The Microsoft GEO Guide That Rewired How I Do AI Search (And Took Superpower From 0 to 36,000 AI Mentions)

Why this document matters

The foreword is from Paul Longo, General Manager of AI in Ads at Microsoft Advertising. He frames it around two questions he says he hears from retail leaders every week: how do we make sure AI actually understands our products, and how do we keep our brand's story clear as the rules keep changing? His thesis is that data quality, context, and credibility have become the new currency.

That's the whole game in one sentence. The rest of the guide is the practical unpacking of it.

The core reframe: from discovery to influence

The guide's central argument is that traditional SEO was about clicks, while Generative Engine Optimization is about visibility inside LLM-powered ecosystems, meaning whether an assistant chooses to recommend you at all.

Microsoft's own definition of GEO: optimising content for generative AI search environments so it's discoverable, trustworthy, and authoritative.

The example they use has stuck with me. The SEO version of a product is:

"Waterproof rain jacket."

The GEO version of the same product is:

"Best-rated waterproof jacket by Outdoor magazine, no-hassle returns for 180 days, three-year warranty, 4.8-star rating."

Same jacket. One is a keyword. The other is a set of reasons an AI can cite. Once you see the difference, you can't unsee it, and it changed how I wrote every piece of product content for Superpower.

Importantly, Microsoft is explicit that this isn't a rip-and-replace. Their line is that SEO and catalog investments built the foundation you expand on in LLM-based search. Up-to-date feeds and crawlable, structured content are still the base. What's new is treating your entire catalog and site architecture as content: every detail, benefit, and price signal machine-readable, current, and context-rich.

How AI assistants actually rank products

This is the section worth the price of admission (it's free, to be fair). Microsoft lays out the three overlapping surfaces:

• AI browsers (Edge, Atlas, Chrome with built-in intelligence) that can "see" the page you're on in real time
• AI assistants (Copilot, ChatGPT, Gemini) that meet you in conversation
• AI agents that don't just advise but act: navigating sites, filling forms, completing purchases

Their point is that these aren't three separate boxes; they're overlapping capabilities. The practical question isn't "which box does this live in?" but "what data can this capability access, and how do we make it accurate and trustworthy?"

Then they walk through what happens when someone asks Copilot for a good waterproof jacket for a three-day hike. The system draws on three data layers:

1. Knowledge graph: pre-trained knowledge, real-time web search, product database
2. Page-level data: dynamic content like pricing, on-page structured data, rendered page content
3. User info: brand affinity, location, sizing

That feeds a reasoning phase: natural language understanding, freshness, breaking down and fanning out queries, text relevance, commercial signals, contextual relevance. The output is a natural-language answer with explanations, trusted sources and citations, and product recommendations.

Their worked example for a "rain jacket under $200" query shows exactly what each source contributes:

• Crawled data provides general knowledge ("Patagonia and North Face make quality rain jackets"), category understanding (rain jackets need good waterproof ratings, hiking jackets need to be lightweight), and your brand positioning ("Brand X is known for hiking equipment").
• Feeds provide current prices (your model is $179, competitor is $199), availability (you have stock, competitor is backordered), and key specs (waterproof rating above 1500mm, sealed seams, GORE-TEX/PU/PVC fabric).
• AI decision: your product makes the top three because the feed shows a competitive price and in-stock status.

Read that last line again. The feed won the recommendation. Not the blog post, not the backlinks. That single insight reshaped where I spent Superpower's effort.

The three ways you have to show up

The guide says your business needs to exist in three distinct forms for AI shopping:

1. Crawled data: what AI systems learned in training and retrieve from indexed pages. This shapes baseline brand perception: categories, reputation, market position.
2. Product feeds and APIs: structured data you actively push to AI platforms, giving you control over how you're represented in comparisons. Feeds deliver accuracy, detail, and consistency.
3. Live website data: what agents see when they actually visit: rich media, reviews, dynamic pricing, transaction capability.

And a line I've quoted to every client since: traditional SEO remains essential because AI systems run real-time web searches frequently throughout the shopping journey, not just at purchase time. Your site still has to rank to be discovered, evaluated, and recommended.

They also make the point that the agent layer has a hard dependency on your live site. The purchase flow they describe, where the agent adds to cart, applies a promo code, calculates shipping, completes the purchase with saved payment, and returns confirmation and tracking, only works on a functional live site. Their warning: without your live site working properly, the sale fails even if your feed and crawled data were perfect.

