How to Run an AI SEO Audit in 2026: Technical, Content and AI Visibility Checks

Key takeaways

• An AI SEO audit is a normal SEO audit plus two extra questions: can AI crawlers reach your pages, and does the answer actually name you?
• Do the work in order: access first (robots.txt, CDN, WAF, rendering), then content, then AI visibility. Fixing content on pages bots can't reach is wasted effort.
• ChatGPT cites product and landing pages far more than blog posts. According to Promptwatch Data (Sep 28 to Oct 4, 2026), product pages are 22.4% of citations and landing pages 16.8%.
• A spreadsheet and free chat interfaces cover a baseline. Tools become worth paying for once you need to repeat the audit monthly.
• Run a light audit quarterly and a full one after migrations, redesigns, or traffic drops.

What an AI SEO audit actually is

A classic SEO audit asks whether your pages are crawled, indexed, and ranked, then turns the problems into a prioritized fix list. An AI SEO audit keeps all of that and adds a question the older guides skip: when someone asks ChatGPT, Perplexity, Gemini, Claude, or Google AI Overviews about your category, does the answer mention or cite you?

Google itself says established SEO practices still apply to AI Overviews and AI Mode. So this isn't a replacement. It's a third layer on top of two you already know.

One honest caveat before we start. Much of what circulates about AI visibility is half-sourced. One audit guide I read points out that a popular "well-organized headings earn 2.8x more citations" stat traces back to a company citing itself, and the report it links never mentions headings. Treat round-number claims with suspicion, including mine. Where I can, I use first-party data with a stated time window.

Before you start: scope and baseline

Write down what "the site" is. Domains, subdomains, locales, CMS, CDN, and JavaScript framework all matter, because each one changes how bots see you. Then record a baseline so later fixes can be judged against something.

• Search Console: organic clicks, impressions, top queries, top pages, recent changes
• Bing Webmaster Tools: same data, since Bing feeds several AI products
• Analytics: sessions and conversions from any AI referrers you can already identify
• A list of your 10 to 20 revenue pages. Everything below prioritizes these.

Pick your priority pages now. A full audit of a large site can take days, and you will not fix everything. The pages that make money get audited first.

Layer 1: technical checks

Technical failures rarely look technical. They show up as "we need more content" or "rankings dipped." A useful way to separate them is three buckets: a visibility problem (the page can't be crawled or indexed), an interpretation problem (crawled, but rendered or understood inconsistently), and a priority problem (understood, but another page or site is chosen instead). Each bucket has different fixes and timelines.

Crawl and index

1. Crawl every accessible URL with a crawler such as Screaming Frog or Sitebulb. Look for broken links, redirect chains, duplicate titles, missing headings, thin pages, and orphaned content.
2. Compare what you crawled against what you expected to exist. The gap is usually informative.
3. Open the Page indexing report in Search Console. Exclusions can be valid, but valuable pages should not be sitting there by accident.
4. Test robots.txt. It controls crawling, not indexing, so don't use it in place of noindex.
5. Keep only canonical, indexable, 200-status URLs in your XML sitemap.
6. Check canonical consistency, HTTPS everywhere, and no mixed content.

Rendering

Crawl twice: once as raw HTML, once rendered. If key content only appears after JavaScript runs, some AI crawlers may never see it. Check this on templates (product page, category page, article), not random URLs. Template-level fixes scale; one-off fixes don't.

Performance

Validate Core Web Vitals on your key templates, especially mobile. Slow, unstable pages hurt users first and everything else second.

Structured data and entities

Confirm schema matches what is visibly on the page. Check Google's structured data documentation for what's supported. Then check entity consistency: organization name, logo, and sameAs links should match across templates, author pages should exist where expertise matters, and you shouldn't call the same product five different things across the site.

For entity and schema automation, InLinks and WordLift are two options worth a look.

Layer 2: AI crawler access

This is the check most classic audits miss, and it's the one that can quietly zero out everything else. If AI crawlers can't reach your content, they can't interpret or cite it.

