Author: Johan Peen, COO · Last updated: September 2026
Connecting an LLM like Claude directly to your ad platforms or raw data warehouse produces fast answers, but unreliable numbers.
In our client work at s360, reliable AI-assisted reporting has one consistent prerequisite: a governed semantic layer between the model and the data, so the AI queries certified metric definitions instead of guessing at raw tables.
Almost every team we talk to travels the same road to that conclusion. It has four stages, and each one fails predictably. Knowing them in advance can save you months.
The journey usually starts with a shortcut: someone installs a community-built connector for Google Ads or Meta in Claude, asks a few questions, and gets impressive answers in seconds. Two weeks later, the reports that went up the chain turn out to be full of silent errors.
The errors are not random. They come from the model working on raw API data with no shared definitions. Ad platforms disagree by design. Meta’s default attribution window is 7-day click and 1-day view, while Google Ads defaults to data-driven attribution, so their “conversions” were never the same number to begin with. Add accounts running in different currencies that get summed without conversion, and micro-conversions counted alongside primary leads, and you get ROAS figures that match no platform dashboard anywhere.
There is also a governance problem hiding under the convenience. There is no native Google Ads connector in Claude’s directory. Google’s official Google Ads MCP server, open-sourced in October 2025, is a self-hosted, read-only developer tool. Most directory connectors are community-built wrappers that route your ad data through a third party’s infrastructure, with no metric definitions and no contract about where the data goes.
The predictable outcome: the numbers get challenged once, trust evaporates, and the experiment gets scrapped.
The usual second stage is a managed aggregator such as Supermetrics. This is a real improvement: these platforms normalize API schemas, convert currencies, and blend channels into clean, consistent reporting. For cross-channel spend and performance dashboards, they do exactly what they promise.
What they cannot do is speak your business language. The semantics are generic, and the vendor’s definitions of a conversion, a lead, or revenue, not yours. And marketing questions stop being platform questions quickly: the moment leadership asks about CAC against unit economics, pipeline contribution, or pacing against budget targets, you need ad data joined with CRM records and finance data. That data lives in your warehouse – for most mature teams, BigQuery or Snowflake.
This is the pivotal stage, because the idea sounds so reasonable: the data is already centralized, and Claude writes excellent SQL. But direct text-to-SQL over raw tables is where an experienced data team starts asking uncomfortable questions, and each one points at a real architectural gap.
The summary answer: Claude will happily write syntactically perfect SQL on top of raw tables, but access to data is not the same as understanding data logic.
A semantic layer is a governed translation layer that sits on top of the data warehouse and defines, in one version-controlled place, what your business metrics mean: how revenue is calculated, which conversions count, how tables join, and who may see what. Tools in this category include the dbt Semantic Layer (powered by MetricFlow), Cube, and Looker’s LookML.
For AI analytics, the semantic layer changes the model’s job completely. An LLM writing raw SQL is doing probabilistic generation: predicting plausible queries, inventing join paths and metric formulas as it goes. An LLM in front of a semantic layer only maps a natural-language question to certified metrics and dimensions; the layer then compiles the query deterministically, with correct joins, partition filters, and access policies applied before any SQL runs. Same question, same answer, every time. And when a question falls outside the modeled scope, the system can say so instead of hallucinating a number.
The connection between the two is increasingly standardized through the Model Context Protocol (MCP), the open standard that lets Claude and other LLMs discover and query governed tools rather than raw database catalogs.
The setups that actually work don’t let agents trawl raw tables. Getting there is a data project before it is an AI project:
| Approach | Answer reliability | Cost risk | Governance | Maintenance burden |
| Community connector on ad platforms | Low – no shared definitions | Low | Weak – third-party data flow | None (and it shows) |
| Managed aggregator | Good within its scope | Low | Vendor-defined semantics | Low |
| LLM direct on raw warehouse | Unpredictable | High – unbounded scans | Depends on auth model | High – breaks with schema drift |
| Semantic layer + MCP | High within modeled scope; refuses out-of-scope | Controlled | Central RLS/CLS before queries run | Real but owned and versioned |
Yes, but not through an official one-click connector. Google’s official Google Ads MCP server (open-sourced October 2025) is read-only and self-hosted, requiring your own developer token. The connectors in Claude’s directory are third-party or community-built, which means your data routes through their infrastructure. Either way, you get raw API access, not governed reporting.
Because the model is aggregating raw data without your definitions. Different attribution windows (Meta’s 7-day click/1-day view vs. Google’s data-driven attribution), unconverted currencies, and micro-conversions mixed into primary conversions each produce numbers no platform dashboard will confirm. The fix is normalized data and shared metric definitions, not a better prompt.
