Top AEO and GEO News for October: New Frontier Models, Schema's Real Job, and the End of the Prompt

I spend most of my working hours thinking about how content gets found, cited, and trusted by machines rather than by people. That is what Answer Engine Optimization (AEO) and Generative Engine Optimization (GEO) come down to: making sure a generative model selects your page as one of the retrieved sources it grounds an answer in, and then attributes it visibly enough that a human clicks through.

This is not a speculative discipline. Researchers at Princeton University and IIT Delhi formalized it in a 2024 KDD paper that coined the term and built GEO-bench, a 10,000-query benchmark spanning 25 domains. Their headline finding was that targeted, non-adversarial content changes lifted a source's visibility in generative engine responses by up to 40%. The authors were blunt about which tactics carried over from classic search and which did not, writing that "traditional SEO strategies will not apply to Generative Engine settings, highlighting the need for GEO" (Aggarwal et al., 2024).

We are now living inside the world that paper described. Here are the three October developments I think matter most, in listicle form so you can skim or dig as needed.

1. The Frontier Model Race: Who Is Actually Answering the Queries

If you optimize for AI answer engines, you need to know which models are generating the answers. The Artificial Analysis Intelligence Index is the closest thing this industry has to a shared scoreboard. Version 4.1.1 aggregates nine separate evaluations, including GDPval-AA v2, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, and AA-Omniscience, into a single composite figure.

Here is where the current leaderboard stands:

Kimi K3 is the headline for open-weight users. Moonshot AI released it on July 16, 2026 and shipped the full 2.8-trillion-parameter weights to Hugging Face eleven days later, and it now holds the top open-weight position on an index tracking 98 open-weight models out of 178 total.

The instability underneath those rankings is the part I find most instructive. A detailed comparison of K3 against its rivals points out that the ordering flips depending on which reasoning-effort tier you look at. Compare max tiers and GPT-5.6 Sol edges K3 by a single point. Look at the aggregate snapshot using each model's default tier and the order reverses, with K3 at 59.7% sitting above Sol at 58.9%. The author's conclusion is worth quoting directly: "Both statements are true. Neither is a clean 'K3 beats Sol' claim."

That is a genuinely useful piece of intellectual humility, because a great deal of AEO advice treats leaderboard position as settled fact when the underlying measurement is tier-dependent and rescaled every few months.

The numbers that actually affect your budget

Cost per task is where the competitive picture separates most sharply. Artificial Analysis calculated that Kimi K3 runs roughly $0.94 per Intelligence Index task, close to GPT-5.6 Sol at $1.04 and about half the price of Claude Opus 4.8 at $1.80. On the Coding Agent Index the gap widens further: K3 averages $3.18 per task, which the same analysis measured as 55% cheaper than GPT-5.6 Sol max at $7.08, 73% cheaper than Fable 5 max at $11.72, and 59% cheaper than Opus 4.8 max at $7.70.

One technical wrinkle deserves attention from anyone building citation-tracking or retrieval-augmented generation (RAG) pipelines. Artificial Analysis reported that K3's accuracy rate on AA-Omniscience climbed from 33% to 46% versus its predecessor, but noted that "hallucination rate has regressed compared to K2.6, rising from 39% to 51%." Higher intelligence and higher reliability are not the same axis. If your GEO strategy assumes that a fact cited correctly today survives contact with next quarter's checkpoint, that data point should give you pause.

There is also an efficiency trend worth naming. K3 consumed 132 million output tokens on the index versus K2.6's 166 million, a 21% reduction alongside a 13-point intelligence gain. Mixture-of-Experts (MoE) architectures, where a routing mechanism activates only a subset of parameters per forward pass, are the reason a 2.8-trillion-parameter model can be priced competitively at all.

My takeaway: stop optimizing for one model. The public snapshot shows roughly 186 evaluated systems with only about 6.3 points separating the top ten, which means well-structured and well-attributed content has a real shot at retrieval regardless of which engine handles the query. Optimize for the query, the entity, and the evidence trail.

2. Schema Will Not Get You Cited, But It Might Get You Trusted

The second story is Search Engine Journal's piece on schema markup and AI citations, and I consider it one of the more level-headed treatments of structured data published this year, precisely because it declines to oversell.

The central claim is deliberately modest. Schema markup will not make an AI system cite you, but it does help search engines and answer engines understand who you are, verify your claims, and decide whether to feature you at all. The author notes that the payoff is slow, observing that work done on client sites long before the rise of AI search is only now paying dividends.

The framing device the article uses has stuck with me. It compares your structured data to an online dating profile: your first chance to introduce yourself to a search engine, and if the profile does not match reality when someone actually arrives, the piece says, "it's like catfishing." That is a vivid way of putting something I have told clients in blunter language for years. Schema is a promise, and generative engines increasingly audit whether you keep it.

The workflow the article recommends

The implementation sequence is straightforward and, importantly, validation-first:

1. Pick the most specific entity type that honestly describes the page or organization.

2. Assign a stable ID so the entity can be resolved consistently across your domain.

3. Pull facts from the visible page and your platform of record, never from aspiration.

4. Check those values against trusted third-party sources before publishing.

5. Validate, then scale. The article's warning here is sharp: scaling before you validate "just multiplies the same error across a thousand pages." Start with one page and treat it as a model.

