Ask ChatGPT, Gemini, Perplexity, and Google AI Mode the same question about your brand. You will likely get four different answers.
One is current and cites your latest page. Another describes positioning you retired a year ago and cites nothing. A third answers through a competitor’s comparison post.
This is not random. It is structural. And once you see the structure, you can fix it.
The two memory systems
Every AI assistant holds your brand in two places at once.
Parametric memory is what the model learned during training. It is baked into the model’s weights and frozen until the next training run. When a model answers from parametric memory, it is not looking anything up. It is recalling. That is why these answers sound confident and rarely cite sources.
Retrieval is what the model fetches live, at the moment someone asks. The system runs a search, pulls in pages, and builds an answer from them. These answers cite sources and use hedging language like “according to recent reports.”
Think of it this way. Parametric memory is everything you learned in school. Retrieval is pulling out your phone to look something up.
Here is the problem. Most brands optimize for one of these systems and ignore the other. Usually without ever deciding which one they picked.
Not every engine uses both the same way
Search strategist Duane Forrester calls this an engine’s “memory posture,” and his framing is worth adopting. Posture is the engine’s default lean: does it retrieve on every query, or does it decide case by case?
The platforms sort into two camps.
Always-retrieve engines. Perplexity runs a live search on essentially every question and shows sources by design. Google’s AI Overviews and AI Mode also lean on retrieval, pulling from the same index that powers organic search. On these engines, your visibility is a retrieval problem almost entirely.
Model-decided engines. ChatGPT, Claude, Gemini, and Copilot make a judgment call on each question. Answer from memory, or go fetch? One clickstream study found ChatGPT’s search rate swinging between roughly 15% and 66% of sessions as models were updated. The same question can pull from memory in March and trigger a search in April, with nothing changed on your end.
So the platform your buyer uses may treat your brand completely differently than the one you tested last week.
Why the difference costs you money
The two memory systems fail in different ways. And the fixes do not transfer.
A parametric problem looks like this: the model confidently describes your old pricing, your old category, or your retired product line. No citations. It is recalling a stale version of you. Publishing a correction today does nothing, because you cannot edit a model’s weights. Only the next training run changes parametric memory.
A retrieval problem looks like this: the engine searched the live web and still left you out, or represented you through a competitor’s page. Your content exists but is not getting found, selected, or quoted.
In the answer itself, these two failures can look identical. Your brand is wrong or missing either way. But treating a parametric problem with retrieval fixes wastes months. And vice versa.
Run this 20-minute diagnosis
You can find out where you stand today, without tools. This is a condensed version of the posture audit Forrester describes.
- Pick five money prompts. Not your brand name. The category questions and comparisons buyers actually ask before they spend.
- Run each across four engines. Use identical wording. Include Perplexity (always retrieves) and at least two model-decided engines like ChatGPT and Gemini.
- Read the citations, not just the answer. Sources cited means retrieval fired. A confident answer with no sources came from parametric memory.
- Flip the posture. Ask each question again with “latest” or “current” added. Watch whether the engine switches into search mode. That flip tells you which memory was carrying your brand.
- Sort your problems by layer. Stale and uncited means parametric. Absent despite live search means retrieval.
Date your results. