that just says "add schema." First, the markup has to be accurate — schema that disagrees with the visible page is worse than none, and it's the kind of error that ships silently. Second, an author identity in Article schema has to resolve to a real, verifiable person (a Person with a role and a profile URL that actually loads), because a named authorship claim pointing at nobody is an E-E-A-T liability, not a signal. What to look for in the audit output: valid JSON-LD (not deprecated microdata), no validation errors, and — the part checklists miss — schema that matches what the page actually says.

A crawler that is allowed in still has to be able to read the page. This category is the classic technical layer, reframed for retrieval. Server-side rendering is the big one. Most live-answer crawlers execute little or no JavaScript, so content that only appears after a client-side render is, to them, a blank page. The audit checks whether your primary content is present in the raw HTML — view-source, not the rendered DOM. A React or Vue site that ships an empty shell and hydrates on the client can be technically "online" and functionally invisible. llms.txt is the emerging convention (proposed at llmstxt.org) — a Markdown file at your root that gives models a curated map of your most important pages, the way robots.txt gives crawlers rules and sitemap.xml gives them a list. Adoption is early and no major engine treats it as authoritative yet, so the audit scores it as a low-weight positive signal, not a requirement — present and valid is better than absent, but it will not save a site that fails Steps 1–3. The rest of this category is SEO hygiene that still applies: a working XML sitemap, correct canonical tags, no accidental noindex, and reasonable response times.

The newest category, and the one that will matter more every quarter. Retrieval crawlers read your pages; AI agents try to do things on behalf of a user — compare products, check availability, start a task. Agent readiness asks whether your site exposes anything an agent can act on. The audit scores this separately (a standalone 0–100) because it's forward-looking: presence of .well-known interfaces, machine-readable structure an agent can navigate, and — for the platforms that have gone furthest — an MCP (Model Context Protocol) surface that lets an agent query your product directly. Most sites score low here today, and that's fine; it's a leading indicator of where you'll need to invest, not a pass/fail. We treat our own MCP connector as the reference implementation of what a high agent-readiness score looks like in practice.

The single biggest error in GEO scoring is averaging the five categories equally. That produces a comforting composite that hides the failure that's actually costing you citations — exactly the mistake we make fun of in blended "AI visibility scores." Weight by evidence. In our own audit the weights are visible on the score card above: AI-crawler access and content each carry 30 points, structured data 20, technical hygiene 15, and llms.txt just 5 — and agent readiness sits outside the composite entirely. Crawler access is also a gate: fail it for an engine and your score there is effectively zero, however good the content. Quotable content carries the heaviest weight of the passable categories, because it has the strongest causal evidence behind it (the up-to-40% finding above). That weighting is why our own site scored 93/100 and still surfaced seven ranked fixes worth up to +11 points: AI-crawler access and technical hygiene were perfect (100), but content accessibility (92) and structured data (76) pulled the weighted composite down. The composite was high; the audit still refused to let a strong technical score paper over the gaps. A score you can't act on is decoration — a good GEO audit hands you a ranked list of fixes, each tagged with the points it would earn back, not a single number.

An audit is a snapshot; AI visibility is a moving target. Three things turn a one-off score into an actual programme. Fix in priority order. Take the ranked list and start at the top — usually a crawler-access gap (highest leverage, often a one-line robots.txt change) or the content gaps from Step 2 (highest evidence). Don't scatter effort across low-weight items to lift a vanity number. Benchmark against competitors. Your robots.txt decisions and content depth only mean something relative to the other pages an engine reads when answering your prompts. Crawl the competitor set: where do they allow a crawler you block, and where is their content more quotable than yours? That gap is your roadmap. Re-run on a schedule. Provider bot names change, you ship new pages, competitors move. We re-audit our own site weekly. Run the free GEO audit on any domain to get the score and the ranked fix list in seconds, then bring it inside CrunchJunkie for the weekly re-audit, the competitor crawler benchmark and the content briefs that close the gaps this checklist finds.