Weekly Site Audit for AI Search: How to Read the 21 Checks, Health Score and Change Log (2026)

A practical walkthrough of a weekly AI search site audit: what the 21 checks mean, how the health score is calculated, why it can mislead, and how to use New badges and the change log to catch regressions.

Key Takeaways

  • A weekly audit that fetches every URL fresh and runs 21 checks tells you how pages respond to crawlers now, not how they looked when last indexed.
  • The health score is the percentage of checked URLs with no error. Warnings and notices do not count against it, so a high score can hide a lot of work.
  • Work errors first, then warnings on pages AI crawlers actually visit, then notices. Skip the rest until the first two are clean.
  • The red New badge is the most useful part of the change log: it marks regressions that appeared since last week.
  • Ignoring URLs changes the denominator. A score that rises after you add ignore rules has not been fixed, only reshaped.

Traditional audits ask whether Google can crawl and index a page. An AI search audit asks a slightly different question: can an AI crawler fetch this page, read it, and use it in an answer? A page can fail that test for boring reasons. A 5xx at the wrong moment. A missing title. A shell of HTML that only fills in after JavaScript runs.

Sites change every week. Developers ship, editors delete pages, certificates expire, a redirect rule gets added and breaks a whole folder. A one-off audit is a snapshot. A weekly one is a diff, and the diff is where most of the value is.

This guide uses the Site Health audit in Promptwatch as the worked example, because its 21 checks, score and change log are documented in detail. The same reading habits apply to any crawler-based audit, and I point out where Semrush and others differ.

Favicon of Promptwatch

Promptwatch

Track and optimize your brand's visibility in AI search engines
View more
Screenshot of Promptwatch website

How the audit runs

Each weekly run takes every URL in the project's page inventory, fetches it fresh, runs 21 checks, and produces three things: a health score, a per-URL issue list, and a comparison with the previous run. You find it under Site, then Health.

One detail trips people up. The audit does not reuse the content crawl behind the Pages view. The Redirects and Failed tabs on the Pages page list what the content crawl saw, not what the audit saw. If the two disagree, that is expected. The audit shows how pages respond today.

A weekly email also goes out every Tuesday per project. It contains the health score, the number of issues fixed since the previous audit, and the top issues AI crawlers may trip over, with a link back to Site Health. You can turn it off in email preferences.

The 21 checks, grouped by severity

The checks fall into three severity levels. The tooltip definitions are blunt, which I like: errors mean AI crawlers cannot fetch or read the page, warnings mean the page is harder to read and cite, notices are minor cleanup.

Errors (7)

  • Page returns 4xx
  • Page returns 5xx
  • Broken internal link on page
  • Redirect loop or chain longer than 3 hops
  • Missing page title
  • Missing H1
  • SSL certificate invalid or expired

These are the ones that stop a crawler cold. A missing H1 sits oddly next to the status codes at first glance, but it makes sense: without a title and a primary heading, a retrieval system has little to anchor the page to.

Warnings (9)

  • Redirecting URL in sitemap
  • Orphan page (no internal links)
  • Missing meta description
  • Multiple H1s
  • Thin content (under 300 words)
  • Slow server response (over 1000 ms)
  • Oversized HTML (over 2 MB)
  • SSL certificate expiring in under 14 days
  • Content needs JavaScript to appear

The last one deserves attention. It means the HTML a crawler receives is a shell. In a browser the page looks fine. For a crawler that does not execute your scripts, there is nothing to read. Promptwatch Data's markdown in AI search report (February 2026, last 7 days of citations) found HTML makes up 99.94% of citations across ChatGPT, Claude, Perplexity and AI Overviews. So the HTML your server returns is what gets cited, and rendering problems come first on the fix list.

Notices (5)

  • Linked but not in sitemap
  • Title length out of range (15 to 70 characters)
  • Canonical points to an external domain
  • Near-duplicate of another page
  • Linked only from site-wide navigation

Notices are housekeeping. Some are worth a look, such as an external canonical you did not intend, but none of them affects the score.

Reading the health score

The health score is the percentage of checked URLs with no error. The page also shows the change against the previous scored run and an "N of M URLs checked" line, so you can see how much of the site the number actually covers.

The docs say notices do not affect the score, and the definition implies warnings do not either. Only errors count. That makes it a strict-but-narrow measure: it answers "how many pages are unreadable to a crawler" and nothing else.

