Technical GEO Audit

Make your site readable by AI agents

When a buyer sends an agent to research you, it gets one fetch. Gumshoe runs its coded assessments against that same page and hands you the specific changes that let the agent read it.

See the audit

What an agent needs from your page

Agent Access

Is the door open when an agent arrives? robots.txt checked against named AI and search agents, plus meta robots and X-Robots-Tag noindex, plus the HTTP status.

• robots.txt
  • GPTBot Allowed
  • ClaudeBot Allowed
  • PerplexityBot Blocked

Fetch health

An agent gets one fetch, usually while a buyer waits for the answer. A bot wall, an error response, a redirect to the homepage, or a near-empty body stops the audit and gets reported instead of a score.

• Fetched cleanly
• the audit runs and returns a score
• Otherwise, reported instead of a score
  • Bot wall
  • Error response
  • Redirect to homepage
  • Near-empty body

Structured Data

Is there JSON-LD, and does every block parse? Plus the metadata an agent reads to work out what the page is: canonical, title, OpenGraph, Twitter card, published date.

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  ...
}
</script>
<script type="application/ld+json">
{
  "name": "Acme",
  "url": ,
}
</script>

Every block has to parse. One broken block is a finding.

Page Structure

Can an agent find the answer without guessing? One h1 and no skipped heading levels, main content volume against the markup around it, and a main or article landmark.

• H1
• H2
• H2
• Answer

What the audit gives you

1. Summary Score
   A score, the reasoning behind it, and the strengths and gaps that produced it.

   • Score: 78%
   • Evaluation: Very Good
   • Notes: Crawlable and structurally sound, but under-signaled where answer engines preview and attribute the page.
   • Strengths:
     • Returns 200 and is indexable
     • Valid JSON-LD is present
   • Gaps:
     • No Open Graph title or image
     • Very little text relative to markup

2. Component Breakdown
   Every assessment grouped into its category, each with a score and a pass or fail count.

   • Machine Readability & Data Layer up to 60% of score
     • Score: 71
     • Pass/Fail: 7 pass · 2 fail · 1 not applicable
   • Content Clarity & Answer-Readiness up to 27% of score
     • Score: 84
     • Pass/Fail: 5 pass · 1 fail · 0 not applicable
   • Authority & Trust Signals up to 9% of score
     • Score: 45
     • Pass/Fail: 1 pass · 1 fail · 0 not applicable
   • Multimodal & Agent Optimization up to 4% of score
     • Score: 90
     • Pass/Fail: 1 pass · 0 fail · 1 not applicable

3. Assessments
   Open an assessment and it shows what it looked for, how to fix it, and the evidence.

   • Open Graph essentials
     • Result: Fail
     • Importance: Answer engines use these fields to preview and attribute the page.
     • Solution: Add og:title, og:type, og:image and og:url to the head.
     • Evidence:

     <meta property="og:title"> not found
     

   • Canonical tag
     • Result: Pass
     • Description: Points at this URL, so the models converge on one version.

How you use it

1. Point it at a page Paste a URL and the audit runs against that one page. Start with the pages an agent is most likely to open when a buyer asks about you: home, pricing, product, comparisons.

   • Example: acme.com/pricing
   • Action: Run

2. Assessments run in code Every assessment runs against the fetched page in code, so the score is repeatable rather than a model's opinion. The summary sitting on top of it is AI-generated.

   • Focus on:
     • robots and status
     • structured data
     • headings and content

3. Findings with evidence Each finding names what the assessment looked for, what it found, and the change that fixes it. The evidence is what the fetch actually returned, close to what an agent gets back.

   • Examples:
     • Open Graph essential evidence
     • Text to markup ratio evidence
     • Publish date evidence

4. Fix, then run it again Ship the change and re-run the audit. The score moves with the page, and the audit carries the latest run and the date it ran, so progress is easy to track.

   • Example Scores:
     • Before: 71
     • After fix: 84

Questions about the Technical GEO Audit

• What is the difference between an AI agent and an AI crawler?
  A crawler visits ahead of time to build an index or training set. An agent visits in the moment because a buyer asked it to research a product, compare vendors, or pull a price... Both need the same things from a page: permission to fetch it, a clean response, and content they can parse.

• What does the Technical GEO Audit check?
  Every assessment is written in code and rolls up into four categories: Machine Readability and Data Layer, Content Clarity and Answer-Readiness, Authority and Trust Signals, and Multimodal and Agent Optimization...

• Which agents do you check?
  Eighteen named AI and search agents are checked against your robots.txt, including GPTBot and ClaudeBot...

• Why would an agent fail to read my page?
  Common causes include a robots.txt rule nobody remembers adding, a bot-protection layer that serves a challenge page instead of content, and pages that only render text after JavaScript runs...

• Why might the score move between runs?
  Every assessment is coded rather than judged by a model, so the assessments themselves are repeatable... The evidence on each check tells you which one moved it.

• Are all assessments weighted the same?
  No, each one carries its own weight, which is why a blocked agent counts for roughly ten times as much as a missing caption track...

• How many audits do I get?
  Starter covers ten page audits a month and Pro covers fifty; Pro also includes the Citation, Content, and Sentiment Audits...

• How is it different from the Content Audit?
  The Technical GEO Audit is focused on whether an agent can access and parse a page, while the Content Audit focuses on whether the content answers buyer queries.