← Series index · · ~440 words · Field notes

How to teach Claude to read the web.

Three stages, eighteen months apart. The tool you give your AI to read the web shapes what work it can actually do. Most marketers are still on stage one.

Gene Hackman as Harry Caul in The Conversation sits beside a brass reel-to-reel tape machine. A small cartoon ‘Firecrawler’ creature with glowing cyan eyes perches on the reel, reading the tape.

Tuesday I asked Claude to run a brand sweep of the Ogilvy network. Heritage, contemporary positioning, Rory Sutherland’s canon, adjacent lineage. Four research streams in parallel.

Three days earlier the same sweep had failed. Same Claude, same prompts, same target. The only thing that changed was the tool the model could reach for.

Stage one: curl

curl fetches a URL. It was written in 1996. For a brochure-ware site from 2008, fine. For anything that loads its content with JavaScript — most of the modern web — what curl sees is forty-three lines of “Enable JavaScript.” Claude’s built-in web_fetch is the same idea in a wrapper.

Friday’s sweep lived here. I tried WebSearch. It returned snippets — enough to know Devika Bulchandani had once been Global CEO. Not enough to know whether she still was.

Stage one tops out the moment you need to know what a real page actually says.

Stage two: browser hijack

A browser MCP opens a real Chrome window and lets Claude drive it. Real JavaScript. Real cookies. This is the stage most “AI agent” demos picture.

What the demos don’t tell you is the cost. The browser is your browser — when Claude is driving it, you can’t use it. A sweep of sixty pages takes twenty minutes through a browser instead of ninety seconds through curl. My honest Friday verdict: not good enough.

Right tool for one specific job: login-walled pages with no other way in. For everything else, it’s a kitchen-sized hammer swung at a panel pin.

Stage three: purpose-built

Tuesday’s sweep ran on Firecrawl. Firecrawl runs a headless browser server-side. Your Chrome stays yours. Pages render. The model gets clean Markdown at a tenth the token cost of raw HTML. Exa does semantic search — “writers similar to Rory Sutherland” returns thinkers semantically adjacent, not keyword-matched. Both MCP-native. Both free-tier.

Tuesday: four agents, three hours, four source-anchored documents. Under fifty Firecrawl credits of a free-tier thousand. Chrome stayed mine.

The findings were not small. Bulchandani moved to WPP COO; Laurent Ezekiel has been Ogilvy Global CEO since — an integrator, not a creative. The contemporary ogilvy.com reads in WPP holding-company prose: “impactful brand experiences, dynamic ecosystems.” Words David Ogilvy would not have written. The classical posture survives mainly inside Rory Sutherland’s Behavioural Science Practice. None of that was reachable Friday.


Every Marketing Engine Pilot ships Firecrawl-first, Exa when discovery matters, browser MCP only for login-walled corners. Setup: twenty minutes. Cost: rarely cracks ten dollars a month. Client owns the keys at handover.

The stage you’re on is the binding constraint on output quality. Not the model. The tool the model has to reach for.

The Ogilvy sweep that failed Friday and succeeded Tuesday ran on the same Claude. The difference was twelve characters of MCP config and a free-tier API key.

The cheapest upgrade in the stack

Firecrawl and Exa take twenty minutes to install and cost almost nothing. The Marketing Engine Pilot wires them in as standard, so your AI starts working from what the market did this week, not last year. Two co-build slots open.

Talk to Anthony →
Written: 24 May 2026, Dubai
Subject: the research layer most AI installs skip
Series: field notes from running the stack