A stacked bar for each run splits URLs into Healthy, Notices, Warnings and Errors. This view is more honest than the single number. You want to see the warning and notice segments shrinking week over week, not just the error segment staying thin.

Why one number is never the whole story

A German SEO consultant, Christian Ott, published an audit write-up in October 2026 that makes the point well. His own audit script gave seo-kreativ.de a health score of 100 on 18 August 2026, covering 209 URLs, and reported 1,092 findings in the same run. Both facts were true, because the score only counted the share of URLs without a serious finding. His advice: take your largest group of findings, check one example live, and if it is a false alarm, check whether the whole group shares the same cause.

Screenshot of an SEO audit checklist article discussing why a health score of 100 can coexist with over a thousand findings

That habit transfers directly. A score of 98% in a weekly AI audit is reassuring, but 2% of a 5,000-page site is 100 broken pages, and if your best-converting product page is one of them, the percentage is cold comfort.

How other tools compute it

Scores are not interchangeable. Semrush's Site Health Score is a 0 to 100% figure based on errors and warnings relative to the number of checks performed, with errors weighing more than warnings. SE Ranking's runs 1 to 100 and reflects issue volume and severity. Promptwatch's is simpler: errors only, per URL.

ToolScore rangeWhat countsAI-specific extras
Promptwatch Site HealthPercentage of URLs with no errorErrors onlyCrawler-focused checks, plus crawl and citation data on the same platform
Semrush Site Audit0 to 100%Errors weighted above warnings, relative to checks performedAI Search Health widget, Blocked from AI Search widget, llms.txt notices
SE Ranking1 to 100Issue volume and severityWebsite Audit API for Slack alerts on score drops
ScrunchContent alignment scorePage content versus tracked promptsPage-level Deep AI Audit
Favicon of Semrush

Semrush

All-in-one digital marketing platform
View more
Favicon of SE Ranking

SE Ranking

All-in-one SEO platform with AI visibility toolkit
View more
Screenshot of SE Ranking website

Thresholds differ too. Semrush flags low word counts under 200 words where Promptwatch flags under 300, and both flag HTML over 2 MB and redirect chains over 3 hops. Do not compare scores across tools. Pick one and trend it.

Reading the issues table

Below the score is an issues table with one row per URL and failed check. Columns are URL, Issue, Severity, Status, TTFB and Checked. You can filter by severity tab (All, Errors, Warnings, Notices), narrow by issue type with the "Search checks" box, and page through 25 rows at a time. Re-check and Export buttons sit above it.

Click a URL and you land on its Performance Insights page, with Health and Content & SEO sections. That is where you go to see what is actually wrong rather than guess from a row label.

A practical way to use the table: sort mentally by issue type, not by URL. If 400 rows say "Broken internal link on page" and they all point to the same deleted URL, you have one fix, not 400.

The change log: New badges and Runs failed

This is the part I would read first every Tuesday.

A red New badge marks a URL that did not fail that check in the previous run. In other words, a regression. Something changed on the site since last week. A deploy, a CMS edit, a plugin update, an expired certificate. The badge turns an audit from a pile of problems into a short list of what is new.

The Failing URLs list adds a Runs failed column. A URL that has failed five runs in a row is a neglected problem. A URL that failed for the first time this week is probably a fresh break and probably cheap to fix, since someone on your team touched it recently.

The email's fixed-issues count closes the loop. If your developers say they fixed the redirect chains and the fixed count does not move, the change did not land.

What to fix first

The recommended order is simple and I would not fiddle with it.

  1. Errors first. Start with 4xx and 5xx in Failing URLs, then broken internal links, redirect chains, missing titles and H1s, then certificate problems.
  2. Warnings on pages that matter. Cross-reference with your most-crawled pages in AI crawler insights. A heavily crawled page with thin or JavaScript-dependent content wastes crawler attention. For speed, use the slow server response check and the Slowest pages card.
  3. Notices last. They are cleanup.

The cross-reference in step 2 is the useful bit. Not every warning deserves a ticket. A thin-content warning on an old tag archive is noise. The same warning on a page AI bots visit daily is a missed opportunity.

