Technical SEO Audit with AI: Checks You Can Run Yourself

Technical SEO has a reputation problem: marketers treat it as developer territory, so the basics go unchecked for years. The data says the basics are exactly what is broken. When Ahrefs audited just over a million domains, 72.3% had slow pages, 72.9% had missing or empty meta descriptions, and 59.5% had missing or empty H1 tags. Those are not exotic rendering bugs. They are the kind of problems a marketer can find in an afternoon, and AI has removed most of the excuse for not looking.

My claim in this post is narrow and I want to state it precisely: ChatGPT or Claude cannot crawl your site, but it can read every artifact your site produces - the robots.txt file, the page source, a crawl export, a PageSpeed report - and turn each one into findings and a prioritized fix list in plain English. That translation layer used to be the consultant. This post is the set of checks I actually run this way, part of my AI SEO playbook.

What the model can and cannot see

The ground rule first, because it decides everything else: a chat model has no access to your website. When you paste a URL into ChatGPT and ask “audit this site,” anything it tells you without fetching the page is pattern-matched guesswork wearing an audit’s clothes. I have watched it confidently report title tags that do not exist.

So every check below follows the same shape. I collect an artifact myself - a file, a source view, an export, a screenshot of a report - paste it in, and ask for analysis of exactly what I pasted, with the instruction to say “not visible in what you gave me” rather than fill gaps. Collecting the artifacts requires a browser and the ability to copy and paste. That is the whole technical bar, which is why I think this belongs in a marketer’s hands and not in a developer’s backlog.

Check 1: read your robots.txt like Google does

Your robots.txt lives at yourdomain.com/robots.txt, and reading it takes one paste. This check earns its place through the failure it catches: rules that quietly block sections you want indexed, usually left over from a migration or a staging setup. The 2025 Web Almanac, which analyzes millions of sites, found 13.3% of desktop sites have no robots.txt at all - which is survivable - while a wrong one is far worse than none.

My prompt:

Here is my robots.txt. For each rule, explain in plain
English what it blocks or allows and for which crawlers.
Then flag: (1) rules that could block pages a marketing
site would want indexed, (2) rules that do nothing,
(3) whether AI crawlers like GPTBot or ClaudeBot are
blocked. Do not guess about anything not in the file.

[PASTE FILE]

One correction the model itself will usually make, because Google’s own documentation is blunt about it: robots.txt “is not a mechanism for keeping a web page out of Google.” Blocked pages can still be indexed if other sites link to them. If something must stay out of search results, that is a noindex tag, and asking your developer for one is a ten-minute request.

Check 2: canonicals, the duplicate-content traffic cop

A canonical tag tells Google which version of a duplicated or near-duplicated page is the real one, and about a third of sites do not use them - the Web Almanac puts implementation at 68% of desktop pages. I check canonicals second because the failure is invisible in the browser: the page looks fine while its signals are split between www and non-www, slash and no-slash, or a tracking-parameter twin.

The artifact here is the page source. Right-click any important page, choose “view page source,” copy everything in the head section, and paste it with: “Find the canonical tag. Does it point to this exact URL? List every other version of this URL that might exist (http vs https, www, trailing slash, URL parameters) and tell me what to check for each.” Run it on your five most important pages, not your whole site.

Worth knowing while you read the answer: Google treats the canonical tag as a strong signal, not a command - if you specify nothing, Google picks a version itself. Usually it picks sensibly. The audit is for the cases where it cannot, because your URLs genuinely duplicate.

Check 3: titles, descriptions and headings at scale

This is the check with the biggest surface area, and the one where AI changes the economics most. One page’s title tag is a glance. Three hundred pages’ title tags are a spreadsheet job nobody does, which is how the Ahrefs numbers above get so bad: 63.2% of sites have titles that run too long, more than half have multiple H1s.

The artifact is an export. The free tier of any desktop crawler (Screaming Frog’s crawls 500 URLs at no cost) produces a CSV of every URL with its title, meta description and H1. If installing a crawler is past your line, your developer or SEO can generate this file in minutes, and the analysis stays yours. Paste the CSV in chunks:

This CSV has URL, title, meta description and H1 for my
site's pages. Report: (1) exact duplicate titles and
descriptions, grouped, (2) titles over 60 characters and
descriptions over 160, (3) pages with a missing or empty
H1 or description, (4) titles that are generic enough to
be any site's ("Home", "Blog", "Products"). Rank the
findings by which pages look most important based on
their URLs. Output a fix table I can work through.

