Research · § 00

Field notes from inside the LLM brain.

ktau publishes what it learns. We share the models we build, the data we observe, and the implications for anyone serious about AI visibility. Open access. No gates. Cite freely.

Your brand is on the web. That doesn't mean the model ever read it.

A community fiber network in Vermont has its own website, coverage maps and customer reviews. Ask a leading AI model to name internet providers and it names Xfinity three times out of four — and the local provider never. We followed 248 brands from the open web, through the filtering that turns the web into AI training data, into the models and about 179,000 AI answers. The brands that get recommended are not the ones mentioned most. They are the ones mentioned in the right context — and whose pages make it into what the model learns from.
Read the post →

The AEO playbook doesn't survive a fair test. The model's memory does.

We followed 158,099 AI answers and 468,045 citations for 90 days across ChatGPT, Gemini and Perplexity, then replayed every question through the engines' own APIs. FAQ blocks, schema markup, summary boxes and author bylines stop mattering the moment you compare pages of the same website.

Across websites, every recommended practice appears to work: a page with an FAQ block receives 18% more citations than one without, a summary box 21%, an author byline 15%, schema markup 8%. Compare pages of the same website and the effects vanish — −3%, +7%, −5%, −5%, none of them significant. The practices are concentrated on websites that are cited for other reasons. Everything we could measure about a page explains 2.8% of the variation in ChatGPT's citations; the engine's own retrieval explains 27%.

Figure 1 · The five recommended practices, across websites vs within the same website (consumer market, 12,031 pages on 1,117 websites)
% MORE CITATIONS · PAGE HAS THE PRACTICE VS NOT −10 0 +10 +20 +30 Has FAQ block +18.1% −3.4%Has summary box +20.7% +6.7%Has author byline +14.9% −4.6%Has schema.org markup +7.5% −5.4%Faster page (+1 SD) +11.7% −1.1% Across all websites Within the same website (none significant)

What does decide a citation: whether the page answers the questions people ask in its own title and paragraphs (+14% and +16% per standard deviation, up to +26% on third-party pages), and whether the engine's search retrieves it — every one of the 8,646 URLs ChatGPT cited was a page its own search returned, and 97% of those pages are absent from Google's top 50. The model's memory shapes both steps: ChatGPT writes brands it already knows into 60% of its searches, 47% of its citations come from websites it would name without searching, and in controlled experiments revealing brand names behind identical product facts reorders its rankings up to 3× more than re-running the prompt. Pages are cited when they are searched; brands are named when they are remembered.

Read the full note →

From Invisible to Recommended: A Controlled 30B-Parameter Study of Training-Time Brand Imprint

We retrain a 30-billion-parameter language model from scratch to isolate the parametric channel — and show that 250 engineered, factual documents can move a zero-visibility brand to a 42% recommendation rate on its target topic, with no measurable loss in benchmark accuracy.

Measured by retraining a 30B model from scratch with ktau’s 250 targeted, factual documents — and holding standard-benchmark accuracy at or above baseline — a brand with zero presence goes from 0.000 to a 0.421 recommendation rate and 0.443 mention rate on its broad topic, and 0.290 / 0.300 on the specific attributes it set out to win. The model never saw the evaluation queries: it generalized.

Figure 1 · Brand visibility lift from a 0.000 baseline — recommendation, mention, and rank
BRAND VISIBILITY · 0.000 BASELINE → AFTER CONTENT 0.50 0.25 0.00 0.421 0.443 0.474 0.290 0.300 0.240 TOPIC FOCUS · BROAD TARGETED · X + Y BRR recommendation BMR mention MRR mean recip. rank
Read the full note →

Presence in LLMs Is an Optimization Problem

Why showing up in generative recommendations is something brands win deliberately — and have to keep winning.

When someone asks an LLM to recommend a product, a vendor, or a tool, the answer is neither random nor fixed. It is the output of a system that can be measured and influenced. That reframes a question brands used to leave to general reputation or to classical search marketing: appearing in a generative model's recommendations is not a matter of luck. It is an optimization problem — and, like any optimization problem, it rewards brands that approach it deliberately and offers little to those who treat it as a one-time task.

This note sets out how we think about that problem at ktau.ai, and what our research has found about moving a brand to the top of LLM recommendations.

Read the full note →

The recommendation surface: a working model of how LLMs choose.

The category treats the LLM as a black box. We propose a different model — one with measurable structure, identifiable leverage points, and a defined optimization objective.

The dominant pattern in generative engine optimization treats LLM recommendations as outputs of a black box. Tools measure citations, mentions, and sentiment, and infer the rest. We argue this leaves too much on the table. The LLM's recommendation is the outcome of a probabilistic surface shaped by training-time priors, retrieval-time signals, source authority, attribute matches, and conversational context. Treated formally, the surface becomes a measurable object — and the strategy stops being a hunt for keywords and starts being engineering against a known objective.

Figure 1 · Recommendation surface, brand-specific

In this paper we describe the model formally, lay out the techniques ktau uses to probe it at scale — query simulation, source-influence mapping, attribute-level decomposition — and walk through three case studies showing how leverage-point optimization outperforms volume-based content strategies. We also discuss the limits: what the model captures well, what it does not yet, and how it evolves as the foundation models underneath continue to change.

Read full paper →

Retrieval vs. training: why most GEO optimizes for half the game.

Most GEO tools optimize for the moments an LLM searches the web mid-answer. That's the easier half — and the one that resets with every new model.

Generative Engine Optimization tooling has converged on a single optimization target: the retrieval step. Probe the LLM. Count citations. Tune content to be picked up the next time the model reaches out to the web. Retrieval matters. It is also half the surface. The other half is what the LLM learns about your brand the next time it trains — the durable, permanent impression that produces recommendations without any retrieval at all.

Figure 1 · Two windows of impact

This paper makes the case that retrieval-time optimization, on its own, is a treadmill. Each model release resets the gain unless the content also imprints into training. We define the two modes precisely, identify the small set of content-engineering choices that earn both, and argue that the next generation of GEO platforms must treat training-time imprint as a first-class objective. We close with practical guidance for content teams and an open question for foundation-model labs.

Read full paper →

Source influence: an empirical map of the sites that move LLM recommendations.

We probed 47 categories with 1.2M queries across six engines. The result is a tight power law — and a per-category list of the sites that actually matter.

Common wisdom in GEO is that "authoritative sources" matter. We wanted to know which ones, in which categories, and by how much. Over Q1 2026 we ran 1.2 million simulated buyer queries across six major LLMs (ChatGPT, Perplexity, Gemini, Claude, Copilot, Google AI Overviews) covering 47 commercial categories. For every cited answer, we traced the underlying source. The finding is robust and uncomfortable: in any given category, fewer than 30 domains supply more than 60% of the LLM's information.

Figure 1 · Source concentration · power law across 47 categories

The map is uneven across categories. SaaS leans heavily on G2 and Capterra; healthcare on regulatory and clinical sources; CPG on review aggregators and lifestyle media; finance on a tight set of analyst sites. We publish the per-category top-10 source list, the methodology, and a public dataset for replication. The implications for content placement, source partnerships, and earned-media strategy are immediate — and uncomfortable for anyone still treating GEO as a content-volume game.

Read full paper · download data →
§ Subscribe

New research, monthly. Direct from the team.

One email per month with new papers, datasets, and field notes from the ktau research team. No gating. No promotion. Unsubscribe in one click.