When to use MCP (and when not) – with Gus Pelogia (Indeed) | Ep.269
Live from brightonSEO: Björn talks with Gus Pelogia, Product Manager at Indeed and advocate for product-led SEO, about his talk “When to use MCP and when not to use MCP”. Gus explains why the real value of MCP comes from connecting several tools in a single flow: keyword research, People Also Ask data, prompt tracking and leadership reporting, all from one Claude conversation without exporting files. He walks through a practical example: clustering 4,000 headphone keywords into brand, feature and price buckets in minutes, then turning them into prompts for AI visibility tracking. The episode also covers where MCP falls short. Gus explains why 15 parallel SERP checks are better handled by a script written with Claude Code, why MCP results can differ from API data, and how he checks whether outputs are accurate. Other topics include agents replacing website visits, the “death of the web” debate, and why you should never trust the numbers blindly. Key Takeaways MCP is strongest when it connects multiple tools in one workflow. Research, analysis, charts and reporting stay in one place. Setup is easy: add the connector, log in to the tool, and start asking. You don’t need prompt engineering, just talk to it like an assistant. Main KPI is time saved. Research that used to take a day now takes about 30 minutes. Don’t use MCP for repetitive, large-scale jobs like live results for 15 cities. Have Claude Code write a script instead. MCP does not always match API data, and token costs can add up quickly with complex tasks. Daily monitoring (robots.txt changes, page-type update checks) still runs on scripts and Google Sheets, not MCP. To build trust in results, ask the model to verify its own data and explain how it checked. Agents will reduce website visits, but websites remain the source of truth for validation. Chapters 00:00 Intro: Live from brightonSEO 00:22 Gus Pelogia, Indeed & product-led SEO 01:37 Talk recap: MCP is king 02:00 The magic of connecting multiple tools in one flow 02:46 Reporting to leadership with Claude 03:52 Use case: clustering 4,000 headphone keywords 04:55 Turning keywords into prompts for AI tracking 05:48 Setting up MCP connectors in Claude 07:00 Why prompt engineering isn’t the hard part 07:51 Ad-hoc vs. automated reporting & local results 08:51 Success KPI: time saved 09:36 When NOT to use MCP: 15 cities, 15 calls 10:09 Better: Claude Code writes the script 10:40 Testing the Gemini MCP for web search grounding 11:28 MCP vs. API: data quality & token costs 12:23 Agents & the future of websites 13:12 The “death of the web”? 14:01 Agents for job search on Indeed 15:07 What’s next: automation beyond MCP 15:45 Tracking robots.txt & page updates at scale 17:09 Can you trust the numbers? 18:20 Don’t stop using your judgment 18:37 Outro