I treat the output as a to-do list, not a verdict, and I rewrite the important ones myself. The model finds the problems at a speed I cannot match. It does not know that one of those pages drives a third of my signups, so the ranking pass at the end is a convenience I always re-order. The same export, batch, review economics apply to image alt text, which gets its own workflow because the fix runs through a vision model instead of a CSV.

Check 4: schema, generated and validated

Structured data is the one technical item I have already written up in full, so this section stays short: my guide to generating schema markup with ChatGPT covers producing JSON-LD safely and validating it before it ships. In audit mode the direction reverses. I paste the existing schema from the page source and ask what type it declares, what required properties are missing, and whether anything in it - a rating, a price, an author - claims something the visible page does not support. That last question matters more than validity. Schema that overclaims is a manual-action risk, and a validator will not catch it because dishonesty parses cleanly.

Check 5: Core Web Vitals, translated

Speed is the finding marketers most often get handed and least often understand, so my audit treats the AI as a translator rather than a tester. The test itself is free: PageSpeed Insights takes a URL and returns your scores. What the model adds is reading the report for you.

The thresholds worth knowing come from Google’s Core Web Vitals documentation: a good Largest Contentful Paint is 2.5 seconds or less, a good Interaction to Next Paint is 200 milliseconds or less, and a good Cumulative Layout Shift is 0.1 or less, each measured at the 75th percentile of real visits - meaning most of your traffic has to hit the target, not your fastest visitor once. I paste the full PageSpeed output and ask: “Which of the three Core Web Vitals fail, what is the single biggest contributor to each failure in this report, and write me the request I should hand my developer for each - specific enough that they know what to change and can estimate it.” The output of this check is not a fix. It is a well-formed ticket, which for a non-technical marketer is the difference between “the site feels slow” and a request a developer acts on this sprint.

Check 6: the sitemap sanity pass

Last and quickest. Your sitemap usually lives at /sitemap.xml, and the audit question is simple: does it list what you actually want indexed, and only that? I paste it (or its index file) and ask for a count by URL pattern, anything that looks like a test, tag or paginated page, and anything important I know exists that is absent.

Keep the model honest about what a sitemap does, because Google’s documentation is explicit: it helps discovery but “doesn’t guarantee that all the items in your sitemap will be crawled and indexed.” A clean sitemap is hygiene, not a growth lever. If your pages are indexed and you have under a few hundred of them, this check is a five-minute confirmation, and what Search Console says about your AI-search visibility will tell you more about where the traffic actually went.

What I hand to a developer, and how

Three of the six checks can end in a developer request: canonical fixes, speed work, and any noindex or redirect changes. The audit’s real deliverable is the quality of those requests. “Improve site speed” dies in a backlog. “LCP on the homepage is 4.1s against Google’s 2.5s threshold, and the report attributes 2.3s of it to the uncompressed hero image - can we serve it as WebP at display size?” gets scheduled, because the model turned the report into exactly that sentence and I checked it against the numbers.

Everything else - titles, descriptions, headings, schema content, sitemap cleanup - is contentwork, and I keep it. The division is the audit’s quiet benefit: it separates what you were waiting on a developer for from what you were only assuming you were.

How often I run this

Quarterly for the full six checks, and immediately after anything structural: a redesign, a migration, a new section. Single-page spot checks whenever a page that should perform is not, in which case this audit runs alongside the content-side diagnosis I covered in E-E-A-T for AI content. If you do SEO for a living, the deeper weekly practice in my guide for SEO specialists subsumes this one.

FAQ

Can ChatGPT crawl my website and find all these issues itself? No, and be suspicious of any answer it gives about a URL it has not fetched. Some tools bolt browsing onto the model, but the reliable pattern is the one in this post: you collect the artifact, the model analyzes it. The artifact is your evidence that the finding is real.

Do I still need an SEO tool subscription? For a small site, the checks here plus free PageSpeed Insights and Search Console cover the technical fundamentals. Paid crawlers earn their fee on scale - thousands of URLs, scheduled recrawls, change tracking. I would start free, and buy the tool the month the manual version stops fitting in an afternoon.

Is any of this different for AI search engines? The artifacts are the same, with one addition: the robots.txt check should confirm you are not blocking AI crawlers you want citing you. The content side of that question - being quotable once they can crawl you - is its own discipline.

What is the first check to run if I only run one? Robots.txt, because it is the only one where a single line can be costing you every page in a section, and it takes the least time to read.

This audit is part of my AI SEO playbook. The schema half lives in the schema markup guide, the keyword side starts at ChatGPT keyword research, and the wider practice begins at the complete guide to AI marketing.