Only with the right authentication architecture. Warehouse security policies apply to the connecting account – an agent using one shared service account bypasses per-user row-level security in practice. Safe setups use per-user credentials or a semantic layer that enforces access policies before queries are compiled, plus logging of every query the agent runs.
For cross-platform questions, effectively yes. An LLM on top of individual ad platforms inherits each platform’s incompatible definitions. Combining ad data with CRM, revenue, and budget data requires a central store – BigQuery or Snowflake in most stacks – and a modeled layer on top that defines how it all joins.
The main options are the dbt Semantic Layer (powered by MetricFlow), Cube, and Looker’s LookML – all of which expose governed metrics that LLMs can query, increasingly over MCP. The right choice depends on your existing stack; the principle matters more than the tool: the AI should query certified definitions, not raw tables.
We help marketing and data teams get from stage one to stage four: the warehouse foundation, the semantic layer, the MCP integration, and the evaluation setup that makes AI-assisted reporting trustworthy enough to send up the chain. If you’re exploring AI-assisted dashboards or analytics agents, or you’ve already been burned by the quick version, contact us to talk about how to structure this for your stack.
Read more news
Search audience can be considered as split into two webs: one navigated by humans, and the other by agents. Both demand data they can trust. This session explores what data integrity means in modern Search – from entity relationships and schema to emerging standards like ACP, UCP and more – and why the structural quality of your data now determines visibility across both. Whether you’re optimising for a person or an agent making decisions on their behalf, this session will explore what to consider for both.
About Alex:
Alex is an expert-level SEO consultant, WordPress developer, and marketing strategist with over 20 years of experience helping brands grow their online presence, conversion, and customer loyalty. He specialises in technical and structural SEO, product development, and strategy, and has worked with thousands of websites across sectors.
He is Principal SEO at Yoast, the world’s leading SEO plugin for WordPress with over 13.5 million installs, where he focuses on product innovation and brand authority. He is a co-presenter of The SEO Update by Yoast video podcast, as well as being a guest on numerous other industry podcasts. Alex also co-founded the organic marketing and WordPress development agency FireCask, and is a long-standing conference speaker and trainer on topics including technical SEO, AI Search, and WordPress SEO.
In my session, I’ll share Google’s vision for a fundamental shift in digital behavior: where search evolves from retrieving information to executing intent. We’ll explore how the transition from “searching” to “delegating” is reshaping Search advertising.
About Anders:
As the Search and Performance Lead for Google Northern Europe, Anders is responsible for helping the region’s largest advertisers bridge the gap between core business objectives and Google’s advanced performance suite. He leads cross-functional initiatives focused on maximizing ROI through AI-driven solutions and deep consumer insights.
As a regional lead for Agentic Commerce and the Future of Search, Anders works closely with Google’s global product teams to ensure Northern European retailers are well positioned and prepared for the next era of Search advertising.
Marketing in 2026 demands more than speed. It demands autonomy. This session breaks down how to architect agent-powered workflows that automate research, content creation, and optimization end-to-end.
You’ll learn when to reach for static prompts vs. dynamic agents vs. full AI workflows, and see real examples of small teams scaling output with simple tools like Apps Script + Claude Code, replacing entire workflows, not just individual tasks.
The talk will also challenge the “AI hallucination” narrative head-on, showing why a well-designed AI system is measurably more accurate than manual human processes.
About Tom:
Engineer by craft, marketer by obsession. Tom founded SEOwind to build AI systems that scale content. Today, SEOwind produces 2,300+ articles monthly for 500+ clients, primarily through agency partnerships. He speaks at SEO and marketing conferences on AI-driven SEO and growth.
Search is no longer a single ecosystem. SERPs and LLMs often surface and cite completely different sources, which means the links that influence rankings aren’t always the same signals that influence AI visibility.
Based on large-scale research and real client data, this session explores why different systems reward different signals and shows how to design link strategies that perform across them.
About Tamara:
Tamara Novitović is Head of SEO at Bazoom, where she builds and tests modern authority strategies across competitive niches and international markets.
She focuses on redefining link value, user-first placements, and hybrid tactics built for both SERP visibility and AI citations.
She is also a regular speaker at international conferences and contributes to industry conversations on search, content, AI, and link strategy.
Search is undergoing a massive transformation. With Google AI Mode and UCP live in the US we are now entering a new phase in both paid and organic search.
Mathias will give you an overview of what we know will happen as well as a checklist of things you can work on that will benefit you both short and long term. Gain a grounded perspective on the structural changes hitting the industry and how to prepare for 2027 and beyond.
About Mathias:
Mathias Hillerup is the co-founder of s360, having started the agency in 2011. Over the last decade, he has helped scale the company into a leading digital consultancy by bridging the gap between marketing, technology, and data.
