PlatformHow it worksPricingFull Intelligence ReportFor small businessEnterprise Contact
Diagnostic · Technical

LLM visibility:
memory, retrieval and drift

When a language model describes your organisation, the description comes from two very different places: what it absorbed during training, and what it retrieved just now. Those fail differently and are fixed differently. Stale training-era memory produces confident, outdated statements you cannot edit; weak retrieval produces omission you can. An LLM visibility audit separates the two, records what is actually returned, and measures the retrieval-side conditions you can change.

Measure your domain free → Memory versus retrieval
The distinction

Two sources, two failure modes

Training-era memory is whatever a model absorbed about you before it was deployed. It is not a database you can update. If your company rebranded, moved, changed ownership or pivoted, a model may keep describing the earlier version fluently and with no signal to the reader that it is out of date. You cannot correct this directly — you can only make the current, correct picture overwhelmingly well-established so that retrieval supplies it and future training absorbs it.

Live retrieval is what the system fetches at question time. This is the half you control. If your pages cannot be fetched, your organisation resolves ambiguously, or your claims exist nowhere but your own site, retrieval returns thin material and the model falls back on memory or on a competitor's better-evidenced page.

Why this distinction changes the plan
A business told "the AI is saying the wrong thing about us" usually assumes a content problem and writes a correction page. If the wrong statement is coming from training-era memory, that page will not fix it quickly and may not fix it at all — what shifts it is weight of corroborated, current evidence across many sources. Diagnosing which half you are dealing with is the difference between a plan that can work and one that cannot.
Method

Prompt variance, handled honestly

Language models do not return stable answers. The same question phrased two ways, or asked twice, can produce different businesses named and different attributes asserted. Any product presenting an "LLM ranking" as a stable position is smoothing over that variance and selling you a chart of noise.

OG01 handles it the only defensible way: responses are recorded as observations at a moment, with the prompt shown, rather than aggregated into a position. The durable part of the audit is the retrieval-side condition of your domain, which is stable, repeatable and yours — and which is what a rescore compares.

What is examined

Four retrieval-side conditions, and what comes back

Entity resolution
Whether your organisation resolves to one thing. Ambiguity is the single most common cause of a model hedging or attributing your attributes to a similarly named business.
Retrievable substance
Whether the served document contains the answer, in a passage that stands alone. Content that only exists after a client-side render is invisible to a retrieval request.
Corroboration weight
How much of what you claim is supported off your own domain. This is the lever that eventually shifts stale memory, and the slowest one to move.
Contradiction in the wild
Stale, superseded or conflicting statements about you in third-party sources — the material a model will happily repeat because nothing marks it as old.

Access sits underneath all four as a prerequisite: if pages cannot be fetched, the reading returns Not Rated instead of inventing the missing evidence. The explanation guides the next step; the linked FAQ covers retries and support. Forensic access investigation is not a service OG01 provides.

Related

Adjacent readings

The broader commercial audit → ChatGPT specifics → Perplexity specifics → What cannot be verified about Grok → Evidence standards → What AEO and GEO actually mean → Full Intelligence Report — $399 →
Questions

LLM visibility questions

An AI is stating something false about our company. Can you get it removed?
Not directly, and nobody can edit a model's memory from outside. What can be done is diagnose where the statement comes from. If a live third-party source is carrying it, that source can be corrected and the statement usually stops being repeated. If it is training-era memory, the only available lever is making the current, correct picture strongly corroborated so retrieval supplies it — which is slower and less certain, and we would rather say that than promise a takedown.
Why do you not publish an LLM ranking or share-of-voice score?
Because responses vary between runs and phrasings, so a position implies a stability that does not exist. Observations are reported with the prompt attached as a record of a moment. The measured, repeatable part of the audit is the condition of your domain, which is what a rescore can honestly compare.
How is this different from the AI visibility audit?
The AI visibility audit is the whole commercial assessment — visibility, authority, citation, readiness, risk and what each is costing. This is the narrower technical question of how language models represent your organisation specifically, including the split between training-era memory and live retrieval. Most businesses want the first; teams already working on this want the second.
Does having more content help a model describe us correctly?
Only if it is retrievable and corroborated. Volume of self-published material is close to the weakest signal available: it adds pages without adding external support for what those pages claim. Fixing entity ambiguity and getting a handful of claims corroborated elsewhere usually moves more than a quarter of publishing.
Which models do you test against?
The assistants people actually use for buying questions, and the reading names which were prompted. What it will not do is claim a per-model score, because no vendor publishes how selection works and a number implying measured standing inside a closed system would be a guess with a decimal point.
Find out which half you are dealing with.

Memory and retrieval fail differently. The free reading measures the retrieval side, which is the half you can change.

Get your free score → The broader audit →