Two things are true at once: 87% of SEO practitioners now use AI regularly or as a core part of how they deliver work, and the single most instructive published failure of the past month was an AI model given write access to a website. The gap between those facts is not a reason to stop using the tools. It is an argument about where the human sits in the workflow, and the answer that keeps holding up is that the human belongs between the model and anything that goes live.
What the adoption numbers actually say
Two surveys landed in the same week and they line up.
Keyword.com’s State of AI in SEO 2026 survey, with 97 usable responses skewed toward lean teams and service providers, found 87% using AI regularly or as a core part of delivery. Claude led tool usage at 78%, ahead of ChatGPT at 57%.
Semrush’s survey of how marketers use AI for SEO shows where that usage goes, and this is the more interesting half:
- 60% use it for keyword research
- 48% for brainstorming content ideas
- 38% for content briefs
- Only 18% for planning topic clusters
- 15% for finding internal linking opportunities
- Just 11% for SERP or content gap analysis
Nearly everyone has pointed AI at the commodity tasks, which is to say the tasks their competitors have also automated. As the Search Engine Land analysis of that data puts it, that distribution “produces more content but no advantage.” The work that is harder to scale, and therefore worth more, is the work sitting at 11% and 15%.
The failure worth studying
The clearest cautionary tale of the month was published by someone who ran it on his own site. Writing at Search Engine Land, he described asking Claude to review Google Search Console, recommend target keywords for a tool his company runs, and build the pages those keywords needed.
What Claude built was the homepage, three times. It cloned the homepage into two new URLs, changed the title tag and H1 to match the new keywords, and reused most of the body copy. On paper each page targeted a new term. In practice it was one page of content living at three addresses, competing with itself.
He left the two clones live deliberately, as a running experiment. Six months of Search Console data show the result: both pages have earned zero impressions and zero clicks. Every query they were built to win still goes to the homepage instead, which sits at average position 9.2 for “content grader,” 10.6 for “seo grader” and 5.3 for “ai content grader.” Bottom of page one, for terms a dedicated page should own outright.
The detail that makes it worth writing about is that it happened twice, on unrelated projects, months apart, with no shared prompt or workflow. His son ran the same kind of request for a different site and got the same behaviour: a batch of new pages, each a copy of the homepage with a changed title tag.
His framing of the underlying problem is the part to keep:
Claude is genuinely useful for SEO research, analysis, and drafting, but it can be quietly wrong when it’s left to execute on its own.
Research needs a model that can hold a lot of context and generate plausible options quickly. Execution needs something that knows when a plausible option is wrong for this specific page, this site architecture and this keyword map. Those are different jobs, and current models are reliably good at the first and unreliable at the second. The failure mode is not stupidity. It is confident production of something that looks finished.
Several practitioners quoted in the same piece describe variations of it. SEO consultant Robert May:
The problem’s not AI, it’s using AI to create hundreds of pages that should never have existed in the first place.
Scott DeSapio, on the structural version:
One page trying to rank for five different searches usually ranks for none. Each URL should serve one clear search intent.
And a crawlability version, measured by a developer posting as @rentierdigital:
[C]laudebot downloads your [JS] bundle in 24% of its requests and never executes it. it cannot read the thing it helped you build.
That last one deserves a moment. An agent that builds a JavaScript-heavy page dependent on client-side rendering can produce something that looks complete in a browser and is difficult for the crawlers, including its own, to read. That is the retrievability layer failing, and it is the layer most sites think they have finished.
Does Google penalise AI-written content?
It does not, and the evidence on this has become fairly settled. Google’s position is that using AI to produce content is not against its guidelines provided the content is helpful and made for people; its systems reward quality regardless of how a page was produced, and demote content built to game rankings.
The measurement backs that up. Ryan Law’s team at Ahrefs studied 331,000 pages and found that 5.3% of pages ranking in positions one to three are fully AI-generated, with no evidence of a filter blocking AI content from the index. Moz cited that study in a piece arguing that the detection debate is the wrong debate:
I think we’re focusing on the wrong thing by obsessing over AI detection. The bigger issue is content quality and how we build ownership as discovery becomes more fragmented.
The same article separates the failures people conflate. Sports Illustrated publishing reviews under fake author profiles was an authenticity problem. The Chicago Sun-Times publishing a reading list containing books that did not exist was an accuracy problem. AI was involved in both, and neither was caused by AI. Both were caused by content reaching publication without passing an editor.
Its sharpest line is aimed at a workflow a lot of teams are currently building:
Our agentic workflow scores content quality, and anything below the mark is sent back to the loop until it passes (congratulations, you’ve automated mediocrity)
A closed loop of models grading each other converges on whatever the grader rewards. It does not converge on something worth reading.
Where does watermarking fit?
On 11 August 2026, Anthropic began adding machine-readable watermarks to Claude’s outputs. Search Engine Land’s analysis of the reaction is worth reading for the context most of the commentary skipped.