What to review

• robots.txt rules for GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, and Google's crawlers
• CDN bot filtering (Cloudflare and others often block unknown bots by default)
• WAF and firewall rules
• JavaScript-only rendering on key pages

OpenAI advises publishers to allow OAI-SearchBot if they want content considered for ChatGPT Search summaries and snippets. Note the distinction: training crawlers and answer-time fetchers are different bots. You may reasonably block one and allow the other. Decide on purpose rather than by accident.

Use your logs, not assumptions

robots.txt tells you what you intended. Server and CDN logs tell you what happened. Compare the two.

Promptwatch Data tracked IP-verified AI crawler requests from Jun 8 to Sep 6, 2026. OpenAI made up 79.8% of verified requests in the week of Aug 31 to Sep 6, down from 94.8% in Jun 8 to 14. See the full AI crawler traffic report. Those are relative shares, not absolute volumes, and provider totals mix training crawlers with answer-time fetchers. Still, the report gives you a practical test: if a provider is large in the benchmark but near zero in your own logs, you are probably blocked or unreachable for it.

Also watch for newcomers. Promptwatch Data shows Meta-WebIndexer going from about 2% to nearly 38% of tracked AI crawler requests between mid-July and August 9, 2026. A robots.txt written in 2024 won't mention it.

Verify bots by IP, because user agents are trivially spoofed. If you're pulling logs manually, that means checking against each operator's published IP ranges. Promptwatch does this verification for you and shows which pages each crawler read and which errors it hit.

llms.txt

Adding an llms.txt file is cheap, so do it if you like, but treat it as a clarity gesture. Nobody has shown it moves rankings or citations. Don't spend a sprint on it.

Markdown versions

Promptwatch Data found that markdown files are just 0.05% of AI search citations. Markdown serves AI agents, not AI search. Your normal HTML pages are what get cited.

Layer 3: content checks

Cannibalization, decay, and gaps

Review content for overlap and rot. Find zombie pages with no traffic, thin pages that should be merged, and two pages fighting for the same query. Then find gaps: topics your audience searches that you haven't covered. This is a content audit in miniature, and it should end with four buckets: keep, refresh, merge, prune.

Match content type to what AI cites

This is where the AI layer changes decisions. Promptwatch Data's ChatGPT citation share by content type (Sep 28 to Oct 4, 2026) shows what gets cited:

Product and landing pages together are 39% of citations. For commercial prompts it's even sharper: product pages hit 26.0% and landing pages 20.6%. If your audit finds thin, vague product pages and a 200-post blog, that's the problem right there. Audit your product and landing pages for complete specs, prices, and clear positioning before you write another article.

Other implications from that data: keep documentation public and crawlable, add FAQ pages, and for branded prompts make sure reviews exist on trusted third-party sites.

Freshness and decay

Content doesn't stay cited forever. In Promptwatch Data's content half-life study, every page ChatGPT cited on July 6, 2026 was followed for 90 days:

News hit its half-life at about 36 days, and almost a quarter of news pages were never cited again. Product pages, reviews and how-tos reached half-life at 74 to 77 days. So build a refresh schedule by type: keep prices and specs current on product pages, treat landing, pricing and docs pages as long-term assets, and refresh listicles and guides before they drop out. Use news for short bursts, not as a foundation.

Structure and evidence

Check priority pages for answer-first openings, a clean H2/H3 hierarchy, tables and lists where they help, embedded citations and statistics, and real author signals. Content that is factual, specific, and well referenced is more likely to be used as a source. Marketing fluff before the answer is the common red flag.

For trust, make authors, company details, policies, and contact information easy to verify. Cite original sources. Trust comes from evidence, not repeated claims about expertise.

Content optimization tools like Clearscope and Surfer SEO can help with the on-page side.

Layer 4: AI visibility checks

Now the part that's new. You're measuring whether the answer names you.

Step 1: build a prompt library

Write about 30 prompts in buyer language, not internal jargon. A workable split: 8 category-definition prompts, 8 comparison prompts, 8 recommendation prompts, 6 use-case prompts. Include branded and unbranded versions.

