Why Torumata instead of a do-it-yourself language model

It's easy to say that Torumata is basically the same as pointing ChatGPT at your own website. It isn't. Not because the model is stupid, but because it doesn't see the whole site, can't measure, and its result can be neither repeated nor backed up. Torumata fetches the site like a crawler, measures 31 things deterministically and uses a model only where it makes sense: to assess content, with safeguards against made-up claims.

That holds for a company website or online shop of 20 to a few hundred pages, where you want to know what AI search engines and assistants actually get from it, compare the result a month later and hand it to an agency. For a single page, or text you're still writing, a do-it-yourself model remains the right choice.

What analysing a site for AI means

Between a website and an assistant's answer stand three gates, and only behind them the content. The order of the layers is the order of importance: a page the crawler can't get has no content, however well it's written.

Google states in its own documentation that a page must be indexed and eligible for regular Search to appear in AI Overviews or AI Mode (Google Search Central). That is the only necessary condition anyone has documented. For content there is a single peer-reviewed experiment, Princeton, KDD 2024, 10,000 queries (arXiv 2311.09735): quotations raised visibility in answers by 41 %, specific statistics by 33 %, cited sources by 28 %. That is a documented effect. Structured data, an llms.txt file and domain authority have no documented effect: you can do them, but not promise anything.

When a language model analyses the site on its own

A model in a chat gets only what its page-reading tool hands it, plus its own impression. Six reasons why that doesn't add up to a site analysis:

1. It sees a handful of pages, not the site. The tool fetches pages on request, not the whole site through links and the sitemap. On a 120-page site the model usually reads the homepage and a few subpages and guesses the rest.
2. It doesn't render JavaScript. Menus, products and prices built by JavaScript aren't in the HTML the server sends. The model then reports "no content", or content it made up from the page name. AI crawlers don't render JavaScript either (except Gemini and Applebot), so this difference needs measuring, not guessing.
3. One lock on a gate and the model gets it wrong. In our October 2026 measurement the claude.ai tool reported for 19 of 20 sites without HTTPS that the site prohibits automated access. Those sites prohibited nothing; they just had no encrypted connection (article with data). A hosting firewall ends the same way.
4. It can't measure. Text contrast, which pages return 404, whether canonical matches the sitemap, how many pages share a title. The model only estimates these numbers, as confidently as if it had measured them.
5. It doesn't separate what it knows from what it guessed. GPTBot, OAI-SearchBot and ChatGPT-User are three crawlers from one company, each with its own robots.txt rules. It knows the state from training and won't admit it never read the file. It describes the benefit of structured data or llms.txt the way marketing does, not the way studies measured it.
6. The result can't be repeated or handed over. Two identical prompts give two different lists, with no record to compare against next month.

This isn't a criticism of the model: Torumata uses the same kind of model to assess content. The difference is what it's given and what happens to its answer next.

How Torumata does it

1. A crawler fetches the site through the sitemap and links, every page without and with a browser.
2. 31 deterministic checks: robots.txt rule by rule for every AI crawler, 404s, canonical, duplicates, contrast and more, each finding with its measured value. The checks manual describes what is measured.
3. A language model only for content, over a sample of pages, with measured facts as input, a cap of $3 per audit and a provider in the EU region.
4. A report and files to deploy: a findings table, CSV, PDF, llms.txt, JSON-LD and a robots.txt patch generated from what was measured on the site.

Safeguards: every step saves its results as it goes, so if the model or the browser fails, earlier measured findings remain. The model never receives the raw site as instructions: content is wrapped as untrusted input and every answer is validated against a schema.

Whether findings hold is measured too. In September 2026 we checked the findings of production audits of three domains by hand against the live site. Over three rounds of fixes the share of false findings fell from 10.3 % (55 of 533) to 0.17 % (1 of 581). None of the defects found was a model hallucination; all were in deterministic checks and could be fixed with a test that keeps them from coming back.

Point-by-point comparison

Where a do-it-yourself model is better

• One page, one text. Whether a paragraph answers a customer's question concretely, a model judges best with the text in front of it.
• Content still being written. A draft FAQ, a new page's structure, headline variants: nothing to measure, only to write.
• A five-page site without JavaScript. Going through it by hand takes as long and nothing important slips through.
• Explaining a term or a finding. Why canonical matters, a model explains better than a table.
• A second opinion on a finished report. Let a model read Torumata's output and ask what to do first.

The common thread: a model is good where it has the whole input in front of it and nobody needs to repeat or prove the result.

What Torumata can't do yet

The honesty we ask of findings applies to the product too. As of 8 October 2026; every point is on the fix plan:

• It detects a CDN in front of the site, but not whether the network lets AI crawlers through; the report tells you to check.
• The hosting firewall can't be measured from outside; the report states it as an open question, not as "fine".
• The model's content recommendations have no list of pages or role that fixes them; deterministic findings always do.
• Public application parts of a site can add meta description findings instead of one finding about the non-public area.
• Comparing two audits, a finding on a page that dropped out of the crawl counts as fixed; watch the missing pages.
• Page speed (LCP) is sampled and varies even for one page; the trend across audits counts.

Time and cost

For a 100-page site. Do-it-yourself time is an estimate from our own trials, not measured at customers; prices as of 8 October 2026.

The point isn't that Torumata is cheaper. The point is that do-it-yourself time is paid again with every repetition, and the results of two runs can't be compared, because the model got different input each time.

Conclusion and recommendation

A model and Torumata answer different questions: "what do you think of this text" versus "what is measurably wrong on this site, and who fixes it". For a whole site, a do-it-yourself model fails at the first step: it doesn't see the site the way crawlers do, and nobody gives it measured facts to assess.

Recommended approach: run a free Torumata audit (12 pages, no card), decide on a full audit based on the result, hand the findings to your contractor as a table, and repeat in a month. Use a chat model for what Torumata doesn't do: writing new content and explaining findings.