An SEO asked their AI assistant to review Search Console, pick target keywords, and build new pages. What came back looked promising: new titles, new H1s, new meta descriptions, new URLs. Except the body copy on every page was cloned from the homepage.
This is not a failure of AI. It is a failure of delegation.
Episode 142 of SEOTalk Spaces was about drawing that line clearly, which SEO tasks AI should handle, which it should support, and where handing over control entirely produces expensive mistakes.
Key takeaways
- AI handles repeatable SEO tasks well — data analysis, root cause diagnosis, content structuring. Strategic thinking still requires a human
- A prompt drawing only from training data produces a common denominator — unique content requires proprietary data fed into the system before the prompt fires
- An AI council with a built-in detractor agent produces more defensible content than any single-model, single-pass workflow
- Agencies are primarily using AI to protect margins, not improve outcomes — brands need to ask harder questions about what they are actually getting
- Every SEO professional now needs an AI-native approach — not to replace judgment, but to make it faster and better-informed
Where AI Actually Saves Time in an SEO Workflow
Our host Parth opened with something practitioners rarely say plainly: AI has saved him 60 to 70 percent of his time on repeatable weekly and monthly tasks. Root cause analysis across GSC and GA4, tracking which technical fixes moved the needle, synthesising data across multiple dashboards that previously required manual review.
The key word is repeatable. These are tasks with a defined process, clear inputs, and measurable outputs. AI executes them faster and with fewer human errors than a team member switching between tabs.
What the freed-up time enables is more significant than the automation itself. Parth described using that headroom to think upstream – giving product recommendations to a SaaS client that resulted in a 35 percent revenue increase. That kind of cross-functional thinking is not something you hand to a model. It requires understanding the business, the customer, and the gap between what the data shows and what the team has not yet noticed.
What to do
List every SEO task you do weekly. Mark each one as repeatable (same inputs, same process, same output) or strategic (requires judgment, context, or cross-functional knowledge). Automate the first group. Protect the second.
Why a Good Prompt Is Not Enough
One of the question raised during the conversation that we generally tend to avoid was: if AI has been trained on everything, why can’t you just ask it a question and get a great answer, every time?
Parth’s answer was specific. Every LLM draws from the same central training data. Two practitioners with similar prompts will get similar outputs because the source is the same. The result is a common denominator: content that restates what already exists in aggregate rather than adding anything new.
The fix is not a better prompt. It is a better system. When AI is given access to proprietary inputs before the prompt fires – your brand’s sales data, support queries, internal search data, customer interview transcripts – the output is grounded in context no other brand has. The prompt becomes a direction, not a question asked into the void.
Parth also described building content section by section rather than generating an entire piece in one go. The context window cannot hold a full brief, research, and output at once without quality degrading. The system handles the process. The human designs the system.
Watch out
A prompt drawing only from AI training data produces a common denominator output. If your content workflow does not feed proprietary data into the system before generating, you are producing a slightly different version of what hundreds of other sites have already published on the same topic.
The AI Council Idea
One interesting idea discussed during the podcast, can we build a panel of AI agents with different roles, including at least one detractor — an agent whose job is to challenge the output, find the gaps, and push back on weak reasoning. In any good editorial team, the person who asks “but why does this matter?” is among the most valuable in the room. That function can be built into an AI workflow.
The practical version: run your content brief, outline, or draft through a panel of agents – one evaluating topical completeness, one checking brand voice, one stress-testing the argument. Content that survives that process is more defensible than anything produced by a single model in a single pass.
The caveat Parth raised is worth noting: at scale, this creates a token cost problem. An AI council that challenges every section of a 2,000-page product catalogue will burn through budget quickly. The approach works best for high-value pages where the quality argument matters most – money pages, pillar content, comparison pages.
Is Content Uniqueness Still Possible at Scale?
One concern across the industry, that got spoken about was – if everyone uses the same models with similar prompts, content across the internet risks converging toward identical outputs.
