AI tools that don't crawl are guessing

There's a fast way to check how AI models talk about your brand, and it's tempting: open ChatGPT, paste in your URL, and ask "how's my SEO?" You'll get an answer in seconds, formatted nicely, with a handful of specific-sounding recommendations. It feels like an audit.

It isn't one. Here's why.

What actually happens when you ask an AI to look at your website?

When you paste a URL into a chat interface and ask for feedback, the model doesn't crawl your site the way a search engine does. Depending on the tool, it either fetches a single page, skims whatever a general web-search plugin happens to surface, or, worse, draws on stale training data that has no idea what your site looks like today. There's no sitemap discovery, no systematic pass through every page, no consistent starting point. Ask the same question tomorrow and you'll likely get a different answer, because the underlying "audit" was never a fixed, repeatable process to begin with. It was a best guess, generated fresh each time.

That's fine for a quick gut check. It's not a foundation for deciding what to fix on your site this week.

What does a real crawl actually involve?

A proper site audit starts the same way a search engine does: by finding out what pages actually exist. That means starting at the homepage and following internal links methodically, page after page, the way a crawler builds its picture of a site. It also means comparing what that crawl finds against the sitemap the site publishes, wherever it publishes it:

• A robots.txt directive
• sitemap.xml
• sitemap_index.xml
• wp-sitemap.xml, including index files that point at child sitemaps

Pages the sitemap promises but no link actually reaches get flagged as orphans, because a page nothing links to is a page a real crawler will never find. Not a sample. Not "the homepage and a couple of links that looked important." A systematic pass, so the picture of your site is complete before any judgement gets made about it.

This matters more than it sounds like it should. A page with a broken canonical tag, a duplicate meta description, or missing schema markup doesn't show up if nobody looked at it. An AI model asked to "review your website" from a skim of three pages will simply never mention the problem on page forty, because it never saw page forty. A systematic crawl doesn't have that blind spot.

Then, and only then, the AI testing

Once there's a real, complete inventory of a site, testing how AI models talk about the brand becomes a much more precise exercise. Instead of asking a model to freeform-review a URL, the right approach is to run the actual queries a buyer would type, such as "best project management tool for small agencies", against ChatGPT, Claude, Gemini and Perplexity, and record exactly what each one said: whether the brand was mentioned, whether the site was cited, and who got named instead.

Pairing that record with a real inventory of your own site is what turns a guess into a plan. If a competitor gets recommended instead of you, you know exactly which query and which model, and you know from the crawl whether you even have a page that answers it. That's an answerable, fixable question, not an AI's vague impression.

Why this is the harder way to build it, and the only honest one

It would have been much faster to skip the crawl step entirely and just prompt an AI model to summarise a site on demand. It's also close to worthless as an audit: the same shallow skim every tool that takes the shortcut ends up doing, dressed up with a confident tone.

The alternative is slower to build and slower to run, because a full crawl of a real site takes real time. But it's the difference between a recommendation that's a hit-or-miss guess and one grounded in what's actually on your pages: a fixed methodology, the same queries and the same scoring every run, so when the number moves next week you can trust it reflects a real change rather than a differently worded question.

If a tool can't tell you how it arrived at a recommendation, ask it. Pasting a URL into a chat window will always get you an answer. It just won't tell you whether anyone actually looked at page forty first.

A few things worth knowing

Does "systematic" mean someone manually checks every page? No. The crawl and its findings are fully automated and deterministic: no AI involved in the scoring, so the same site produces the same findings on every run. The judgement calls come later, when you decide which flagged issue to fix first.

What happens to a page my sitemap lists but nothing on the site links to? It gets flagged as an orphan. A search engine, and an AI model doing the same job, will never find a page nothing points to, however good the page itself is, so a systematic crawl treats that as a fixable finding rather than letting it disappear silently.

Can I trust a week-over-week score if the AI's answers keep changing? Only if the queries and the scoring method stay fixed, which is the whole point of running this as a repeatable process rather than a one-off prompt. A model's answer can genuinely shift between runs; what shouldn't shift is what you asked it and how the answer got scored, so a real change in the number reflects the site, not a reworded question.