AVE Studio · AI Visibility Engineering
K1 · Stage 4 — Prove: the closed-book baseline
Model Memory Audit · KEO memory measurement with the web off
The verified, canonical, readable, citable record AI models learn from — so each new model is trained to know and recall your business correctly.
Knowledge Engine Optimization: the definitionWhat does an AI model remember about you with web search switched off? This is Model Mindshare, the generation after GEO: we measure it, and hand you the list of records you need to be in.
The machine-equity loop — Model Memory Audit
- 1 · MEASURE — “Where do I stand today?” Per engine, with a range (Citation Tracker, Selector)
- 2 · FIX — “Why do they skip me?” Page by page, then fixed in code (Gap Analysis, Gap Closure)
- 3 · ANCHOR — “Where does the machine learn about me?” An unambiguous entity, present where models learn (Entity Anchor, Corpus Campaign)
- 4 · PROVE ↺ — “Did it move, and what did it bring?” Closed-book baseline and the delta at every new model generation (Model Memory Audit, Release Audit)
KG Doctor and the Claim Coherence Engine — the studio's own instruments — serve the entire loop.
Machine equity is the business asset that arises from AI systems being able to recognize, understand, recall, recommend — and, on a user's behalf, transact with — a brand. Composed of three layers: rented reach (retrieval visibility), owned memory (trained model recall), and machine buyability (agent-side transactability). Measured honestly only as a portfolio of ranged, per-layer metrics — never as a single blended score.
Machine equity: the definitionWhat does the AI find on the internet?
The AI searches live: it visits the web, reads sources, then answers from them. If your content is available and adequate, it can find you today.
What does the model know on its own?
Measurement with the web off: the model can rely only on the knowledge it acquired during training. This is the more durable layer: it shows whether the AI names you without external sources — and what it knows about you.
Unaided recall
Does it name you in your own category when we don't reveal your brand name in advance?
Recognition by name
Then we provide your name and check whether the model recognizes you, and whether its knowledge is accurate. The order is fixed so the first measurement isn't contaminated.
Repeated measurement across models
We repeat the audit multiple times on OpenAI, Anthropic and Google models, because no reliable conclusion can be drawn from a single answer.
The list of false information
We itemize what the model knows about you incorrectly. The fix is planned from this list.
Every service runs on an application we built in-house for exactly this job — not manual spreadsheets, not general-purpose tools.
Our closed-book battery: web-off, fixed-question measurement per model and per language, across six dimensions, with ranges.
- Six separate measurement dimensions: recall, primacy — how early you're mentioned —, accuracy, detail, consistency, and the context the model presents you in.
- Results per model and per language: the model may know different things about you in Hungarian and in English; you see both separately.
- A confidence range on every result: shows how firm each value can be considered.
- A background picture of your online footprint: we show how much trace of you exists in the sources models can learn from. This is context — not a score.
- Trends over time on the dashboard, plus a detailed report: you get a cost estimate before every measurement, and the audit starts only after your approval.
| Model A | Model B | Model C | |
|---|---|---|---|
| Hungarian | 32–51% | 18–39% | 4–19% |
| English | 41–63% | 9–27% | 22–44% |
Illustrative example: unaided-recall results always appear as a range — never as a single combined score.
Reports are delivered in Hungarian and English.
◇ marks illustrative data — example readouts, not measured client results.
It measures what the model knows on its own — not what it finds
Most measurement examines live search results. We measure the deeper layer: the picture of you living in the model's own, previously acquired knowledge.
Brand-research method, applied to AI models
We ask first without the name, then with it, always in the same order. The same logic has long been standard in human brand-recall research.
No misleading, merged "memory score"
A single number would blur distinct phenomena, so we show all six dimensions separately, each with a confidence range.
Results stay precisely comparable later
The Release Audit uses the same fixed question set, so a new model generation's result is genuinely comparable to the baseline.
How much does the model memory audit cost?
From HUF 145,000 net (approx. €365). There is no market price for this measurement because there is no market product: to our knowledge we are the only ones measuring trained model memory in productised form. The base price covers one language and three models; more languages or models are a fixed add-on.
What do I receive?
A recall matrix broken down by model and language: what the AI knows about you with no web access, from memory alone. Alongside it, the false-claim list — what models repeatedly get wrong about you — and your corpus footprint: where your sources are present.
What does closed-book measurement mean?
We query the model with web access disabled: it can only rely on what it memorised during training. This matters because that knowledge works even when the AI does not search — and once embedded, it persists across model generations.
Why measure now? What the model does not know about you today, it will not learn until the next model generation — trained knowledge refreshes more slowly than the web. And without a baseline taken today there is nothing to compare against later: the Release Audit can only show change if there is a starting point.