How to Choose a GEO Audit Tool: 8 Checks Before You Trust the Score

Choose a GEO audit tool by checking how it reaches its verdicts, not how many checks it lists. Run one page you know well through it and look for eight things: correct crawler facts, a live test of crawler access, a view of the HTML without JavaScript, no points for llms.txt or schema, a score you can trace to evidence, honest handling of AI answer variance, use of the free first-party data, and fixes you can act on and verify. Tools that fail several of these will send you chasing changes that don't matter.

First, know which kind of tool you're looking at

“GEO tool” now covers four different products. They answer different questions, and comparing a prompt tracker's number with a page audit's score is comparing a thermometer with a stethoscope.

The rest of this guide is about the second and third kinds, the ones that claim to tell you what to fix. For tracking, see how to measure AI visibility without fooling yourself.

1. Does it get the crawler facts right?

Every AI vendor now runs separate crawlers for separate jobs, and the most common error in GEO tools is treating them as one. Check any tool's crawler advice against these documented facts:

• OpenAI: GPTBot collects training data; OAI-SearchBot decides whether a site appears in ChatGPT search; ChatGPT-User fetches pages when a user asks. Blocking GPTBot does not remove you from ChatGPT search (OpenAI).
• Anthropic: ClaudeBot is for training, Claude-SearchBot indexes for Claude's search, and Claude-User fetches on a user's request (Anthropic).
• Perplexity: PerplexityBot indexes for Perplexity's answers, and Perplexity-User fetches on request (Perplexity).
• Google: AI Overviews and AI Mode use the normal Googlebot index. Google-Extended controls Gemini training and grounding outside Search, and is not a Search ranking or inclusion signal (Google).

A tool that tells you to “allow Google-Extended to appear in AI Overviews”, or warns that blocking GPTBot removes you from ChatGPT search, has the basics wrong. The full breakdown is in GPTBot vs OAI-SearchBot.

2. Does it test access live, or only read robots.txt?

robots.txt states a policy. A CDN, firewall or bot-management product enforces one, before robots.txt is consulted, and when the two disagree the firewall wins. In our study of the top 5,000 sites, 2,672 sites allowed OAI-SearchBot, PerplexityBot and Claude-SearchBot in robots.txt and answered all three requests. Of those, 263 (9.8%) still turned at least one away, and 77 (2.9%) did so in a way aimed at AI crawlers by name.

A tool that only parses robots.txt can't see any of that. Look for one that requests the page with each AI crawler's user agent and compares the answer with a normal browser request. Look also for honesty about the limit: real crawlers are verified by IP address or signed requests, so a request that only borrows a crawler's name can be treated differently. A good tool labels this result as a simulation, not a verdict.

3. Does it look at the page without JavaScript?

The crawlers behind ChatGPT, Claude and Perplexity read the HTML your server returns; Vercel and MERJ's analysis of crawler traffic found no evidence that OpenAI's or Anthropic's crawlers execute JavaScript (Vercel). A page that builds its text in the browser can look complete to you and nearly empty to them. If a tool audits pages in a full browser and never compares that with the raw HTML, it's reporting what you see, not what the crawler gets. See can ChatGPT, Claude and Perplexity read your website?

4. Does it award points for llms.txt or schema?

Google's own guide to its AI features says you don't need an llms.txt file, special schema.org markup or AI-specific files to appear in them, and that optimizing for generative AI search is still SEO (Google). No AI search engine has said it uses llms.txt to choose which pages to cite. Structured data still matters for rich results and for describing entities clearly, but no controlled test has shown that adding it changes which pages AI answers cite.

So a tool whose top recommendations are “create llms.txt” and “add FAQPage schema” is spending your attention on the cheapest things to measure, not the things most likely to matter. It's fine for a tool to report them; it's a red flag when they move the score. The evidence for each is graded in GEO myths, graded.

5. Can you trace the score back to evidence?

A composite score is only as good as its weights. Ask three questions of any “GEO score” or “AI readiness score”:

1. What is each point based on? Look for a public methodology that lists every check, its weight and its source.
2. Does a page-killing problem cap the score? A page set to noindex, or one AI crawlers can't fetch, shouldn't score 80 because its headings are tidy.
3. What happens when something can't be measured? A timed-out request should be reported as unverified, not as a failure that drags the score down.

If the score is written by a language model reading your page, run the same page twice. If the number moves without the page changing, treat it as an opinion, not a measurement. Our own weights and evidence grades are on the methodology page.

6. Does it treat AI answers as samples?

AI answers change from run to run. A vendor study by Otterly.AI that collected 252,407 answers found brand mentions noticeably more stable than the specific pages cited (Otterly.AI, vendor study). That means any single “are you cited?” check is one draw from a distribution. Good tools report how many runs, prompts and engines sit behind a number, show the actual answers and cited URLs, and separate what was observed from what was inferred. Be wary of a tool that turns one run into a rank.

7. Does it use the data Google and Bing now give you free?

Since 31 August 2026, Search Console's generative AI performance report has shown every verified site its impressions in Google's AI features, by page, country and device, though without queries or clicks (Search Console Help). Bing Webmaster Tools' AI Performance report shows citations, cited pages and sampled grounding queries for Copilot and Bing's AI summaries (Bing). A tool that ignores these, or sells you a modelled estimate of the same thing, is asking you to pay for less accurate data. Setup guides: Search Console and Bing.

8. Does it tell you what to change, where, and how to verify it?

This is the check that separates a report from a tool. Industry surveys keep finding that teams buy AI-visibility data and then struggle to act on it. Look for findings that say what was measured, what to change, where the change lives (a template, a robots.txt rule, a CDN setting) and how to confirm it worked. Settings that live outside your code, like a firewall rule, should be called out as such rather than handed to a developer. And the tool should make re-checking a page trivial, so “fixed” means verified, not assumed.

On rewriting tools specifically: the original GEO study found that adding citations, quotations and statistics raised visibility in a test environment (Aggarwal et al.), and AutoGEO found that engines consistently prefer specific, well-structured, neutral and current content (AutoGEO). Neither supports promises of a fixed percentage lift on your site. Ask any tool claiming one for the controlled comparison behind it.

A 10-minute scorecard you can run on any tool

Start with one important page. Run a free audit here and in any other tool you're considering, compare the findings against this table, and keep the one that tells you the truth about your site, even when the truth is “this part is fine, the problem is elsewhere”.