In a November 2026 audit prepared for a technology publisher, we checked all 66 landing URLs listed in its XML sitemaps. Every one returned HTTP 200. Each had a title and a self-referencing canonical tag. None declared noindex in the fetched HTML.
The publisher was still concerned about low Google Search impressions.
Those findings narrowed the investigation, but they did not explain the visibility problem. We could describe what the website served. Without the publisher’s Search Console data, we could not establish which pages Google had indexed or which searches they appeared for.
That distinction matters when you verify AI-generated SEO advice. A model can turn a collection of observations into a confident diagnosis. The reviewer has to decide whether the evidence supports that diagnosis before anyone changes the site.
Several checks in this audit show what that review involves.
A working page does not prove Google indexed it
A 200 response establishes that a server returned the requested page successfully. It does not establish that Google indexed the content, selected that URL as canonical, or considers it relevant to a particular search.
The same limit applies to canonical tags. Finding a self-referencing canonical tells you which URL the website prefers. Google’s selected canonical is a separate observation. Google’s canonicalization guidance describes these signals and how they help indicate a preferred URL.
In our review, the defensible finding was that the sitemap pages were accessible to our crawler and did not declare noindex. “All pages are indexed” would have gone beyond the evidence.
A useful technical SEO audit should preserve that distinction: name the affected URL, show what was observed, and identify the next check when the answer depends on Google data.
When an AI draft says a page is “not indexed,” ask for the URL Inspection result or other direct evidence behind that statement. If the draft only checked the page’s HTML, revise the claim.
Also Read: ChatGPT vs Claude vs Gemini for Coding: Which Wins?
Conflicting link checks need another test
We checked 138 unique external URLs linked from article bodies. Four returned 404 responses in direct requests and again during verification.
That initially looked straightforward. However, two FTC consumer resources remained readable through a separate web search and fetch service. The evidence disagreed.
We recorded the conflict and left those references in a browser-verification queue. We did not tell the publisher that every flagged citation was conclusively dead.
This is a useful test of an AI-generated audit: does it retain uncertainty when different tools produce different results?
A response from an automated request describes that request. Access controls, request handling, temporary failures or cached material can complicate the interpretation. A recognizable source name does not resolve the disagreement either.
Before replacing a citation, record the exact address, response, final destination and normal-browser result. Then check whether the source still supports the sentence. Replacing a legitimate reference with a less relevant page can make an article worse even if the replacement loads successfully.
Keep sample findings inside their limits
The publisher’s 52 articles linked to 738 distinct tag URLs. We fetched 94 tag archives, selected alphabetically, and found mixed settings: 43 allowed indexing and 51 declared noindex. Of the 43 indexable samples, 38 listed just one article.
That justified a taxonomy review. It did not justify saying all 738 tag pages were indexed or that tags caused the low impressions.
The distinction affects the fix. Deleting every tag would ignore potentially useful archives, navigation and existing search performance. The recommended next step was to inventory the terms, check their value and decide which archives to retain, improve, consolidate or noindex.
When reviewing an AI draft, underline every phrase containing “all,” “every” or “sitewide.” Check whether the underlying work covered that scope. A sample should remain a sample in the recommendation.
If noindex is the chosen treatment, implementation needs care. Google’s noindex documentation explains that Google must be able to access a page to see the directive. Blocking the same URL in robots.txt can prevent that.
Check what a screenshot can actually establish
We also inspected the homepage and one article at a 390-pixel mobile viewport. The menu expanded and exposed the category links. On the article, document scroll width matched the viewport width, and the article occupied 340 pixels.
That supported a narrow finding: no horizontal overflow was observed in the inspected article view.
It did not establish that every page, table, embed and form worked on every phone. It also said nothing about Core Web Vitals.
The public PageSpeed API requests returned quota errors, so the report did not include a performance score. There was no measured basis for one.
If an AI audit assigns a speed grade after viewing a screenshot, ask for the test URL, device settings, date and results. If a test failed, the draft should report the failed measurement rather than fill the gap with a plausible number.
Give the revision a source record
The most useful material to give an AI assistant is a small record of verified findings. For each recommendation, include:
- The exact URL and inspection date.
- The relevant response, HTML excerpt or rendered observation.
- Any conflicting result.
- The action supported by that evidence.
- The check required after implementation.
Then ask for a bounded rewrite: use these observations, retain their limits, and flag unsupported explanations.
For example, “38 sampled indexable tag archives contain one article” is reproducible. “Tag bloat is destroying rankings” adds a causal claim the crawl did not establish. Editing the latter into more polished language would not repair its evidential weakness.
The final review should also distinguish proposed work from completed work. Our audit identified fixes; it did not establish that those changes had been implemented or that traffic had improved afterward.
Before acting on AI-generated website advice, pick one recommendation and try to reproduce it. Open the named URL, repeat the relevant check and compare the result with the wording. If the recommendation survives that process, you have a concrete starting point. If it does not, correct the evidence record before expanding the task list.
About the contributor: SERP Radius is a Toronto SEO agency providing technical SEO, on-page optimization, keyword research and SEO consulting.
FAQS
AI can group crawl findings and draft recommendations. Verify important claims against the exact URL, the tool output, and the date of the check before acting. Google’s guidance on generative AI content
Enter the exact URL in Search Console and review its indexed status. Use Test Live URL to check whether Google can access the current page and whether it might be indexable. Google’s URL Inspection guide
Compare both URLs in Search Console. If the pages are duplicates and your preferred URL is correct, point the canonical tag, internal links, and sitemap consistently to that URL. Google’s canonicalization guidance
Open the exact URL with Chrome DevTools’ Network panel open. Enable Preserve log, reload, and inspect the final status and destination after any redirects. Chrome DevTools Network reference
Test the exact URL in PageSpeed Insights and match the reported device type. Use Lighthouse lab results to diagnose issues and field data, when available, to assess real user experience. Google’s PageSpeed Insights guide