Automating SEO with Claude, Content Refreshes & Google's Future
The one with Dragan Berak
What we talked about
- Working in-house and consulting at the same time
- Rising cost of AI models and token usage
- Claude artifacts, skills and projects for SEO
- What not to automate: hallucinations and human in the loop
- MCP connections to Ahrefs, GSC, GA4, Screaming Frog
- The illusion of completeness in LLM output
- Challenges for junior SEOs
- Where Google Search is heading: AI Overviews, AI Mode, indexing
- LLMs.txt and AI 'about us' entity pages
- Content refreshes for evergreen vs commercial content
- Verifying AI output
- How to start learning about LLMs
- Auditing for SEO vs AI search
The questions we dug into
Q1Why combine an in-house head of SEO role with consulting instead of an agency?
In-house lets you go deep into many layers of one business and use enterprise tools (you still 'sell' to product and dev teams for resources). Consulting exposes you to different websites, teams and problems — staying in one vertical limits professional growth. He's curious and likes variety.
Asked by Nik · answered by Dragan Berak
Q2What's new with AI and AI search from your perspective?
AI is getting more expensive: new Claude models burn tokens and he hits limits more often, so he's careful about which tasks need reasoning models. The economics look shaky (he cites OpenAI making ~$4B while spending ~$9B).
Asked by Nik · answered by Dragan Berak
Q3How do you automate SEO workflows like audits with Claude?
Learn artifacts (outputs) and skills (repeatable workflows): once satisfied with an output, ask Claude to turn it into a skill, refine it, share it with your team and apply it to other domains. Use separate projects with context files (brand guidelines, instructions) for standardized outputs.
Asked by Sara · answered by Dragan Berak
Q4What shouldn't you fully automate with AI?
Dragan: LLMs are advanced text predictors without a search index; they must always give an answer and will hallucinate, even when connected to data via MCP. Keep a human in the loop and don't use LLMs for keyword research; they're best at spotting patterns in structured tables. Sara: you can't verify what the model left out, creating a false impression of completeness.
Asked by Sara · answered by Dragan Berak & Sara
Q5What specific Claude setups and automations do you use?
He uses projects per workstream (on-page, content strategy), client-branded report skills, competitor backlink gap analysis via the Ahrefs connection, content briefs (~90% there, then human editing), technical analysis using Screaming Frog crawl data via MCP, and daily/weekly digests from GSC, GA4, Ahrefs and log files (up to ~10 GB/day of access logs to see where Googlebot spends time).
Asked by Nik · answered by Dragan Berak
Q6Where is Google Search heading, with rising zero-click searches?
Google is still figuring it out and must monetize. With huge search revenue, it won't switch to AI Mode easily; AI Overviews are here to stay. At Google's invite-only Zurich event, the focus was technical optimization and saving resources — he expects more indexing issues and is seeing many pages drop out of the index.
Asked by Sara · answered by Dragan Berak
Q7What do you think about llms.txt, schema and similar technical suggestions?
Google's messaging has been contradictory (dismissing llms.txt, then mentioning it for agents at Google I/O). He believes it will matter for how businesses appear online, and is revisiting his skepticism toward AI-facing 'about us' / entity-map pages after research from Steve Toth.
Asked by Nik · answered by Dragan Berak
Q8How should you approach content refreshes when LLMs seem to favor fresh content?
It depends on content type. Evergreen informational content that LLMs can summarize isn't worth fighting for. For commercial money pages, add layers of information — new angles, clearer explanations, tables, and especially proprietary data — which has increased impressions and LLM citations. 'Escape competition through authenticity' (Naval).
Asked by Nik · answered by Dragan Berak
Q9How do you double-check AI output for mistakes and hallucinations?
Push back constantly ('Are you sure? Where did you find this?') and verify data, especially numbers — even MCP-connected GSC data can mismatch.
Asked by Sara · answered by Dragan Berak
Q10Where should someone start learning to optimize for LLMs?
Don't try to be advanced from day one. Watch Andrej Karpathy's ~3.5-hour video on how LLMs work to understand they're statistical models, not omnipotent beings, then use them cautiously.
Asked by Sara · answered by Dragan Berak
Q11Do you separate SEO and AI search in audits?
No — good AEO/GEO is good SEO. If a site isn't crawlable, it won't be indexed. Titles and meta descriptions matter even more. First learn the business model and ICP problems; audits should be tailored like a doctor's diagnosis before prescribing.
Asked by Nik · answered by Dragan Berak
Lines that stuck with us
- Dragan: 'The hottest programming language right now is English' (quoting Andrej Karpathy).
- Dragan: 'Behave as if every single pixel on the site is a ranking factor.'
- Dragan: 'Good AEO/GEO is really good SEO.'
- Sara: LLMs create 'the impression of completeness' — you can't see the opportunities they missed.
- Dragan: 'Escape competition through authenticity' (Naval).
Note: This page was AI-generated from the podcast transcript and manually reviewed for accuracy and context. While we’ve made every effort to ensure it reflects the original conversation, mistakes or omissions are still possible. If you spot something that doesn’t look right, please reach out and let us know.