AI tools can offer convincing explanations without showing how they reached an answer — so SEOs need independent verification before acting on what they say. That is the core argument of a Search Engine Land piece by Greg Jarboe, published November 6, which brings together three perspectives on AI reasoning. Thore Graepel, a UCL machine-learning chair and former AlphaGo and DeepMind researcher, argues large language models do not reason. Arthur Holland Michel argues we put too much faith in AI's ability to say no — refusals are inconsistent. Grace Huckins asks whether the benefits of open models outweigh their risks. The piece lands in the same week that Anthropic, on September 29, released a report warning that GLM-5.3, an open-weight model from Chinese AI company Z.ai, could significantly increase the cyberattack capabilities available to malicious actors.
Graepel's argument is the most fundamental: what looks like reasoning in a language model is pattern completion, not reasoning. The model produces the next plausible sentence, not a chain of verified logic — and the difference matters when the output is a recommendation about your site. A model that sounds certain is not a model that is right, and the confidence of the prose is no evidence at all.
Michel's point is about safety: refusals are inconsistent, so an AI that declines a harmful request once may comply the next time, with slightly different wording. Huckins's question is about the ecosystem: open models bring benefits, but they also distribute capabilities — including, per Anthropic's September 29 warning about GLM-5.3, capabilities that could help malicious actors run cyberattacks. Each perspective points at the same gap between what AI appears to do and what it actually does.
For SEOs, the application is concrete. An AI audit that claims a page has a problem should be checked against the page itself — the live HTML, the Search Console data, the actual rankings. An AI-generated content plan should be checked against keyword data and the competitive landscape. An AI's explanation of why traffic fell should be checked against the analytics, not accepted because it sounds plausible.
The habit to build is the one good analysts already have: trust, but verify, with the emphasis on verify. That means keeping the primary sources — Search Console, analytics, the live SERP — as the ground truth, and treating AI output as a hypothesis to test rather than a conclusion to act on. The tools are powerful assistants and unreliable authorities, and the difference is the verification step.
The link-building version of this problem is the AI-generated outreach list, the AI-scored prospect, the AI-written pitch. Each can be wrong in ways that look right — a prospect that does not exist, a score based on invented data, a pitch that misstates the recipient's work. Verifying before sending is the difference between a campaign and an embarrassment.
More broadly, the piece is a useful corrective to the industry's enthusiasm. AI has made SEO work faster, but it has not made it more certain — and in some ways it has made it less certain, by generating confident-sounding output at a scale that outruns checking. The SEOs who thrive will be the ones who use AI to do more verification, not less.
The verification habit also protects against a subtler failure: the AI that is right for the wrong reason. A model can recommend a correct action — fix the page speed, add the structured data — on the basis of a completely invented diagnosis, and the recommendation's correctness does not validate the reasoning behind it. Acting on the recommendation is fine; believing the diagnosis is not, because the next recommendation from the same flawed reasoning may be wrong. The discipline is to separate the prescription from the explanation, verify each independently, and keep the explanations that survive contact with the data. Over time, that habit builds a mental model of which AI outputs to trust and which to check — and that model is the real skill, more durable than any single tool.
Sources: Search Engine Land, November 6, 2026.
Related reading on this site: OpenAI Turned On Automatic Ad Changes in ChatGPT Ads by Default, OpenAI Expands ChatGPT Ads: Visual Formats, Amazon DSP and Brand Safety Controls, facebook-answers-6-million-pages-scaled-content, election-day-news-clicks-8x-no-ai-overviews, geo-referral-data-misses-buying-decisions, and Early-November Google Volatility Hits 9.4 Out of 10 — With No Update Confirmed.
Linkbaits helps you create content that earns citations and backlinks. Start with the free tools and build your first linkable asset this week.
Start free trial