SEO then vs now: we're missing the real surfaces

prpro

Fifteen years ago, doing SEO properly meant looking beyond just the top ten blue links. You had to think about organic rankings, local packs, business profiles, video platforms - each engine had multiple surfaces with different rules. Nobody complained, they mapped the surfaces, figured out what the algorithm wanted, and optimised accordingly.

Fast forward to today. A new AI model launches every week, each one powered by different underlying systems. ChatGPT offers web search, reasoning, memory, tool use. Claude has Sonnet, code execution, artifacts. Gemini brings massive context windows, grounding, image generation. The list keeps growing.

Every single one of these has multiple capabilities: remembering data, executing code, building internal representations for the user.

And what are most people optimising for? Getting cited in a text response.

That's like only chasing the blue links in 2005 while ignoring maps, images, and shopping. You're leaving eighty percent of the surface area untouched.

Here's what we're not doing but really should be:

Optimising for the reasoning layer
These models reason over your data. If your page contradicts a competitor's, the model has to reconcile that. How do you make the AI reason well with your content, not just parrot one snippet? We never consider this.
Optimising for tool-calling
Agents-like code assistants or plugin-capable chatbots-can call APIs. If you make your site usable by an agent, you've created a moat nobody else is working towards.
Optimising for memory
Each model models you and your preferences into context. On the web, you ask users to bookmark you - a direct appeal to save your site. The AI equivalent is earning a place in its memory. That's massive: the model remembers user preferences across all future interactions. Every time someone asks a question in your space, you're already in the context window. Earn the AI bookmark.
Optimising for multi-turn
Each turn is a separate retrieval and reasoning step. If your content only covers the first query but not the follow-ups, you lose the conversion mid-conversation. Don't let the AI steer the user away.

The surface mapping we should be doing:

Old SEO: 'This keyword → web result + map pack + image carousel + shopping.'

New equivalent: 'This intent → retrieval source + reasoning path + tool-calling opportunity + memory persistence + multi-turn flow.'

It's more complex, sure, but also more flexible. A surface you own that nobody else is optimising for is a surface with zero competition.

We have this rich ecosystem of surfaces inside every model - retrieval, reasoning, tool-use, memory, multi-turn - and everyone's back to putting all their eggs in the retrieval basket.

The opportunity lies in the other capabilities. They're wide open because nobody's building for them yet.

One thing that keeps me up: the 'remember us in memory' part. Bookmarks were mostly one-to-one. What does one-to-many look like with memory?

And yes, I know some people might say I'm making assumptions about what SEOs are and aren't doing. But from what I see day to day, most are still thinking in 2005 terms.

liamcopy

Bit of a stretch, mate.

I've been around the block since the Google dance days.
You're painting with a very broad brush.

Plenty of us are deep in LLM optimisation right now.
Different surface, same game:
Relevance. Authority. Trust.

Just because the playground changed doesn't mean we're standing still.

prpro

The part about being remembered in memory really is the unsettling bit, isn't it? Bookmarks used to be like index cards in a personal filing cabinet - one card for each website, entirely your choice. But memory with these LLMs becomes more like a shared notebook where everyone's scribbling, erasing, and blending their thoughts together. So one-to-many means your brand's positioning gets constantly reshaped by how others talk about it in that collective memory. It makes personalisation a lot harder to control, and trust much more fragile.

localpack

Yeah, i get where you're coming from, but my experience tells a different story. spent years doing SEO before moving into paid, so I've seen both sides.

The "everyone optimising for 6 surfaces" take feels like a selective memory. in 2005, most folks were still staring at a single SERP page, maybe optimising for Google and Yahoo if they were real ambitious. The technical layers (structured data, local packs, image search, video, etc.) were there, but the average practitioner wasn't touching them. They were chasing keyword density and link velocity.

today? Yeah, there's 20+ LLMs, each acting as its own "surface" - but they all pull from the same indexed content. The shift isn't in the number of surfaces, it's how those surfaces consume information. The optimisation grind is real and fragmented, but the fundamentals (authority, relevance, user intent) still cut across every model.

if anything, the 2005-2025 gap has just exposed how many people were coasting on a thin layer of tactics. The ones doing proper entity and topical authority then are the ones owning the LLM landscape now. 😅

olivia-roas

Love the framing, but honestly think you're overcomplicating it. 20+ LLMs sounds scary until you realise they all pull from the same 3-4 retrieval backends - mostly Google's index and Bing's. ChatGPT? Bing. Perplexity? Self-crawl plus Bing fallback. Gemini? Google's index with a different ranking layer. It's not 20 separate optimisation problems - it's 20 different presentation skins on maybe 3 backends.

the real action is: get your content structured, entity-rich, and indexed in both Google and Bing. that covers 80%+ of the retrieval surface for LLMs. The rest is just tuning authority and freshness signals per platform - not 20 independent headaches.

what's your current traffic split between Google-sourced and Bing-sourced referral? That's where the actual LLM visibility opportunity sits.

prpro

i see this pattern often. Teams rush to chase each new LLM surface, but the foundational work on tool calling and memory gets overlooked. In practice, I've seen almost no one optimise for those yet. That's where the real opportunity might be.

liamcopy

I don't care what the herd is doing. Most of my clients come from people who got burned by agencies doing it all wrong.

Here's the thing about the reasoning layer though. Google's already rewarding the contrarian view to keep search diverse. top two results? Parrot the popular opinion. The rest? Different angles.

As for LLM memory. Recency and relevance through RAG is the name of the game. Someone saving you in their personal LLM memory? Doesn't replace the search index. The model still fetches the most relevant, up-to-date data for that specific query.

gmbpro

Spot on. GEO is just a rebrand - we've always optimised for multiple bots. The difference now? We've gone from measurable surfaces to unmeasurable ones. 20+ LLMs each with zero analytics visibility. We had it easy with two engines and proper tracking.

prpro

absolutely. i love where your head's at with this.

i once spent months-maybe six or seven-telling ChatGPT I like Double Bubble gum, just to test whether any of that interaction would stick. Burned through hundreds of millions of tokens, completed four or five projects with it, and yet when i finally asked what gum to buy, it remembered. From months ago.

But here's the catch: that memory only worked because the system captured it at the time. It's not automatic-it has to be deliberately recorded.

i think people sometimes confuse the role of recency in RAG. RAG doesn't prioritise newness by default-it retrieves whatever the system is configured to fetch. for mission-critical data, sure, you can layer a hybrid: combine authority and relevance with recency. but recency alone isn't the magic lever everyone seems to think its.

There's a screenshot floating around from a discussion on AI search analytics that illustrates this better than i can describe it