Step 2: test across platforms

Run every prompt in ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews. Log whether you're mentioned, your position in the answer, the sentiment, and which sources were cited. Do the same for your top three competitors. A spreadsheet is enough for a baseline, and the manual version takes roughly 6 to 10 hours.

Two limits of the manual approach. Answers vary run to run, so one test proves little. And consumer interfaces can differ from API outputs, so test the real UI.

Step 3: see which pages get cited

For every citation to your domain, note which page it was. Pages that rank well in Google but are never cited by AI are your best opportunities, and pages cited often deserve protection and refreshing.

Step 4: map off-site influence

AI answers lean on sources you don't control: review sites, forums, editorial listicles, news. Count your presence on G2, Trustpilot, Capterra, Reddit, YouTube, and LinkedIn, and map the same for competitors. Under 10 reviews or a rating below 3.5 is a red flag in most categories.

Be careful with social assumptions, because they shift quickly. Promptwatch Data shows reddit.com's share of ChatGPT citations fell from about 4% to 0.5% on August 14, 2026, and for Sep 28 to Oct 4, 2026 ChatGPT got 0.46% of citations from social platforms while Google AI Overviews got 12.2%. Your off-site strategy should differ by engine.

Step 5: check sentiment and accuracy

Read what the answers actually say. Wrong pricing, outdated features, or a competitor positioned as the default are all findings. Each one maps to a page you can fix or a third-party source you need to influence.

Choosing tools for the audit

You can start free. Once you need to repeat this monthly across many prompts and countries, tooling pays for itself. Here is how the categories break down.

A few notes. Many AI visibility tools are monitors: they tell you whether you were mentioned. That's useful for a baseline, but the audit's real output is a fix list. Promptwatch goes further than most, with crawler logs, page-level citation data, a site audit, and an Action Board that turns findings into prioritized tasks, so it suits teams who want to audit and fix in one place. Its Essential plan starts at $95/month and there's a free Explore tier; check promptwatch.com/pricing for current details.

For technical SEO at enterprise scale, Botify is the established choice.

If you already live in a traditional suite, Semrush and Ahrefs Brand Radar bolt AI tracking onto familiar workflows, though with more limited AI-specific data.

Lighter-weight trackers such as Peec AI and Otterly.AI cover prompt monitoring at lower cost.

To compare more options, the GEO software directory at bestgeosoftware.com and the AI rank tracker list at ai-rank-tools.com are good starting points.

Score it and prioritize

A score makes the audit operational instead of subjective. Here's a simple 100-point rubric you can adapt:

Then rank every finding on two axes: impact (how many important URLs or templates are affected) and effort. Fix the gates first, meaning crawl, index, and bot access. Everything else gets easier after those are stable.

A 30/60/90 plan

• Days 1 to 30: unblock crawlers, fix indexation errors, repair product and landing page facts, correct inaccurate AI claims at the source.
• Days 31 to 60: refresh decaying pages, add FAQ and comparison content, implement schema, start review and editorial outreach.
• Days 61 to 90: re-run the prompt library, compare against baseline, and decide what to double down on.

How often to repeat it

Run a light audit quarterly to catch decay early. Run a full one after any site migration, redesign, traffic drop, or major algorithm update. Re-run your prompt set at least monthly if AI search matters to revenue, since citations move faster than rankings. Promptwatch Data has shown ChatGPT behavior shifting overnight in 2026, including the site: operator rollout on August 8, where site: queries jumped from about 0.4% to about 17% of fanout queries.

When to bring in help

If you'd rather not build the process yourself, an agency can run it. 1001 SEO Media handles technical SEO, content, and AI search work together, and the audit above is close to how that kind of engagement starts. Whoever you pick, ask for ticket-ready findings your developers can ship, not a PDF of tool scores.

The short version

Check access before content, and content before AI mentions. Verify bot behavior in your logs instead of trusting robots.txt. Put your best effort into product and landing pages, because that's where the citations are. Then measure the answers, fix what you find, and run it again in a quarter.