Parth’s counter was grounded in how the system is built, not how the prompt is written. Brand voice guidelines, negative vocabulary lists, proprietary data feeds, internal search data, sales team insights – all of these create structural differentiation that persists across every piece the system produces. A single writer producing content for two competing brands could still produce unique work before AI. The same principle applies now: the uniqueness comes from the inputs, not the model.
What makes textual content genuinely unique in the current environment, Parth argued, is proprietary data: original research, real numbers, studies backed by millions of data points. That is the layer that neither competing brands nor AI systems can replicate. A brand that publishes original research owns a citation source. Everything else is a variation of something already on the web.
Is SEO Becoming an AI-Native Role?
Finally, we spoke about the SEO role in itself. A comparison that has circulated in marketing for years: the CMO role was supposed to evolve into a chief marketing technologist as technology became central to the function. Is something similar happening to SEO?
Parth’s position: yes, and it is already underway. When onboarding clients, he now collects go-to-market strategy documents, performance marketing numbers, and sales team data as standard. The insights from connecting those dots inform SEO decisions in ways a purely keyword-focused approach would miss.
Deepali added an important note from her experience: over-reliance on AI creates a different kind of risk. The astrophysicist who generates code with Fortran but cannot explain what the code does is a useful warning for any professional. If you can no longer do the work without the tool, you have lost something more than efficiency.
The practical conclusion from the panel: SEO is not becoming a purely technical or AI-native role. It is becoming a role where AI literacy is a baseline expectation, the same way spreadsheet literacy was a decade ago. The practitioners who understand enough about AI workflows to design them, evaluate their outputs, and connect them to business outcomes will operate at a different level from those using AI as a glorified autocomplete.
Join the conversation
Where have you drawn the line between AI-assisted SEO and AI-executed SEO in your own workflow? Is there a task you handed to AI that you have since taken back and why?
Frequently Asked QuestionsWhich SEO tasks are safe to fully automate with AI?
Repeatable tasks with defined inputs and measurable outputs are the safest starting point: pulling and analysing GSC and GA4 data, tracking technical fix performance, generating content outlines from a pre-built brief, and running internal link audits. These follow consistent processes where AI reduces time without requiring judgment calls about strategy or business context.
Why does AI-generated SEO content often look the same across different brands?
When a prompt draws only from an LLM’s training data, the output is grounded in what already exists across the web in aggregate. Without proprietary inputs fed into the system before the prompt fires, the output is a variation of content that hundreds of other sites have already published on the same topic. The model is the same. The training data is the same. So the output converges.
How do you make AI content genuinely unique at scale?
Feed proprietary data into the workflow before the prompt fires. This includes customer support queries, internal search data, sales team insights, brand voice guidelines, negative vocabulary lists, and original research. Structure the brief section by section rather than generating everything in one pass. Content built on inputs competitors cannot access will be structurally different from what any shared training data can produce.
Are agencies improving content quality with AI or just protecting margins?
Based on the panel’s direct experience, protecting margins is the more common initial motivation. That said, the efficiency gains are real: faster turnaround, fewer feedback loops, more consistent first drafts. Whether that efficiency translates into better outcomes for the client depends entirely on whether the workflow was designed with quality inputs and editorial review, or just speed.
What is an AI council and how does it improve content quality?
An AI council is a panel of agents with different assigned roles — including at least one detractor agent whose job is to challenge the output and find weaknesses. Rather than generating content through a single model in a single pass, the content flows through multiple agents that evaluate it from different angles. The result is more thoroughly tested before it reaches editorial review. It works best for high-value pages rather than bulk content production, because of the token cost involved.
Does every SEO professional now need to understand AI?
Yes, but not at an engineering level. The practical requirement is understanding how AI workflows are designed, when they are producing weak outputs, and how to connect AI-generated insights to business decisions beyond traffic and rankings. Practitioners who understand enough to design a system, evaluate its outputs, and course-correct it are operating at a different level from those treating AI as an advanced autocomplete tool.