Promptwatch's AI crawler logs show which pages bots read and where they fail, which is how you decide which warnings sit on high-value pages. Promptwatch Data's AI crawler traffic report (verified requests, June 8 to September 6, 2026) puts OpenAI at 79.8% of requests in the week of August 31 to September 6, down from 94.8% in early June. If a provider is large in that mix but near zero in your logs, you are probably blocked or unreachable for it, and a site audit alone will not tell you that. Check robots.txt, CDN and WAF rules.

Using ignore rules without fooling yourself

Since November 13, 2026 you can tell the audit to skip a page, a folder, or any URL matching a regex. Ignored URLs are not fetched, scored or listed from the next audit onward. You set rules from the ignore menu on issue rows and failing URLs, and they are listed under Project settings, then Sitemap. The dialog previews which URLs match.

Good candidates: old versioned docs, tag archives, paginated listings. Use a path rule for whole sections, since new pages under that path are skipped as well.

The pitfall is the one that matters. Ignoring URLs shrinks the denominator, so the score can rise with no real fix. Ignored pages stay in the Pages list and are still read for content, and the homepage is always audited. The Actions cards and the audit email do not count newly ignored pages as fixed, which is the right call. Treat a jump in score after adding rules as an accounting change and note it somewhere, or you will confuse yourself in three months.

After you fix things

Re-check the touched URLs, or run a full audit. Then watch the score and the New badges on the next run. If New badges appear on pages you just edited, your fix introduced something.

Pair the audit with a visibility baseline. As one consultancy puts it in its AI search readiness checklist, an audit without a baseline produces fixes nobody can evaluate. Track a fixed set of prompts weekly across ChatGPT, Perplexity and Google's AI surfaces, record whether your brand appears, and rerun the same prompts after fixes ship. Single spot checks vary from run to run, so one chatbot answer proves nothing.

Automating alerts

If you manage several sites, do not rely on remembering to open a dashboard. Promptwatch's MCP server includes a listSiteAuditRuns tool that returns runs newest first with health score, coverage and issue totals by severity, which suits alerting or a dashboard. SE Ranking's Website Audit API supports a similar pattern: a Slack alert when a high-priority page's score drops past a threshold. Workflow tools like n8n or Zapier can glue this together.

Favicon of n8n

n8n

Open-source workflow automation with AI agents and code flex
View more
Screenshot of n8n website

A simple Tuesday routine

  1. Open the weekly email. Note the score change and the fixed-issues count.
  2. Go to Site, then Health. Filter to Errors. Look for New badges first.
  3. Group by issue type. Check one example live before assigning work.
  4. Filter to Warnings. Cross-reference against your most-crawled pages and fix those.
  5. Skim notices once a month.
  6. Re-check touched URLs after deploys.

That is about 20 minutes once the backlog is cleared.

Other crawlers worth knowing

A weekly AI audit does not replace a full technical crawl. For deep structural analysis, desktop and cloud crawlers such as Screaming Frog and Sitebulb go much further on things like rendering comparisons and internal link graphs. Promptwatch also has a separate free AI SEO audit tool that checks robots.txt access for AI crawlers, SSL, sitemap, llms.txt, meta tags, canonicals, JSON-LD and more. It is not the same thing as the weekly 21-check audit, so do not mix their results up. If you want to compare GEO platforms and what their audits cover, the directory at bestgeosoftware.com is a decent starting point.

Favicon of Screaming Frog

Screaming Frog

Industry-leading website crawler for technical SEO audits
View more
Screenshot of Screaming Frog website

What the audit will not tell you

It will not tell you whether AI engines cite you. It will not tell you whether content is good. It is a readiness check: the floor, not the ceiling. A clean score means crawlers can read your pages. Whether they choose to cite them depends on content, authority and what the prompts ask, which is why tracking citations and crawler activity next to the audit gives you a much better picture than the audit alone.

If you want to try this on your own site, the Promptwatch free trial runs for 7 days on monthly billing, and the free AI visibility report is a no-login way to see crawler access first. Confirm current plan limits at promptwatch.com/pricing, since audited page counts vary by plan.

Share:

© 2026 AI Search Tools · Best AI search tools and platforms · RSS

AI Search Tools is an affiliate review site. When you click links to vendors or buy through links on our site, we may earn an affiliate commission at no extra cost to you.

AI Search Tools is a review website based on user reviews on Reddit and G2, and on publicly available information. We keep everything as up to date as possible, but pricing and features can change. Always confirm the details with the vendor before purchasing.