The change is a compliance response to Article 50(2) of the EU AI Act, which requires providers of systems generating synthetic text, images, audio or video to mark those outputs in a machine-readable format. Anthropic, OpenAI, Google, Meta, Microsoft, Mistral and Cohere have all signed the EU’s Voluntary Code of Practice on Transparency of AI-Generated Content. xAI did not.
The method is statistical rather than orthographic. Older text watermarking inserted hidden characters or zero-width spaces, which alter the form of the text and are straightforward to strip once you know to look. Statistical watermarking instead biases the model’s sampling with a secret key, so the output reads naturally while remaining detectable by the provider. Anthropic states it inserts no hidden characters, does not identify individual users, and has no practical effect on output quality.
What this means for content teams, honestly: not much yet, and it is too early to know how useful it will be for detection. It is not a filter Google applies, and it does not make AI-assisted content a liability. It is worth knowing about because it will be cited at you in a meeting.
The measurement wrinkle nobody expected
One finding from this fortnight has practical consequences for anyone tracking AI visibility. Profound tested 1,724 prompts across both Claude and Claude Code between 13 and 23 July, generating 24,135 responses with web search enabled on both. Search Engine Journal covered the results.
Claude used web search in 93% of responses. Claude Code used it in 13%. The brands mentioned for the same prompt overlapped by roughly 20% on average, which is to say two products running the same underlying model recommended largely different things.
Their crawling patterns diverge too. Around three-quarters of Claude Code’s observed page visits went to documentation, informational and pricing pages, against 5% for Claude. Conversely, 60% of Claude’s visits were to robots.txt files, sitemaps and homepages, against 4% for Claude Code. One agent surveys what a site contains; the other goes straight for specifics on page types it already expects.
The caveat matters: this is one vendor’s data, from a company that sells AI visibility tracking, with page types labelled by a model and no stated human review. Treat the direction as more reliable than the decimals. But the implication is sound, and it is that “are we visible in AI” is not one question. Products sharing a model are not interchangeable measurement surfaces.
A workflow that holds up
The pattern that survives all of the above is gates. Run content through discrete stages — idea, keyword research, brief, draft, fact and quality check, an editing pass — and let nothing reach publish until it clears each one. The failure mode of AI content is the firehose: hundreds of pages, no gates, all of it average.
The gate that matters most sits before drafting, and it is one question: does this page add something the top ten results do not already have? Google holds a patent on measuring information gain, the new information a page contributes beyond what is already indexed. If the answer is “nothing new,” the correct output is not a better draft. It is a decision not to write the page, or a decision to go and get the data that would make it worth writing.
Where the budget goes is the other half. Writing at Search Engine Journal, a consultant who builds AI search programmes describes cutting the publishing calendar in half and moving those hours into work a model can quote: original data, named outcomes with numbers attached, expert commentary from people inside the company. Eight generic posts a month lose to one piece carrying a number nobody else has. On budget, the advice is deliberately conservative: move 15% to 20% in the first quarter and let the evidence move the rest, keeping paid search largely intact because it remains the cleanest read on which queries carry buying intent. The line that moves is usually digital PR, and our own read on what earns placements now is that the work has become harder to fake and easier to measure.
That is a position we recognise, and it is roughly where our own practice has landed. We use these tools daily for research, clustering, analysis and first drafts, and we build our own tooling around them when the off-the-shelf version does not fit, as with wiring a knowledge base into Claude over MCP. We do not let them decide what gets published, and we do not let them create pages. That is the line we hold on every account we run.
Frequently asked questions
Will Google penalise my site for AI-written content?
No. Google’s guidance is that AI-produced content is acceptable when it is helpful and made for people. An Ahrefs study of 331,000 pages found 5.3% of results in positions one to three are fully AI-generated, with no sign of a filter blocking them.
What is the most common way AI-assisted SEO goes wrong?
Giving a model authority to create pages. The documented failure is duplication: cloning an existing page, changing the title, and producing a URL that competes with the original. In the case above, both clones earned zero impressions and zero clicks over six months.
Should I track Claude and Claude Code separately?
If your customers use both, yes. In one vendor’s testing, Claude used web search in 93% of responses against Claude Code’s 13%, and the brands mentioned overlapped by only about 20%.
Does Claude’s watermarking affect my content’s search performance?
There is no evidence that it does. It is a compliance measure under Article 50(2) of the EU AI Act, detectable by the provider rather than by search engines.
Where is AI actually underused in SEO?
In the analytical work. Only 18% of marketers use it for topic cluster planning, 15% for internal linking opportunities and 11% for SERP or content gap analysis, against 60% for keyword research.
The short version
Use the tools for the thinking and keep a person responsible for what ships. The published failures of the past month were not caused by models being bad at language. They were caused by models being given the authority to decide something, and then optimising for output that looked finished rather than output that was right.
