Using AI To Assist With SEO Work, Not Replace The Worker

An SEO practitioner building a Chrome extension that combines traditional SEO checks, browser data, deterministic analysis and language models argues AI should assist rather than replace the SEO reviewer, after testing Google's 4-bit Gemini Nano on raw versus rendered DOM HTML differences. The on-device model handled wording changes and their possible consequences reasonably well but was much less reliable at combining several structured technical facts into a final SEO judgement, confusing inputs with outputs and inventing rationale. The author's conclusion is to reserve deterministic checks for binary facts such as 404s, canonicals, robots.txt access and server-versus-rendered DOM link changes, and use language models only for interpretation tasks like concise explanations, change summaries and Jira ticket-ready descriptions.

One of the easiest traps when adding AI or an agent to a process is for it to become: Here is some data. Model, please tell me what to think. I’ve been wrestling with this for a few months now. I’d say where I have ended up is trying to go in the opposite direction – at least in certain contexts. The Chrome extension I’ve been building combines traditional SEO checks, browser data, deterministic analysis, and language models. I am aware you can effectively do that with MCP servers and agents https://www.searchenginejournal.com/safaris-mcp-server/581487/ – it’s almost too easy to audit a site without ever auditing the site itself … or to think you are, at least. But the goal isn’t to delegate an SEO review to AI outputs; the goal is to make the review easier to perform by an SEO using AI. That distinction matters, for now at least – it’s been a constant point of tension in my own practice. This follows on from my previous article, Using Local AI Compute To Reduce Reliance On Frontier Models. https://www.searchenginejournal.com/using-local-ai-compute-to-reduce-reliance-on-frontier-models/ Deterministic Where Possible Some parts of technical SEO are simple, and will be forever simple – at least if you know what you’re looking for: - A URL returned 404, or it didn’t. - A canonical exists, or it doesn’t. - A link destination changed between the server HTML and rendered DOM https://www.searchenginejournal.com/client-side-vs-server-side/482574/ , or it didn’t. - Robots.txt permits a crawler to access a path, or it doesn’t. We do not need an LLM to discover these things; in fact, most LLMs are likely to write a deterministic check to then reliably capture this data themselves. For a process you are going to repeat time and again, a firm framework rather than a probabilistic token-eating word-predictor IS the better option. This gives the reviewer a much stronger starting point than simply pasting a page into a model and asking what might be wrong. AI Where Interpretation Helps There are still plenty of places where language models can make the workflow better https://www.searchenginejournal.com/how-to-use-ai-to-streamline-time-consuming-seo-tasks/566499/ . It would be highly hypocritical of me to be anti-AI altogether Once gathered, a bundle of technical information might be accurate but unpleasant to read, or create friction when processing and understanding it. The skill here isn’t in an SEO doing SEO things; it’s in managing the friction of a JSON read-out or a spreadsheet and turning that into something to gain insight from. A language model can turn it into: - A concise explanation. - A summary of what changed. - A Jira ticket-ready description. - Clearer wording for a consultant. - An interpretation of a genuinely ambiguous semantic change. Those are useful contributions and, crucially, they do not all require the model to become the final decision-maker. You, the consultant/SEO/CMS editor, are the decision-maker; your AI isn’t accountable. This became particularly obvious while testing a small on-device model Gemini Nano against raw/rendered DOM HTML differences. The model was reasonably capable of describing changes in wording and possible consequences. It could recognize that changing an anchor from something descriptive to “Learn more” removed useful context and that could be a problem. But it was much less reliable when asked to combine several structured technical facts into a final SEO judgement. It confused inputs with outputs, it tried to invent rationale, and it failed to fully understand how a combination of facts could lead to an outcome. There was too much context and the judgement itself had to be nuanced – and right now the 4-bit Gemini Nano that ships with Chrome doesn’t seem great for that. Rather than keep adding instructions until the prompt became a technical SEO textbook, I opted