The 5000x Signal: Why Google DeepMind and Microsoft Angels Are Betting on "Model Memory" GEO

A British real estate company was practically invisible inside AI answers. Ask ChatGPT where to buy property in the UK, and it didn't even make the long list.

A few weeks later, tracking data showed something startling: AI crawler activity against the company's content had surged 5,000x. Its position in relevant prompts went from "nowhere to be found" to ranking second.

This wasn't paid placement. The operator behind it is a London-based startup called geoSurge, and its cap table deserves a closer look: European veteran fund Passion Capital, Silicon Valley's Tuesday Capital, and a group of angel investors from Google DeepMind and Microsoft AI.

People who build AI for a living are funding a company that optimizes what AI says. What exactly did they see?

A counterintuitive thesis: RAG is a painkiller, memory is the cure

Most GEO services today bet their technical stack on RAG (retrieval-augmented generation): when an AI answers a question, it searches the web in real time, pulls relevant content into context, and synthesizes an answer. The optimization playbook is to make your content easier to retrieve.

Francisco Vigo, geoSurge's founder, offers a different verdict: even when AI uses RAG, what ultimately gets displayed and cited is still driven by the model's "memory layer."

He breaks a model's information sources into two tiers:

• Internal memory: what the model "learned" during training, functioning like a person's worldview, shaping how it interprets everything it sees;
• Real-time retrieval (RAG): searching the web on the fly to fill gaps where memory is incomplete or uncertain.

The critical part: content retrieved via RAG must pass through the memory layer's filtering and interpretation before it gets cited. A brand the model never encountered during training can be retrieved and still judged "unfamiliar, not trustworthy" — and quietly dropped.

Vigo puts it bluntly: "A model can hardly recommend a brand it has never seen in training."

How it started: the shock of a brand disappearing

Vigo was previously a chief data scientist in fintech and an early power user of large language models. As an ordinary user, he noticed he trusted AI answers more than traditional search results — especially when search pages were stuffed with ads.

Then he ran an experiment: asking models the same set of questions, repeatedly, over a long period. What he found should unnerve every brand owner — each time a model version updated, a brand's visibility inside AI answers swung violently, even when nothing about the business had changed. On the recommended list last month, gone this month.

"How to keep a brand from vanishing out of AI answers" became geoSurge's founding question in April 2025.

The three-part stack: measure, attribute, re-engineer memory

geoSurge's product comes in three modules that form a complete pipeline:

Module one: MEASURE. Tracks a brand's current position across major AI systems like ChatGPT — whether it gets mentioned, how often, and how consistently across time and markets. This is table stakes for any GEO service.

Module two: EXPLORE. Here's where it diverges. EXPLORE exposes the model's internal probability distribution and its token-by-token prediction "thought process," telling clients: why did this position form? In which niches, prompt types, and contexts does the brand have a "probability gap" — the openings where visibility is most winnable?

Module three: BOOST. The core technology is called Corpus Engineering: proactively optimizing and preparing brand corpora so that when models train or iterate, the information gets learned, memorized, and recalled more efficiently and accurately. geoSurge also pipes in large volumes of telemetry to verify that client brand data actually ends up inside training datasets.

The contrast with the rest of the market becomes clear: most GEO vendors deliver a dashboard — telling you where your brand appears and offering content suggestions. That is fundamentally "getting retrieved" optimization. geoSurge works on making the model itself recognize and remember the brand — so it recommends the client from memory, even without any live search.

The sober part: nobody can promise you first place

To its credit, geoSurge is candid about the limits: generative systems are probabilistic by nature, and with each model's alignment constraints, it cannot guarantee any client an absolute #1 ranking. What clients can expect: early effects within weeks of technical integration, measured across four indicators — brand mention rate in relevant prompts, stability over time, contextual accuracy, and the frequency and depth of AI crawler activity.

The British real estate firm's 5,000x crawler surge and jump to second place is a real case from one BOOST cycle — but it's a case study, not a promise.

Three takeaways for brands

geoSurge plans to expand into Asia-Pacific this year. For GEO practitioners everywhere, three things are worth chewing on:

First, rethink what "being cited" means. Having your content crawled is tier one. Being retrieved by RAG is tier two. Being remembered during training is tier three. The first two tiers deliver short-term wins that reset with every model update; the third is the asset that survives version cycles.

Second, publish with the training-data lens on. Content on platforms where pages vanish quickly, versus content on authoritative sources that get included in high-quality datasets and remain stably accessible, face completely different odds of entering training corpora. That's why the weight of "slow sources" — authoritative media, encyclopedias, industry reports — never really declines in GEO.

Third, monitor leading indicators, not just mention rates. The frequency and depth of AI crawler activity against your content is an earlier signal than mention rate — crawler behavior shifts before answer rankings do. In the real estate case, the 5,000x crawl surge was the precursor to the ranking jump.

The direction that Google and Microsoft angels are betting on is now plain to see: the second half of GEO isn't about getting found by AI — it's about getting remembered by it.

(Source: an interview feature on geoSurge published by Chinese tech outlet Intelligent Emergence, October 25, 2026)