A digital agency built on thinking, for the global financial services industry.

Microsoft on GEO vs SEO

  • Marketing
5 min read

Article Overview

Earlier this month, Microsoft published what it calls a ‘playbook’. It’s written for retailers who are gearing up for AI to come and steal their lunch. It’s a far broader-reaching document than the pitch would suggest, though.

Contained within is a breakdown of how the AI search process works, straight, as they say, from the horse's mouth.

Before I get into what’s covered, it’s worth explaining why Microsoft’s opinion on this is worth hearing:

  • Microsoft’s Copilot AI tool holds roughly 4% of the AI chatbot market

  • Given that Copilot comes bundled with Office 365 and MS products, I suspect its share of office-based workers is far higher

  • Microsoft is in cahoots with ChatGPT maker OpenAI, to which it provides real-time search results via Bing

  • ChatGPT has an 80% stranglehold on the AI chat market at the moment

The AI search process, according to Microsoft

Microsoft, in documenting its AI search process, has identified five phases, starting with the prompt; that’s the thing you or I enter into the tool.

For quick reference, here are the five phases:

  1. The prompt

  2. Tool selection (agent, browser, assistant)

  3. Data processing (knowledge graph, page-level, and use info)

  4. Reasoning phase

  5. Response

The user, it says, enters their prompt into one of three flavours of AI tool:

  1. AI browser - Microsoft’s own Edge is an example of a browser that can see and help contextualise what you’re looking at online

  2. AI assistant - think ChatGPT; a tool designed answer questions and help with tasks

  3. AI agent - the lesser spotted AI tool that doesn’t just answer you, it completes tasks for you; in retail, that might be purchasing toilet roll. In financial services (what we care about), it could be automatically switching you to the best insurance policy.

In practice, any AI tool worth its salt is probably a mixture of all three of these. Google’s Chrome browser has a search bar that acts as an AI prompt input. It can take you to ‘AI Mode’ and, in time, will almost certainly be hoping to use your card payment details at will.

If the prompt is phase one, and the AI tool processing it is phase two, the processing phase is where things get interesting for marketing and technology teams.

Microsoft has outlined three sources of information that its models use to respond to a prompt:

1. Knowledge Graph

This is the pre-trained data and real-time information gathered via search. The pre-trained data is a bit like how a human uses memory. Everything we learn enters a pool of knowledge and understanding that we use to unpack the world around us. An AI tool uses its knowledge graph a bit like this.

Twinned with that is access to live search data. In the case of Microsoft (and ChatGPT), Bing is the source of power here. Gemini uses Google, obviously.

This step is the first port of call that Microsoft uses in assessing a prompt.

2. Page-level data

Microsoft, interestingly, also references page-level data in its documentation. Now, it’s worth highlighting that this document is speaking to retailers. Retailers maintain online inventory, and this inventory is usually presented to online tools using standardised data ‘feeds’.

Microsoft does, though, reference Schema.org structured markup, and the concept of structured data carries a heavy focus.

I don’t think this means that implementing structured markup is a silver bullet, but if websites and apps are to interface with AI bots, feeds and structured data are an obvious benefit. This document confirms as much.

3. Use information

Finally, Microsoft suggests that it references what it calls ‘use information’. Read customer reviews and online mentions of a brand/product/service/person/entity.

All three of these sources of information are then put into a big mixing bowl and sent to what Microsoft calls the Reasoning Phase.

Reasoning phase

This is where Microsoft says that it applies things such as:

  • Natural language understanding

  • Freshness

  • Break down and fan out queries

  • Text relevance

  • Commercial signals

  • Contextual relevance

There’s a lot of jargon there, but if we skim over that, what it really tells us is that Copilot (in this example) is trying to understand ‘intent’.

It wants to know what you, its user, actually would like. It is trying to figure out if you want to find a fact or whether you’re trying to purchase some breakfast cereal.

The final phase of the process is then the AI tool’s response; it could be text, image, links, spoken word, directions on a map, a downloadable document, you name it.

Why does this matter?

This is Microsoft, one of the big players in the nascent GEO market, giving some (fairly) plan advice on how to approach this new way of finding things online.

The document published is, again, focused on retailers. It speaks of feeds that provide ‘current prices’, ‘availability’, and ‘key specs’.

I don’t think we’re far from a world, though, where financial products begin to use similarly standardised practices as consumer retail.

Microsoft is also transparent with how it uses what it calls ‘crawled data’ to gain:

  • General knowledge

  • Category understanding

  • Brand positioning

Does this tell us how GEO differs from SEO?

It does.

Sort of.

Microsoft says that SEO is for discovery, and GEO is for influence. That’s a little trite for me, but I understand what they are getting at.

SEO has, historically, been a means to an end. Like (for those who remember) opening the Yellow Pages and finding an address, Google lets you, fundamentally, do the same thing online.

GEO delivers influence, as Microsoft puts it, because AI tools are a two-way interaction. You can refine your prompt, you can question the response, and you can ask for the information to be tailored to your needs.

GEO, then, isn’t just a case of showing up in the top spot. It’s far more about being understood, or to view it another way, not being misunderstood.

For retailers, there’s an easy example; if your product feed is broken, AI tools won’t know what you’ve got in stock.

For the rest of us, it means making sure our brand position is crystal clear. It means everything we say and do is consistent and kept up to date.

These rules have always been in place, but AI tools now mean that mistakes are far more likely to be found and compounded.

As I’ve written before, I believe GEO will require brands to take a ‘responsive information’ approach to sharing information online, and nothing I’ve seen Microsoft say here changes my view on that.