The case study in the piece is concrete. The team built a complete Person schema on an executive bio page and connected it to the correct professional profile and company details. Within days, the AI Overview identified the correct individual where it previously had not.

Equally valuable is the article's list of things schema categorically will not do. It will not compensate for thin content or invented credentials. It will not turn a dead social account into corroboration. It will not replace your visible HTML. The author's ordering is the whole argument in one line: earn the authority first, express it clearly, and make it easy to verify.

What independent data says

Third-party research broadly supports the "trust accelerant, not ranking lever" reading. An analysis of schema types for AI search found that pages carrying three or more genuinely applicable schema types earned roughly 13% more Perplexity citations, and reported that structured content had approximately 2.5 times higher odds of appearing in AI-generated answers, with 65% of Google AI Mode citations and 71% of ChatGPT citations in 2026 studies going to pages carrying schema. That same analysis flagged a controlled Otterly AI experiment that found limited direct lift from adding schema in isolation, which is exactly the nuance you would expect if structured data amplifies good content rather than rescuing weak content.

Separate work on entity resolution argues the problem is broader than markup. A review of Ahrefs' study across 1,885 pages concluded that Google's Knowledge Graph, ChatGPT's entity database, and Perplexity's source ranking each assemble entity profiles from cross-domain signals including earned media, Wikipedia references, consistent business data, and executive profiles, and that "a brand with ambiguous entity signals gets skipped regardless of on-page quality."

My takeaway: JSON-LD is infrastructure for verification, not a citation switch. If your claims cannot survive a cross-check against three independent domains, no amount of markup rescues them. If they can, schema meaningfully shortens the distance between accurate and cited.

3. Is the Prompt Box Already Obsolete?

The third piece is an opinion column by Max Snow, CEO of Yobi, published August 27 in The Current, a publication of The Trade Desk. It is written for media buyers, but I think anyone in GEO or AEO should read it, because it reframes what sits between our content and the humans we want to reach.

Snow's opening diagnosis is that every agentic tool currently pitched to marketers works identically. You describe what you want, the agent attempts to deliver it, and it feels like magic the first time. His verdict is unsparing: "this is the 1.0 version of what's possible, and the industry has mistaken it for the destination."

His objection has two parts. The first is about training data. Almost all AI tools built for the open web run the same general-purpose large language models under different logos, and those models, in his phrasing, "learned to read and write by studying everything published on the internet." That makes them excellent at generating creative and orchestrating workflows. It does not make them good at predicting behavior. As he puts it, knowing who is likely to buy and what moves them "is a prediction problem about human behavior; a model that learned everything from text has never actually seen any of it."

The second objection is about the interface itself, and I find it the sharper of the two. Snow argues that the chat box was designed for queries rather than briefs, and that prompting therefore imports a structural weakness: "the gap between what you meant and what you wrote becomes the gap in your results." Your outcomes end up hinging on your ability to phrase things well, which he calls "a strange skill to bet a media budget on."

His alternative is to delete the prompt and substitute a measurable event. Rather than describing an objective in paragraphs, you point at the KPI that matters, tag it with a pixel already on your site, and let a behavioral foundation model trained on real transactions, search activity, and ad engagement optimize toward it directly. Snow reports that Yobi deployed this against a configured purchase event with a performance agency and produced twice the Google-measured revenue on connected TV inventory versus the brand's YouTube campaigns, using half the budget.

I want to be explicit that this is one vendor case study from a party with an obvious commercial interest, not an independently replicated benchmark, and it should be weighted accordingly. But the underlying distinction between language-trained and behavior-trained models is sound and underdiscussed. Snow's closing prescription holds regardless of whether you buy the product: ask the next vendor demoing an agent "which AI model is being leveraged, and what data was used to train it?"

My takeaway: as agentic search and agentic commerce mature, the systems making selection decisions on a user's behalf increasingly sit behind the language layer. Structured conversion signals, verified entity data, and consistent cross-platform facts are what those systems consume. Prompt craft is a depreciating skill.

What Ties These Three Stories Together

Read side by side, all three point the same direction, and the 2024 GEO paper anticipated it. Its authors found that adding citations, quotations from relevant sources, and quantitative statistics produced the largest visibility gains in their benchmark, while keyword stuffing, the classic SEO reflex, actually scored below the unoptimized baseline. Their most encouraging finding for smaller publishers was distributional: the Cite Sources method produced a 115.1% visibility increase for sites ranked fifth in search results, while the top-ranked site's visibility fell an average of 30.3%. Generative engines, in other words, flatten the advantage that backlink accumulation confers in classic search.

Whether the engine answering the query is Claude Opus 5, GPT-5.6 Sol, Grok 4.6, or Kimi K3, and whether the decision layer is a chat prompt or a behavioral agent working from a purchase pixel, the requirement is identical. Your content has to be accurate, attributed, independently corroborated, and structured cleanly enough for a machine to trust on first pass. I will be tracking all three threads into the fourth quarter, and I fully expect the leaderboard section of this piece to be outdated before the month is out.