The three strategies (this is the playbook)

01. Data structure: make your catalog machine-readable

Schema implementation:

• Deploy Product, Offer, AggregateRating, Review, Brand, ItemList, and FAQ schema types
• Include dynamic fields: price, availability, color, size, SKU, GTIN, dateModified
• Use ItemList markup on collection and category pages so AI understands product groupings
• For multi-region operations, express localised pricing and language via inLanguage and priceCurrency
• Ship JSON-LD with correct types and attributes
• Write descriptive titles pairing product name with key differentiator (their example: "TrailMaster 30L Hiking Jacket - Waterproof 3-Season Gear")

Real-time synchronisation:

• Sync price and inventory in real time between feeds and on-site schema
• Expose dateModified and availability in structured data
• Include explicit start/end dates for promotions
• Keep values consistent across feed, on-site schema, and what users see
• Ensure the rendered DOM contains the same facts consumers see. Never serve different HTML to bots

02. Content enrichment: design for intent and context

The framing here is that AI assistants interpret queries as intents, so your content should answer real-world questions directly.

Intent-driven product information:

• Front-load descriptions with benefits: who it's for, what problem it solves, what makes it better
• Add use-case context AI can match to queries (e.g. "best for day hikes above 40 degrees")
• Create headings and copy that mirror real-world queries
• Build modular, citable content
• Provide Q&A blocks AI can reason over and cite ("Which size should I pick?")
• Display specs as key/value pairs and feature lists
• Include comparison tables ("Model A vs Model B")
• Add "goes well with" data for complementary products

Multi-modal signals:

• Detailed alt text and ImageObject schema describing visuals
• Video transcripts that parse feature explanations
• Mobile and voice experiences that expose identical structured data, not just desktop HTML

03. Trust signals: establish authority and credibility

Verified social proof: verified reviews with Review and AggregateRating schema; review volume and verified purchase ratios; review sentiment that enables natural-language recommendations ("highly rated for comfort and fit").

Authoritative brand identity: brand identifiers and official social/retailer links in structured data; links to expert reviews and articles where you're featured; certifications and partnerships surfaced as factual entities ("Certified B Corp", "Climate Neutral Certified").

Content integrity: avoid exaggerated or unverifiable claims. The guide states plainly that AI systems penalise low-trust language. Keep a consistent brand voice across touchpoints. Provide structured FAQ and help content that grounds conversational answers.

How I applied this at Superpower

When I started on Superpower's AI search presence, the number of AI mentions was zero. Not low. Zero. Here's how the Microsoft guide shaped the program that got it to 36,000+:

I stopped treating GEO as a content problem and started treating it as a data problem. The rain jacket example made it obvious that the feed and the structured data were doing the heavy lifting in the reasoning phase. So the first sprint wasn't articles. It was a full schema audit and rebuild against Microsoft's list: Product, Offer, AggregateRating, Review, Brand, ItemList, FAQ, all shipped as JSON-LD, with dateModified and availability exposed everywhere.

I rewrote every product and category description in the "GEO version" format. Front-loaded benefits, explicit use-case language, who-it's-for, what-it-solves, what-makes-it-better. Every page got Q&A blocks written to mirror the actual questions people ask assistants. Every comparison that a user might ask an AI to make, we made first and published as a table.

I made the three data layers agree with each other. The guide's "never serve different HTML to bots" and "maintain consistent values across feed, schema, and user-facing displays" points became a hard rule. If the feed said one price and the page said another, that was a P1 bug.

I chased citations, not just links. The trust signals section reframed link building for me. The goal became getting Superpower featured in expert reviews and third-party articles that AI systems treat as credible sources, then surfacing those as structured, factual entities on our own pages. It is the same logic behind how ChatGPT decides what to recommend.

I didn't abandon classic SEO. The guide is clear that assistants search the live web repeatedly throughout the journey. Superpower still needed to rank. The technical SEO work continued in parallel. It just now had a second consumer.

The result: from 0 to 36,000+ AI mentions across the major assistants. That growth tracked almost exactly with the rollout of the three strategies above, in the order Microsoft lays them out.

The takeaway

Microsoft's closing point is the one I'd tattoo on every e-commerce marketer: retailers already hold most of the data signals that influence Copilot and Bing ranking. They're just not surfaced in product feeds. Enrich feeds and content with attributes and trust data, and you help AI understand not just what the product is, but why users love it and when it performs best. They call this "AI ranking readiness", a data discipline that directly impacts discoverability in the age of conversational commerce.

If you only read one thing on GEO this year, read this guide. Then go check whether your rendered DOM, your feed, and your schema all tell the same story. That's where I'd start, and it's where Superpower's 36,000 mentions started too.

For how I measure whether any of this is working, see the citation-share method on the GEO page, and for the revenue stakes, how much traffic sites are losing to AI search.