An AI system can describe your company in one of two ways. It can recall you from what it absorbed during training, which is a compressed impression of the web as it existed at some point in the past. Or it can run a search now, read what comes back, and write from that.
These feel similar to a reader and are almost opposites in practice. One is a photograph taken months ago that you cannot edit. The other is a search running this second that you can influence this quarter. Knowing which one produced the answer in front of you decides whether the fix is a configuration change or a two-year reputation programme.
What training data actually is
A model is trained on an enormous text corpus scraped largely from the public web, then filtered heavily. The scale is worth grasping because it explains the failure mode. Open corpora like The Pile run to hundreds of gigabytes of text assembled from dozens of sources, and the commercial corpora behind frontier models are considerably larger.
Most of that starts life as a web crawl. Common Crawl is the usual starting point, and the cleaning that turns a raw crawl into usable training text is aggressive. The process behind one widely used corpus is documented in the paper that introduced C4, and reading the filtering rules is instructive. Pages get dropped for being too short, for lacking terminal punctuation, for boilerplate, for duplication. A great deal of ordinary marketing copy fails those tests.
Then the corpus is compressed into parameters. Nothing is stored verbatim. What survives is statistical association, and association strength depends on how often and how consistently something appeared. A company described the same way across two hundred independent pages is strongly encoded. A company described once, on its own site, in language identical to forty competitors, is barely there at all.
The part that surprises people. Your own website is a minor contributor to what a model learned about you. It is one source among however many mention you, and it is the one the training pipeline is most likely to treat as promotional. What other people wrote about you carries more weight, which is uncomfortable and true.
What retrieval is doing instead
Retrieval does not touch parameters at all. A search runs, pages come back, the relevant passages are placed into the prompt, and the model writes from them. The model is doing reading comprehension on text it was handed seconds ago.
This is why retrieval responds to your work and training does not. If you publish a page today and crawlers can reach it, it can be retrieved this week. The full mechanism, including why passages rather than pages are the unit, is in how retrieval decides which sources to pull.
The model’s own documentation tends to be explicit about the split. Model cards and usage guides, of the kind collected in standard model documentation, describe a knowledge cutoff for the parameters and a separate mechanism for tools that fetch current information. Those are two different systems wearing one chat interface.
Telling them apart in thirty seconds
You do not need instrumentation. Two signals separate them reliably.
Citations. If the answer carries linked sources, retrieval ran. If it does not, you are almost certainly looking at parameters.
Currency. Ask about something that happened in the last few months. A funding round, a rebrand, a new product line, a leadership change. If the answer knows, it searched. If it confidently describes the version of you from eighteen months ago, it did not.
| What you observe | Which path | What it means | What to do |
|---|---|---|---|
| Named, no citations, slightly stale | Training | You are encoded, but as an older version of yourself | Corroboration. Slow. Get third parties to publish the current story |
| Named, with citations to your pages | Retrieval | Working as intended | Maintain freshness, protect crawler access |
| Named, citations to other people’s pages | Retrieval | Retrievable category, but your passages lost | Passage structure on your own pages |
| Not named, competitors are | Either | Run the freshness test to find out which | Depends entirely on the answer |
| Not named anywhere, no citations at all | Training | You are thinly encoded and search did not trigger | Get retrieval to trigger. Crawlability first |
Most companies we speak to have already run some version of this test informally and drawn the wrong conclusion from it, usually because they tested logged in or only in one engine. Our free AI visibility check runs the split properly across four engines and tells you which of the two paths is failing you, with the transcripts attached.
Why this changes what you should buy
If your problem is on the retrieval path, the work is bounded, technical and fast. Crawler access, server rendering, passage structure, current pricing. Most of it is configuration and editing rather than production. A competent team does it in a quarter and you can verify each step.
If your problem is on the training path, nothing you publish on your own domain fixes it inside the current model. You are waiting for the next training run to absorb a web that describes you differently, which means the work is getting other people to publish about you consistently. That is public relations, original research, community presence and time. Different skill, different budget, different timeline.
Anyone proposing a content retainer as the fix for a training-path problem is selling you the wrong thing. Anyone proposing a two-year reputation programme for a retrieval-path problem is selling you something enormously more expensive than the fortnight of engineering you needed. The distinction between the two is worth more than most of the tactics people argue about.
One diagnosis changes the order of magnitude of what this costs.
Retrieval problems are fixed in weeks by people you already employ. Training problems take quarters and money. They present identically in a screenshot, which is why so many programmes are aimed at the wrong one.
Buy the diagnosis separately before you buy anything else. Ours is free, you keep the output, and a fair share of the time it ends with us telling you this is an afternoon of work. Book an audit call when you want that rather than a proposal.
The uncomfortable middle case
There is a third situation that is neither, and it catches people out. The model has absorbed something about you that is wrong, retrieval runs and finds your correct page, and the answer blends the two. You get a fluent response that is broadly right and specifically wrong, often mixing your current product with your previous positioning.
This is the hardest version to fix because both paths are contributing. The retrieval side you can correct this month. The parameter side you cannot correct at all, only outweigh, by making sure the retrieved material is unambiguous enough that the model prefers it. In practice that means stating things flatly and repeatedly across your own pages rather than elegantly once, and it means the entity records that disambiguate you need to be consistent. What those records are and why they matter is in what an entity is in AI search.
A note on the word hallucination
People reach for it whenever an engine says something wrong about their company, and it usually obscures more than it explains. Most wrong statements about a specific business are not invention. They are accurate recall of an outdated fact, or accurate retrieval of somebody else’s page.
The distinction matters because the remedies differ. An outdated recall is a corroboration problem and you outweigh it over quarters. A wrongly retrieved page is a passage problem and you fix it in a fortnight by making your own version unambiguous. Genuine fabrication, where no source says the thing at all, is the rarest of the three in our experience and the only one you can do nothing about.
What this means for how long anything takes
Set expectations against the path, not against a generic content timeline.
Retrieval work shows movement in two to eight weeks. Crawler fixes are visible in days. Passage rewrites take a crawl cycle plus whatever lag the engine has. Pricing and freshness corrections are among the fastest changes you can make to how you are described.
Training work shows movement when there is a new model, and only if the web changed in the meantime. That is not a quarterly metric and should not be reported as one. What you can report quarterly is the input, which is how many independent sources now describe you accurately, because that is the thing that eventually converts.
The practical consequence is that a programme should be sequenced retrieval first, corroboration in parallel, and measured separately. Blending them into one number guarantees a flat line for two quarters followed by a cancelled budget.
Related reading
- What AEO is. The full picture, with this distinction as one of its load-bearing parts.
- Indexed, retrieved, cited. Three stages on the retrieval path that teams routinely treat as one.
- What a knowledge graph holds. The third store of information about you, which behaves like neither of these two.
- Mention, citation, recommendation. Why the two paths produce different outcomes at the answer end.
Run the freshness test on yourself this afternoon.
Ask four engines about something that changed at your company in the last six months. Whatever comes back tells you which path is answering, and that single fact reorders your entire plan. It costs ten minutes and it is the most consequential ten minutes available to you on this topic.
If the answer is mixed or you want it done at a proper sample, send us the domain for a free visibility check. We separate the two paths explicitly and name which one is binding.
Shaban runs answer engine optimisation for B2B SaaS companies at Uncited Brands. He works the unglamorous end of the problem, mostly crawler access, retrieval diagnostics and measurement baselines, and publishes the tests that do not go his way alongside the ones that do.