to instead change the duties I was giving it. For my sanity/hairline and to ensure that the tools do what they’re most capable of. The local model now helps convey the findings/evidence – it doesn’t get to judge them. Reduce Friction Rather Than Remove The Practitioner This is increasingly how I think useful AI tooling should work – at least today. IF the costs of AI really start to grind things to a halt or the bubble bursts , I think this way will become THE WAY forward. OR we’ll find a way to maintain this pace at all costs and this assistive method will become “ artisan SEO https://www.chris-green.net/post/birth-of-artisanal-seo .” Imagine finding a rendered-DOM difference which contains: - Two URLs. - Their HTTP status. - Canonical evidence https://www.searchenginejournal.com/google-updates-javascript-seo-docs-with-canonical-advice/563545/ . - Robots evidence. - Anchor changes. - Local page context. - Reconciliation confidence. - Transformation data. A practitioner can absolutely inspect that information manually, but doing so repeatedly is friction, tedious friction. If a model can turn it into: The link destination changes after rendering. The server HTML already exposes a working URL, both versions ultimately resolve to the same destination, and the anchor text is unchanged. That saves effort without taking the decision away from the reviewer, and a human can then decide whether it matters. For those cases when the evidence genuinely requires stronger reasoning, or there are many, many areas that need to be inspected at once, THEN a more capable AI model can tag in to speed up that process. AI Assistance Also Makes Disagreement Useful There is another benefit to keeping the human in the loop – perhaps an unexpected one with AI. When the model disagrees with you, you can inspect why. When feeding in facts under a specific remit, the sycophantic tendencies of AI chatbots https://www.searchenginejournal.com/llms-are-changing-search-and-breaking-it/560346/ are suppressed because the AI model’s remit is different – almost “purer.” During development, this was extremely useful as model failures were exposed when: - Evidence was too noisy. - Technical terminology was ambiguous. - Two concepts had been collapsed into one. - Deterministic logic was weak. - My own personal experience/knowledge was getting in the way. - There was not enough context to judge effectively. - The model simply wasn’t capable enough for the task. If the entire workflow had been delegated to the model, those weaknesses would have been much harder to see. The Goal Isn’t Autonomy There is an understandable desire to make AI tools increasingly autonomous https://www.searchenginejournal.com/ai-agents-will-game-your-seo-metrics-mit-stanford-research-points-to-the-risk/589948/ . I build processes, workflows, and train people on them – most tasks would be replaced by a capable agent – which creates new problems, trust me Autonomy of AI or an agent is not the only measure of usefulness. A tool that saves me 30 seconds twenty times a day, makes evidence easier to interpret, keeps repetitive work consistent, and gives me better raw material for decisions can be extremely valuable without ever making those decisions itself. That’s what we should aim for, as we still understand what is happening, why it is happening, and why it is a problem. These are crucial when making recommendations and getting them fixed. If you use a model to do everything and you’re just there to deliver the recommendations, what do you do if someone then asks you a question? You won’t understand it well enough. Use software to establish facts. Use AI to reduce the effort required to work with those facts. Use our judgement where judgement is actually required. The result is less spectacular than “AI does your SEO audit for you.” It’s not a 40-page audit which no one will read, let alone understand. It is, I’d argue, considerably more useful, keeps you close to the problem, and makes you a better practitioner who can’t be replaced by AI https://www.searchenginejournal.com/what-not-to-automate-with-ai-the-seo-deskilling-trap/574887/ . More Resources: - The Technical SEO Audit Needs A New Layer https://www.searchenginejournal.com/technical-seo-audit-new-layer/571583/ - Create Your Own ChatGPT Agent For On-Page SEO Audits https://www.searchenginejournal.com/create-your-own-chatgpt-agent-for-on-page-seo-audits/546016/ - Deploying Agentic AI For SEO: A Playbook For Technology Leaders https://www.searchenginejournal.com/deploying-agentic-ai-for-seo-a-playbook-for-technology-leaders/559800/ This post was originally published on Chris Green SEO https://chrisgreenseo.substack.com/p/using-ai-to-assist-with-seo-work . Features Image: Taris Tonsa/Shutterstock