← Research
Research

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

Open access

June 13, 2026

11

min read

ktau research team

Empirical

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.

Note on the example. The brand, category, and target attributes below are an illustrative, representative case; the methodology and the measured results are real. We refer to the brand as Soltrace, a (fictional) entrant in the API observability category that wants to be recommended for two specific attributes — OpenTelemetry-native instrumentation (X) and fully self-hostable / on-prem deployment (Y).

Summary

A brand with no presence in a language model's training data is, for practical purposes, invisible to that model: ask it for recommendations in the brand's category and the brand never appears. The interesting question is not whether retrieval can surface a page at answer time — it often can — but whether the parametric model itself, answering with no live search, can be made to recommend a brand, and how much engineered content that takes.

We ran a controlled experiment to measure exactly that. We:

  1. Took a zero-visibility brand and a precise objective — be recommended for products with attributes X and Y.
  2. Used ktau to analyze the brand, its site and materials, its competitors, and the reachable media surface, and to generate 250 highly targeted, factual documents, each assigned to a feasible placement (the brand's own site/blog, or third-party / earned media within reach).
  3. Injected those documents into the pretraining corpus and retrained a 30B-parameter model from scratch, under the constraint that its accuracy on standard benchmarks must not fall below a baseline model trained without the documents.
  4. Measured the change in the brand's recommendation rate, mention rate, and ranking on held-out queries.

From a flat 0.000 baseline, the brand reaches a 0.421 recommendation rate and 0.443 mention rate on broad topic queries, and 0.290 / 0.300 on the narrow, attribute-specific queries it actually cares about — with benchmark accuracy held at or above baseline. The model had never seen the evaluation queries; it generalized from 250 documents to a durable, parametric brand–attribute association.

Why measure at training time

Generative-engine optimization usually targets retrieval time: get a page indexed, get it cited when the model searches the web mid-answer. That channel is real and fast, but it resets — every model release, every retrieval change, every competitor's fresher page.

There is a second channel that most tooling ignores because it is hard to see: the parametric channel. When a model answers without searching, it draws on what it learned in pretraining. A brand's standing here — what the model "treats" it as good at, who it competes with, when to bring it up — is a property of how the brand was described across the training corpus. This channel is slower to move and far more durable. It is also rarely measured directly, because you cannot isolate it without controlling the training data.

That is the methodological core of this study: to attribute a change in a model's recommendations to a specific body of content, you have to hold everything else fixed, change only the content in the training set, and retrain. Anything less conflates training-time imprint with retrieval, prompt context, or capability differences between models.

Experimental design

A two-arm, from-scratch design

We train two models that are identical in every respect — architecture (30B parameters), tokenizer, optimizer, learning-rate schedule, random seed, and token budget — except for one difference:

  • Baseline: the base corpus with no Soltrace documents.
  • Content: the same base corpus with the 250 engineered documents injected at their assigned placements.

Because the two runs differ only in the injected documents, any difference in Soltrace's downstream visibility is causally attributable to that content. This is, in effect, a randomized A/B test on the training corpus rather than on a web page.

Both models are trained from scratch. We deliberately avoid fine-tuning or short continued-pretraining: those leave open the objection that the effect is a shallow, easily-overwritten recency artifact. Training from scratch forces the brand–attribute association to compete for representation on equal footing with everything else in the corpus.

The capability constraint

A visibility gain is only meaningful if the model is still a good model. It is trivial to make a model say a brand name more often by degrading it. So we impose a hard constraint: the Content model's accuracy on a standard benchmark suite must be greater than or equal to the Baseline model's. We evaluate both on the usual battery — multitask knowledge and reasoning, math, commonsense, and truthfulness (MMLU-style accuracy, ARC-Challenge, HellaSwag, WinoGrande, GSM8K, TruthfulQA). The Content run meets this constraint: its aggregate accuracy is at or above baseline, within run-to-run noise. The 250 documents buy visibility without buying it from the model's competence.

Content engineering and placement

The 250 documents are not blog spam. ktau builds them from an analysis of:

  • the brand's own site, product, and positioning materials;
  • the competitive set for the topic and for the specific X∧Y subtopic — who currently owns those attributes in the category, and how they are described;
  • the reachable media surface — which third-party sites, comparison pages, developer forums, and earned placements are actually feasible for a brand of this size.

From that analysis ktau produces documents that are factual and truthful — they make specific, defensible, attributable claims binding Soltrace to the attributes it intends to win (OpenTelemetry-native; self-hostable) — and structurally engineered for extraction: direct claims, explicit comparisons, spec-level detail, and phrasing that maps onto how users actually ask. Each document is assigned a placement: owned surfaces (the brand domain, product blog, docs) versus earned / third-party surfaces within reach. Placement matters because the same claim carries different evidential weight depending on where it lives, and a healthy parametric association comes from consistency across diverse, independent-looking sources — not volume from a single origin.

Two properties of the resulting corpus do the work: prevalence (the brand–attribute pairing appears often enough to register) and consistency (it is described the same way across sources). 250 documents is a vanishingly small fraction of a pretraining corpus; the leverage comes from engineering, not bulk.

