Inside SellThru's AI SEO Infrastructure — Why We Built It, What It Does, and Where It's Going

At some point in late 2025, we made a decision that most agencies our size don’t make: to stop waiting for off-the-shelf tools to catch up with what the market actually needs, and to start building our own. Not because we enjoy engineering for its own sake, but because the gap between what standard SEO tooling offers and what clients in the UAE and Saudi Arabia actually need had become too wide to ignore.

This article is a transparent breakdown of what we’ve built, the reasoning behind each decision, and where the infrastructure is heading. If you’ve read our piece on how we ranked for “Best SEO LLM Agency Dubai” — this is the engine behind that result.

Three problems kept appearing in our client work that standard tools couldn’t solve well enough.

The infrastructure has three layers: a data layer that pulls live signals from multiple sources, an integration layer that routes and structures those signals, and an AI analysis layer powered by Claude that turns structured data into consistent, calibrated deliverables. Here’s how they fit together.

Five live data connections power every audit we run. Three are standard to any serious SEO practice — DataForSEO for keyword and SERP data, Google Search Console for impression and ranking data, and GA4 for organic traffic and AI referral attribution. Two are specific to our LLM SEO work and are the reason we had to build rather than buy.

The ChatGPT API and Gemini API connections enable active querying — we submit market-specific prompts and analyse the actual responses returned. This is the only methodology that works for MENA. Every passive LLM monitoring tool returns empty data for GCC markets. If an agency is telling you they’re tracking your LLM visibility in UAE or KSA using a SaaS dashboard, ask them where that data is coming from.

The MCP (Model Context Protocol) server is the connective tissue. It’s a server we host internally at mcp.sellthru.me, managed by our AI developer, that allows Claude to query live data sources directly during analysis — rather than working from static CSV exports. This matters because audit outputs should reflect current conditions, not last week’s data pull.

MCP is an Anthropic-developed standard that gives AI models structured access to external tools and data sources. By hosting our own server, we control what data Claude can access, how it’s formatted before it reaches the model, and how outputs are structured. It also means we can extend the system — adding new data sources or report types — without rebuilding from scratch.

Claude processes all data through a structured prompt library we’ve built and refined across dozens of real client audits. The reason we use Claude specifically — over other models — comes down to instruction fidelity. SEO audit methodology is precise: the right metric, from the right source, calibrated correctly, interpreted in the right market context. Claude follows complex, multi-part instructions without drifting from them.

The prompt library currently covers four output types — each a distinct document with its own data requirements, structure, and market calibration rules. The architecture behind this library is what we’re covering in the next piece.

This infrastructure is not just an operational efficiency play. It’s a deliberate positioning bet. The SEO LLM agency market in Dubai is still early. The agencies that establish measurable, repeatable LLM visibility capability now — while the competitive set is small — will be significantly harder to displace when the market matures.

We’ve already seen this play out on our own brand. The infrastructure we built to serve clients is the same system we used to rank SellThru for “best SEO LLM agency Dubai” — appearing in Google’s top three and being cited by name in Gemini’s AI Overview. The methodology is proven on our own domain before it’s applied to a client’s.

For clients, this means their LLM visibility audit is not a one-off project produced by an analyst manually querying ChatGPT. It’s a structured output from a system that runs repeatedly, producing comparable results over time, with consistent calibration across UAE and KSA markets.

The infrastructure described above is the foundation. What sits on top of it is the SEO Prompt Generator — a structured prompt library, currently at version 10, that defines exactly how Claude analyses data for each of the four output types. It specifies which metrics to pull, how to interpret them, what calibration to apply, and how to structure the output.

The prompt generator is what turns a capable AI model into a consistent SEO analyst. Without it, Claude produces variable outputs depending on how questions are phrased. With it, every audit follows the same methodology — regardless of who triggers it.