Why Claude first gave us the wrong SEO advice

In April 2026, we asked Claude, the Anthropic AI assistant we use to build our website, to improve our search rankings. Four days later the job was done, and done well: title and description tags, social sharing images, structured data for Google, a sitemap in French and English, a robots.txt file open to AI crawlers, an llms.txt file for ChatGPT and its cousins, marked-up FAQ, breadcrumbs, lighter images, accessibility.

Five months later, fuushia.com still didn’t show up on Google or in AI assistants’ answers.

The AI wasn’t wrong about what it had done. It was working on the wrong problem, and so were we. Here is what was missing, how we fixed it, and the checklist we wish we’d had in April.

What we asked for, what we got

In essence, the request fit in one sentence: “improve the site’s SEO”. An AI working inside a codebase does exactly what it is asked, with what it can see. It could see HTML files, tags and images. So it optimised HTML files, tags and images.

Everything it produced follows best practice. Nothing was wrong. But all of it was about the shape of a site nobody was seeing yet.

Google says so itself: structured data makes rich results possible but does not guarantee them. And its SEO Starter Guide is clear about priorities: creating content people find compelling and useful will likely influence your presence in search results more than any of its other suggestions. We had done the opposite: everything else, and no content.

What was really missing

At the end of September, we came back to it with a different question: “why isn’t this working?”. The audit surfaced gaps you can’t see in code.

No search engine data. We had visitor statistics, but no Google Search Console account, so no way of knowing whether Google knew our pages or which searches it showed us for. The sitemap existed but had never been submitted. No Bing Webmaster Tools account. No Google Business Profile. We had been optimising blind for five months.

A copy of the site on another domain. After we changed hosting provider in September, the technical domain provided by the platform served a full copy of the site, accessible and indexable. To a search engine, two identical sites are a duplicate-content problem.

A file for AIs that contradicted the site. Our llms.txt, written in April, had never been updated: it talked about “on-site” work when the diagnostic is done by video, and listed no prices even though our pricing is public. An AI that finds two different versions of the same company has good reason to distrust both.

Nothing saying who we are and where. Seven indexable pages per language, three of them utility pages. Slogan-style titles (“Tarifs — FUUSHIA”, i.e. “Pricing — FUUSHIA”, sixteen characters). No city: our area was “France”. And not a single page answering a question a small business owner would actually type.

Why the AI didn’t see it

An AI optimises what you show it. It doesn’t check what it can’t see.

Search Console, Bing Webmaster, the Google profile: none of these exist in a code repository. The duplicate only showed up by querying the live site, not by reading files. And the gap between llms.txt and the rest of the site widened over the months, without anyone asking the question again.

This isn’t a flaw specific to one model. It’s how we used it: a vague request, no real data, no checks against production. A human consultant given only the code would probably have done the same. A good consultant would have started by asking: “do you have access to Search Console?”

What we fixed, in two days

On 29 and 30 September:

• Measure first: Search Console set up (domain property), sitemap submitted, site imported into Bing Webmaster Tools, Google Business Profile created (service areas, no public address, verification pending).
• Remove the duplicate: the technical domain now sends a no-index directive.
• One version of the facts: llms.txt rewritten from the site, with our actual prices (excluding VAT) and the video diagnostic.
• Say where we are: one main heading per page carrying the topic and the area (Bouches-du-Rhône, Paris and Île-de-France), the same areas declared everywhere.
• Write for real questions: the blog went live. This is its fourth article.

The results, unvarnished

What Search Console shows from 29 September to 10 October: 50 impressions, 1 click, average position 13. The search that shows us most often is “test chatgpt entreprise”, which matches our first article, Does your business exist for ChatGPT?.

That’s not much, and it’s too early to draw conclusions. But it is an observed figure, from a real source, where in April we had nothing. To use the framework from our article on the ROI of an AI project: the effect of April’s technical work remains unknown, and will stay that way, for lack of a baseline.

The checklist before touching a single tag

If you want to work on your search visibility, with or without AI, this is the order we would follow today. The first steps need no technical skills.

1. Check that Google knows you: type site:your-domain.com into Google. If nothing comes up, everything else can wait.
2. Open Search Console and submit your sitemap. It’s free, and it’s the only reliable source on what Google sees of you.
3. Import the site into Bing Webmaster Tools: from a Search Console account, it takes a few minutes.
4. Create or complete your Google Business Profile if you serve a geographic area.
5. Look for copies of your site: old domain, your host’s technical domain, the http or non-www version. Only one address should answer; the others should redirect to it.
6. Make your information consistent: website, Google profile, LinkedIn, llms.txt if you have one. Same services, same areas, same prices.
7. Write one page per real customer question, with your city when it matters.
8. Only then, tags, structured data and performance.

Using AI well for SEO

We still work with Claude, including on this. What changed is how we talk to it:

• Give it real data: a Search Console export beats a thousand code files.
• Ask it what it can’t see: “what could be blocking our visibility that you can’t check from the code?” That one question would have surfaced Search Console back in April.
• Have it check production, not files: what a search engine sees is the live site.
• Have it list its assumptions before acting, and read them.

What stays human

Choosing your areas, deciding what you sell, answering your customers’ real questions from your own experience: no AI will do that for you. Above all, asking whether you’re asking the right question. In April, we had a good answer to the wrong question.

Where to start this week

Run the site: test on your domain, then open Search Console if you haven’t already. In a month, you’ll have your first observed figure. To find out what AIs say about you, follow our 15-minute test.

And if you’d like us to look at what’s missing together, using your data rather than guesswork, that’s what the free 30-minute diagnostic is for. Our other engagements are on the services page.