Key Takeaways
- Three levels: Test accessibility (does the bot read the page?), test extractability (does the model isolate the correct answer?), and test citation (is the page included in actual search results?).
- Free: curl with an AI bot user-agent, server logs, Bing Webmaster Tools, assistant interfaces, and referrals in GA4.
- Paid: Monitoring platforms that query multiple search engines using a set of search terms, ranging from 30 to 500 € per month depending on volume.
- The classic mistake: Testing the quote before verifying access and extractability, and concluding that “GEO doesn’t work.”
Publish content optimized for the GEO Not testing it is like publishing a page without ever checking Search Console. Useful tests fall into three categories, and each category has its own tools—many of which are free. This article reviews them, including what they measure, how much they cost, and the order in which to use them. Monitoring over time, once the pages have been published, is the subject of a separate article: track your visibility on ChatGPT and Perplexity.
Level 1: Does the robot read the page?
Before conducting any editorial test, you must ensure that AI bots can access the content and view it in its entirety. Here are four checks, all of which are free:
| Test | Tool | What We Check |
|---|---|---|
| Authorization | The site's robots.txt file | GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot, Google-Extended, and Bingbot are not prohibited |
| Actual Access | A cURL request or a header testing service using the GPTBot user-agent | The server, CDN, or firewall returns a 200 status code and the complete HTML, not a verification page or a 403 error |
| Content visible without JavaScript | Browser with JavaScript disabled, or the “source code” view | The text, tables, and FAQ are included in the original HTML |
| Actual crossings | Server logs or CDN dashboard | Frequency of visits by each AI bot over the past 30 days |
A page that fails any of these tests will never be cited, regardless of its quality. CDN-based blocking has become common since several web hosts and content delivery networks block AI bots by default; this point is discussed in How Does AI Index Content?. Supplement this with Bing Webmaster Tools: ChatGPT relies on Bing’s index, and any page not included in that index is invisible to it.
Level 2: Does the model extract the correct answer?
This test is performed before publication, on a draft, and takes five minutes per page. It involves feeding the text into a model and verifying that it extracts the desired information.
- Extraction test: Paste the text into ChatGPT, Claude, or Gemini with the prompt: “A user asks [target query]. Respond only based on this text, citing the passage used.” ” If the model quotes the expected passage, the page is extractable; if it gives an off-topic response or quotes the introduction, the answer is misplaced.
- Battery life test: Copy a single paragraph and ask, “What is this passage about, and what claim does it make?” A vague answer indicates that the paragraph depends on the context.
- Missing Value Test: "What information is missing from this text for you to consider it a source on [query]?" The answers almost always point to the absence of numbers, dates, or an author.
- Competition Test: Submit your text and the text from the page with the most citations for that search query, and ask which one the model would cite and why.
Free interfaces are sufficient for these tests; the publishers’ developer consoles (OpenAI, Anthropic, Google AI Studio, Mistral) allow you to automate them across dozens of pages for just a few cents per query. The writing criteria that these tests evaluate are described in Structuring content so it can be understood and cited.
Level 3: Is the page mentioned in actual search queries?
After publication and indexing (which can take anywhere from a few days to a few weeks), the test involves entering the target queries into search engines—with web search enabled and in incognito mode—several times, and noting whether the page appears in the search results. Manually, this is feasible for 10 to 20 queries; beyond that, monitoring platforms take over.
| Category | Examples 2026 | What She Does | Price Range |
|---|---|---|---|
| Multi-engine tracking integrated into a marketing tool | GreenRed (GEO Tracking) | Ask ChatGPT, Claude, Gemini, Perplexity, and Mistral 10 questions, count the citations, identify competitors and third-party sources, and track the history | Included in the subscription |
| Specialized GEO platforms | Peec AI, Otterly.AI, Profound, Scrunch | Extended query sets, voice share, source analysis, alerts | €30 to €500 per month, plus in-house positions |
| AI Modules in SEO Suites | Semrush AI Toolkit, Ahrefs Brand Radar, SimilarWeb | Brand Mentions in AI Responses, AI Traffic Estimates, Competitor Comparisons | 100 to 500 € per month in addition to the suite rates |
| Measuring Actual Traffic | GA4, Search Console | Referrals from chatgpt.com, perplexity.ai, copilot.microsoft.com, gemini.google.com; clicks from AI previews (not isolated) | Free |
The choice depends on the number of queries to track and the number of search engines; for an SME, 10 to 30 queries across 5 search engines, with weekly or monthly reports, are sufficient. The configuration for tracking traffic is described in GA4 and Search Console for Measuring GEO.
Interpreting a Negative Test Result
A page that has not been cited four weeks after publication has one of the following causes, which should be examined in this order: blocked robot or page not indexed in Bing (Level 1); misplaced answer or non-standalone paragraph (Level 2); the page ranks beyond the first page of Google and Bing for the query, so it’s never crawled by the search engine; the query is dominated by third-party sources (comparison sites, forums, media outlets) where you also need to have a presence. Only the third case falls under traditional SEO; the other three can be corrected within a few hours.
A realistic test routine
- For each publication: Level 1 (once per site, then whenever the CDN or firewall changes) and Level 2 (every five minutes per page).
- At four weeks: Level 3 manual review of the page's three target queries.
- Every month: automated analysis of the entire set of search queries, comparison with the previous month, and review of competitors' sources.
- Every quarter: review of pages with unstable citations, using the following indicators: GEO KPIs.
How GreenRed Can Help You
Instead of juggling multiple tools, GreenRed's GEO Tracking module brings these metrics together in a single dashboard, compares them over time, and shows you which actions to prioritize. You can try it for free—no credit card required—on the page Rates.
Frequently Asked Questions
Is it possible to test your content without paying for a tool?
Yes, for all three levels, on a small scale: robots.txt, curl, and server logs for access; free API interfaces for extractability; and manual data collection in incognito mode for citation, based on 10 to 20 requests. Paid services become useful when the number of requests and search engines makes manual collection too time-consuming or too sporadic.
Does the test in the interface yield the same result as it does for a user?
Provided that you disable caching and custom instructions, enable web search, and repeat the query three times, yes. The API results sometimes differ from the interface (search tools, default template); to reflect the actual user experience, the consumer interface remains the standard.
How long after publication should the citation be checked?
Wait for your site to be indexed by Google and Bing—this usually takes a few days to two weeks for an active site. A citation test conducted after four weeks is representative; one conducted after one week is not. Google’s AI previews track organic rankings with an additional lag of a few weeks.
Do tracking tools cover Google's AI insights?
Some include them by noting, for each query, whether a snippet appears and which pages it references; others limit themselves to chatbots. Be sure to check this before signing up, because in France, AI snippets account for a growing share of informational searches and are not tracked separately in Search Console.