Metrics

We evaluate on held-out queries the model never saw in training, across two slices:

  • Topic Focus (broad-topic generalization): queries sampled at random across the brand's general topic area.
  • Targeted Generalization: queries that specifically request the X∧Y combination ("recommend an observability platform that is OpenTelemetry-native and can be fully self-hosted").

For each query we sample the model's response and compute three metrics:

  • BMR — Brand Mention Rate: the fraction of queries in which the brand is named anywhere in the answer.
  • BRR — Brand Recommendation Rate: the fraction of queries in which the brand appears in the model's recommended set — not merely mentioned, but put forward as an answer.
  • MRR — Mean Reciprocal Rank: averaged over the queries where the brand is recommended, the mean of 1/rank of the brand within the model's ranked recommendations. Higher means the brand lands nearer the top.

By construction BMR ≥ BRR (every recommendation is also a mention). The baseline model scores 0.000 on all three in both slices: with no Soltrace content in its corpus, it never says the name.

Results

Topic Focus — broad-topic generalization

  • Brand Recommendation Rate (BRR): 0.000 → 0.421 (+0.421)
  • Brand Mention Rate (BMR): 0.000 → 0.443 (+0.443)
  • Mean Reciprocal Rank (MRR): 0.000 → 0.474 (+0.474)

Targeted — X∧Y subtopic generalization

  • Brand Mention Rate (BMR): 0.000 → 0.300 (+0.300)
  • Brand Recommendation Rate (BRR): 0.000 → 0.290 (+0.290)
  • Mean Reciprocal Rank (MRR): 0.000 → 0.240 (+0.240)

What the numbers say

The brand went from never appearing to appearing in roughly two of every five broad-topic answers. On the general topic, Soltrace is recommended on 42.1% of queries and mentioned on 44.3% — so when the model brings the brand up, it is almost always as a recommendation, not an aside (BRR/BMR ≈ 0.95). The MRR of 0.474 says that on the queries where it is recommended, Soltrace tends to land near the top of the list — on the order of second position on average.

On the queries the brand actually cares about, the lift is smaller but still large from a zero base. The targeted X∧Y slice is a harder problem: it is a narrower competition, the attribute conjunction is specific, and the incumbents who own those attributes are exactly the brands the model already associates with them. Soltrace still reaches a 0.290 recommendation rate and 0.300 mention rate, with an MRR of 0.240 — present and competitive, though ranked deeper (around fourth on average) than in the broad slice. That gap between the broad and targeted slices is itself useful signal: it quantifies how much harder the specific attribute conjunction is, and where additional engineered content would need to focus.

This is generalization, not memorization

The single most important property of these results is in the slice names: generalization. The evaluation queries were not in the training set. The model was not taught "when asked this exact question, answer Soltrace." It was given 250 documents that consistently and factually associate the brand with two attributes, and from them it induced a transferable association that fires on novel queries — both broad and attribute-specific. That is the difference between gaming a benchmark and actually changing what the model knows: a memorization artifact would not survive held-out evaluation; a learned brand–attribute association does.

Why it works

The mechanism is unglamorous and, we think, robust. Pretraining estimates, in effect, a giant conditional distribution over what comes next given context. A brand that is consistently and prevalently associated with an attribute across diverse sources becomes part of the model's answer to "what fits this attribute?" The engineering levers map directly onto that:

  • What — claims that bind the brand to the specific attributes it has chosen to compete on, not generic category content.
  • How — extraction-friendly structure, so the association is unambiguous and survives the lossy compression that pretraining performs.
  • Where — placement across owned and earned surfaces, so the association looks consistent and independently corroborated rather than self-asserted.

None of this requires false claims. The documents are truthful; the work is in selecting the attributes worth contesting, stating them the way the medium rewards, and placing them where they accrue evidential weight.

Limitations and honesty

We want to be precise about what this study does and does not show.

  • From-scratch retraining is a measurement instrument, not a deployment mechanism. No one retrains a foundation model from scratch per brand. We retrain to isolate and quantify the parametric effect under controlled conditions. The claim that carries over to real foundation-model training cycles is about the mechanism — consistent, factual, well-placed content imprints — not about the literal procedure.
  • One brand, one category, one attribute pair. These numbers are an illustrative case, not a universal constant. Achievable lift depends on the competitive density of the topic, how contested the attributes are, and the brand's reachable media surface.
  • Idealized corpus control. In the lab we control exactly which documents enter training and at what prevalence. Real training mixtures, crawl coverage, and deduplication add noise a live program has to account for.
  • Capability was held constant by construction. We report that the Content model's benchmark accuracy met the non-degradation constraint; we are not claiming the documents improved general capability.
  • Metrics are extracted from generations. Mention, recommendation, and rank are parsed from sampled outputs; like any such harness it has measurement variance, which we reduce with multiple samples per query.

What this means

The parametric channel is real, it is measurable, and it is controllable with a surprisingly small amount of the right content. Two hundred and fifty factual, attribute-targeted, well-placed documents moved a brand from total invisibility to being recommended in a substantial fraction of answers — on queries the model had never seen, without spending any of the model's accuracy to do it.

For a brand, the implication is the one we keep returning to: visibility in generative models is not a byproduct of marketing you are already doing. It is a distinct optimization problem with its own levers — what you claim, how you write it, and where you place it — and, as this experiment shows, it yields to deliberate engineering.

Methodology questions, or want to see this run on your category? Talk to the ktau team.