He works at the intersection of technical infrastructure and commercial growth, helping international brands adapt to a rapidly maturing digital landscape.
Mathias is known for a pragmatic, business-first approach to technology, focusing on the structural shifts that actually drive long-term performance.
Paid Search has evolved through several major shifts: Google Shopping, Smart Bidding, and the disruption of the pandemic years. Today, a new phase is emerging driven by AI.
As AI takes over more of the decision-making, the old playbook of granular account structures and manual control is becoming less relevant. Success increasingly comes from simplifying setups and feeding the systems with the right data.
In this session, Frederik Hyldig explores how the fundamentals of Paid Search are changing and what that means in practice. He will outline how advertisers should rethink account structure, how to work with automation instead of against it, and why data and feeds are becoming central to staying competitive.
About Frederik:
Frederik Hyldig has over a decade of experience in Paid Search and has advised some of Europe’s largest brands on media strategy and technology.
His focus has always been on the practical application of technology in search. Today, he helps brands move past manual management to build structures that actually work in an automated environment.
Frederik is a frequent speaker and is known for his straightforward approach to SEM and digital strategy.
To stay visible in LLMs, you must master Generative Engine Optimization (GEO). This keynote is a practitioner’s guide to becoming the primary source for AI-driven answers like Gemini and Perplexity.
Moving past theory, I will share concrete technical and content-led tactics to improve LLM visibility and transform your brand into a definitive authority for AI.
Leave with a 2026 checklist to turn AI search shakeup into a measurable, competitive advantage.
About Nikolaj:
Nikolaj Mogensen is the VP and Group Lead of SEO at s360. Over the last 15 years, he has worked hands-on with some of the largest websites in the Nordics, helping them navigate the constant shifts in search. His current work centers on AI and GEO, finding real-world ways to maintain organic visibility as search evolves.
A regular speaker at events like brightonSEO and SMX, Nikolaj focuses on turning technical complexity into practical steps that work for businesses.
SEO is shifting from optimizing pages to building search-driven products. In this talk, Björn Darko shares how large marketplaces turn SEO insights into scalable product features. He’ll also explore how AI search, LLMs and changing user behavior are forcing SEO teams to rethink how they measure visibility, traffic and growth & how to inform your board about it.
About Björn:
Björn Darko is Group Head of SEO & AI Search at Swiss Marketplace Group, overseeing organic growth across Switzerland’s leading marketplaces in real estate, automotive, finance and classifieds. With more than a decade in SEO and product leadership at companies like AVIV Group, Searchmetrics, Ringier etc., he focuses on bridging the gap between SEO, data and product development. Björn also hosts the SEOPRESSO podcast and regularly speaks about Product-Led SEO and the future of search.
A common pattern shows up account after account: over-segmentation, premature automation, and budgets spread too thin to generate meaningful learning. Small budgets rarely fail because of poor tactics. They fail because strategies designed for higher spend environments are applied without adjustment, and what works at scale, collapses when resources are limited. In this session, Ramial covers the mindset shifts and structural principles required to make PPC sustainable in constrained environments.
About Ramial:
Ramial Aqeel is a PPC freelancer who works primarily with SMEs, running paid media on lean budgets. He focuses on clear, practical approaches that help advertisers make limited spend work harder. His work centres on simplifying PPC and supporting smaller businesses in getting the most from every pound they invest.
Oliver will run through the latest AI-enabled features of Screaming Frog’s SEO Spider, as well as some lesser-known classic features that you might have missed along the way. This talk is for anyone looking to take their day-to-day SEO to the next level, though there’s lots to be gained if you’re not too familiar with auditing websites too. No frogs were harmed during its production.
About Oliver:
Oliver has a decade of experience at Screaming Frog, having conquered SEO consultancy and management, he now oversees the agency’s wider strategy. He’s spoken at SEO conferences in 6 countries, so he unironically considers himself an international SEO rockstar now.
In this session, we’ll walk through the mechanisms of AI search systems discussion things like personalisation, query expansion, memory, and response generation. We’ll go over how this is different than traditional search, and what carries over. Attendees will leave with a framework for winning in both traditional and AI search, without compromising visibility in any domain, but instead strengthening the content message in an omnichannel world.
About Lazarina:
Lazarina Stoy is a marketing consultant and the founder of MLforSEO, a learning platform, academy, and community for marketers and organic search professionals. She designs hands-on training for in-house and agency teams, and the forward-thinking individuals within them, integrate AI and ML solutions into their workflows to drive productivity and organic growth – regardless of the platform their audience is searching on. A strong advocate for responsible automation, Lazarina speaks at conferences and webinars and creates practical resources for marketers starting their AI/ML journey.
