GEO checklist: how to audit your website and your brand for AI engines

Go through it in order. The crawling items come first because if an engine can’t read your site, nothing else matters. Every item says why it is on the list: if the reason doesn’t apply to you, skip it. For the general concept of GEO (generative engine optimization), start with the GEO guide.

• robots.txt lets AI search crawlers in (OpenAI’s OAI-SearchBot, Anthropic’s Claude-SearchBot, and Googlebot): if you block them, those engines can’t read your pages to answer.
• You have decided separately what to do about training crawlers (GPTBot, ClaudeBot and the Google-Extended token): blocking them is a decision about your content being used to train models, which is separate from appearing in answers that use search. Google-Extended doesn’t affect your inclusion in Google Search.
• Your CDN or firewall doesn’t block them: services such as Cloudflare or Imperva can stop AI bots by user agent or IP reputation even when robots.txt allows them, and no robots.txt check will catch it. Review your CDN’s bot settings and its logs.
• Your main content is in the HTML the server returns: many AI crawlers don’t run JavaScript, so a site that assembles its text in the browser reaches them empty. Test it with “view source,” not with the element inspector.
• Your structured data (JSON-LD) is in that HTML too: if a tag manager injects it after the page loads, a crawler that doesn’t run JavaScript never sees it.
• The lang attribute on the html tag declares the right language (“en,” or a regional variant such as “en-US” or “es-MX”): a Spanish page marked as English competes poorly for questions asked in Spanish.
• Every page has a canonical that points to itself: a canonical pointing to another page, for example a Spanish version pointing to the English one, asks Google to index the other page instead.
• hreflang connects the language and country versions of bilingual or multi-country sites, in both directions: without it, the search engine may show the wrong version in each market.
• Your important pages are indexed: a page reported as “Crawled - currently not indexed” in Search Console doesn’t take part in Google’s AI Overviews.
• There is no noindex or nosnippet set by mistake: noindex removes the page from search, and nosnippet limits what Google can show from it, including in its AI features.
• Your XML sitemap is current and you use IndexNow: the sitemap lists what you want crawled; IndexNow notifies Bing, and the other search engines that adopt the protocol, every time you publish or change a URL.

• Every page answers first: the first paragraph answers the page’s question in two or three sentences, which is what an engine can lift and cite.
• There is one page for each important sub-question in your category (price, how to choose, coverage, how it works): AI search breaks each question into parts and looks for the best answer to each one.
• There isn’t one page for every variation in wording: Google warns that this can violate its policy against scaled content abuse.
• You have honest comparison pages (against competitors, alternatives, “X or Y”): comparison questions get answered with pages that compare, and if you don’t write yours, someone else will.
• You name your country, your cities and your currency: answers change from one market to another, and the engine needs to know where you operate.
• You give concrete, verifiable facts: prices or price ranges, lead times, specifications, dates. A vague claim can’t be cited.
• The structure is clear: headings that describe the content, lists and tables, so the engine finds the part it needs.
• Pages with data that changes show their update date: it tells the reader and the engine how current the fact is.
• You publish in each market’s language and, if you sell abroad, in English as well.
• You cover the topics people ask about: compare the questions in your category with what your site answers. The gaps are your content plan.

• Your brand has the same name, category and description everywhere (your site, social profiles, directories, marketplaces, press): contradictions make the engine hesitate, and an engine that hesitates names someone else.
• You have organization structured data with name, URL, logo and sameAs links to your official profiles: it helps systems connect your site with your brand, and it has to match what the page shows.
• Your brand has a Wikidata item, if it meets the criteria: it turns your brand into an identifiable entity rather than just a word, which reduces confusion with similarly named companies.
• Your knowledge panel and Google Business Profile are correct: they are the record Google associates with your name.
• Product and price data match across the page, the structured data and, if you run a store, your Merchant Center feed: a different price in each place is a fact the engine can’t rely on.
• You have a page with the company’s facts (founding date, headquarters, countries, leadership team): it is your source of truth when AI repeats an outdated fact.

In Refractia’s data, a brand’s own site accounts for between 0.5% and 2.8% of what AI cites when it talks about that brand. Most of what engines read about you is on third-party sites, so this section carries as much weight as the ones before it.

• You know which domains the engines cite for your category: it is the list of places worth being, and it doesn’t always match the one you imagine.
• You appear in the comparisons and lists for your category in your country: they are a frequent source when someone asks “which one should I choose?”
• You have earned coverage in your market’s media: articles, interviews and reports with data an engine can cite.
• Your reviews are on the platforms your category uses and you respond to criticism: third-party opinions feed into how AI describes you.
• Your third-party profiles are up to date: directories, marketplaces, professional networks, app stores.
• You don’t buy mentions or links: it puts your rankings and your reputation at risk, and it doesn’t create the real association engines look for.

• Search Console is set up and you check it: clicks and impressions from Google’s AI features are included in the Performance report, and the generative AI performance report shows their impressions.
• Bing Webmaster Tools is set up and connected to IndexNow: it is where you see how Bing crawls and indexes your site. It matters beyond Bing, too: ChatGPT search shares queries with third-party search providers, and OpenAI names Bing among them.
• You identify traffic from assistants in your analytics, for example from chatgpt.com or gemini.google.com: it is the measurable part of the visits AI sends you.
• You have a fixed set of questions covering category, comparison and purchase intent, written in the market’s language and naming the country: without fixed questions you can’t compare one measurement with another.
• You measure several engines and separate chat from AI search: chat engines answer from the model’s memory, and AI search from what the web ranks today; each improves with different actions.
• Every rate carries its sample, and non-answers don’t count as zeros: “14 of 20” can be discussed; a “35%” calculated over queries the engine never answered can’t.
• You measure your competitors with the same questions: otherwise you are comparing questions, not brands.
• You measure again after every important change, with the same questions, on the same engines and in the same country: it is the only way to know whether something worked.

An audit ends in a long list. To turn it into a content strategy, work in this order:

1. 01

   Fix what blocks

   Crawling, CDN, server-rendered HTML, canonicals and indexing. These are contained technical changes, and everything else depends on them.

2. 02

   Write the missing pages for the questions that matter most

   Start with the purchase and comparison questions in your category where you don’t show up today, with the answer first and your country’s data.

3. 03

   Get your entity in order

   Name, company facts, profiles and structured data that agree with each other. It is one-time work that you then maintain.

4. 04

   Work on third-party sources

   With the list of domains the engines cite, prioritize where to earn a real presence.

5. 05

   Measure and repeat

   The same set of questions, the same engines, the same country. Whatever didn’t move goes back on the list.

In Refractia, the Technical Workshop runs 56 crawlability and extraction checks for AI, each with an evidence grade: A (documented by Google or mechanically verifiable), B (independent study) and C (convention, shown as an observation with zero weight on the score). It detects when a CDN or firewall blocks AI crawlers even though robots.txt allows them, organizes issues as a backlog (quick wins, defects and observations), and exports to PDF and Excel. GEO Optimization analyzes your pages with a six-pillar GEO Score, page-by-page scores and a content plan, up to 500 pages in full-site mode; the Content Workshop turns that plan into tasks, and Demand compares the topics people ask about with what your site covers.