Claude is fast. Give it a crawl export and it will summarise a site’s technical problems in minutes, a job that once took most of a day. That speed is the appeal, and it is also the risk. A confident recommendation that is wrong reads exactly like one that is right. If nobody checks it, the mistake ends up on your website.
Bryson Meunier, an SEO with more than 20 years in the industry, recently described several occasions on Search Engine Land where Claude went wrong on real projects. The patterns apply to any AI tool. Below are seven of them, each with a quick way to check.
1. It reports problems that are not there
Crawl data is easy to misread. In Meunier’s audit, Claude described a plain HTML site as a JavaScript single-page application. It also told a business to claim a Google Business Profile that had already been claimed.
Acting on either one means paying for a fix that changes nothing.
Check it: Open the page and view the source (Ctrl+U on most browsers). That shows what Google receives before any JavaScript runs. For a business listing, look at the live profile before recommending anything.
2. It states things it cannot know
Claude told every site owner that no sitemap had been submitted to Google Search Console. It had no Search Console access, so it had no way to know.
When an AI tool has a gap in its information, it can fill it with something plausible. If the tool cannot see the source of a claim, treat the claim as a guess.
Check it: Ask what evidence the statement rests on. You can confirm a sitemap exists by visiting /sitemap.xml and reading the robots.txt file. Only Search Console itself shows whether it was submitted.
3. It writes titles that are correct and forgettable
Title tags are what people see in search results, so a dull one costs clicks. When asked to write titles from high-volume keywords, Claude produced repeated patterns such as “Service | Area | Brand”. In one case it replaced a distinctive, keyword-rich title with a flat one.
Check it: Read each title aloud. Ask whether it differs from every other title on the site, whether it uses a phrase people actually search, and whether it still sounds like the business.
4. It misses duplicate URLs, then breaks canonicals
Meunier found over a hundred duplicate titles caused by one page loading at many different addresses. Claude had the crawl data and did not flag the cause until it was pointed out.
In another project it went the other way and set the canonical tag on every demo page to one page. A canonical tag tells Google which version of a page to index, so this told Google that only one page mattered. It repeated the mistake after being corrected.
Check it: Sort the crawl by duplicate titles and read the actual URLs. After any change, view the source and confirm each page’s canonical tag points to itself, unless there is a clear reason it should not.
5. It adds noindex without being asked
This is the most dangerous item on the list. While fixing title tags, Claude also added noindex to a set of demo pages and justified it with crawl budget, on a site of roughly 20 pages. Months later it noindexed an entire web app, even though the project instructions said not to.
A noindex tag tells Google to drop a page from its results. Crawl budget is a concern for very large sites and rarely for a small business site.
Check it: After every AI-assisted code change, compare the site before and after. Search the source for “noindex” and test key pages with the URL Inspection tool in Search Console. Do this on every release, not just the first.
6. It rejects good ideas on false reasoning
Earlier in the same project, Claude and Meunier had agreed not to build a page for a term because the homepage already ranked for it. When a new idea came along, Claude treated it as the same case and suggested dropping it. The new pages were aimed at different searches, and the homepage did not rank for them.
Check it: When Claude says to drop an idea, ask which search term is involved and which page ranks for it today. Then search for that term yourself.
7. It makes link building slower, not faster
Claude sorts and scores potential link targets well. It also invented some contacts, deleted rows from spreadsheets without being told, and suggested sites that turned out to be a poor fit. Its outreach drafts were often too generic to send.
Check it: Visit every prospect and confirm the site, the contact and the topic match. Keep backups of your master spreadsheet. Write the outreach yourself.
Ask why before you approve
If you take one habit from this, make it a simple question: “Why are you recommending this, and what did you see that shows it?”
Claude tends to agree it was wrong once it is challenged, but that only happens if someone challenges it. Use it for analysis and first drafts. Keep a person responsible for deciding what goes live.
I have worked in SEO for over 23 years, and most of the damage I see comes from changes that nobody reviewed, whether a person or a tool made them. If you would like your team to spot these problems themselves, my SEO and PPC training covers crawlability, indexing and on-page basics using your own website. If you would rather have someone look at the site for you, get in touch about an audit.
Graig Upton
Graig Upton is a UK SEO and Google Ads consultant with 23 years of hands-on digital marketing experience. Google certified in Ads, Analytics and Conversions, he has helped businesses ranging from local providers to national brands - including Nando's and Financial Times - dominate search and scale their leads. Based in Preston, working with clients across the UK.