An AI Search Audit in Practice: Our Findings, Fixes, and Retest
See Answer Patch’s own public-page audit: observed crawler-copy errors, practical fixes, and preview verification, with search outcomes kept separate.
An AI search audit should show what was observed, what changed, and what the retest actually proved. This case study uses Answer Patch’s own public website. It documents content and access checks, not a customer success story or a measured increase in traffic or AI citations.
Scope and evidence
On September 13, 2026, we fetched the 66 URLs in our public sitemap and inspected their HTML. All returned HTTP 200. We checked titles, descriptions, canonicals, headings, indexing directives, and links. These requests came from our audit connection; they do not prove how a firewall handles a vendor’s verified crawler IPs.
Download the public observation record for the three worked examples below. The record contains public URLs and selected metadata, without account data, paid reports, or private request logs. This was a manual content and HTTP audit, not an automated Fix Report run.
| Page | Observed issue | Selected change | Verification boundary |
|---|---|---|---|
| Crawler comparison | Older copy conflated Google-Extended with AI Overviews and overstated robots controls | Align explanations with vendor documentation | Verify the corrected page text; this does not test vendor behavior |
| llms.txt guide | Example lines were separated into list items | Provide one Markdown block and downloadable file | Compare rendered text with download; does not establish AI adoption |
| Sample report | Before/after title lacked a complete acceptance workflow | Add evidence, edit location, verification and outcome | Check all steps render; the plumber remains fictional |
Finding 1: successful fetching did not make the advice correct
The crawler comparison was accessible, but its table described Google-Extended as a Gemini and AI Overviews token. Google’s crawler documentation separates Gemini training and grounding uses from Google Search. We revised the comparison and glossary to make that distinction consistent.
We also corrected the statements about blocking user-requested visits. OpenAI says robots.txt may not apply to ChatGPT-User. Perplexity says its user-requested fetcher generally ignores those rules. The updated comparison records the distinction without promising that a robots rule is an access-control boundary.
Finding 2: an implementation guide needed an intact example
The llms.txt article already included format explanations and a fictional example. Its example was rendered as separate list items, making it awkward to reuse as a file. We added a preformatted block, a matching download, and HTTP troubleshooting steps for missing files, redirects, and HTML responses.
The llms.txt setup guide retains its existing URL and definition. The change improves the task a reader can complete. It does not turn llms.txt into a Google ranking requirement, and a successful file download does not show that an AI platform consumed it.
Finding 3: a sample needs a way to check the edit
The sample report already showed a fictional plumber’s title before and after an edit. We expanded it to identify the supporting fact, where the title is changed, and how the published page should be checked. The worked sample remains explicitly fictional and includes no invented customer outcome.
What the preview retest establishes
The implementation retest checks rendered content and downloads in a local preview. Code blocks preserve line breaks and escape markup; the llms.txt example matches its downloadable file. The sample report contains the evidence, edit, and verification steps. The revised crawler text distinguishes search eligibility from training and user-requested visits.
This verifies the prepared changes. It is not a post-deployment Google or AI-platform retest. Production crawl dates, Google index inclusion, referral visits, and citations require later observation. The public evidence record labels traffic and citation outcomes as not measured.
How to repeat this on your own site
Use the audit template to record a baseline, a specific fix, an owner, and a verification result. Start with the page that answers a customer’s question. A broad score without an evidence record is harder to act on.
The Technical Access pillar explains request checks. Our methodology describes what the product scores and how evidence is labeled. The AI search audit provides site-specific priorities when you need a Fix Report. Keep observed page changes and downstream search outcomes separate in either workflow.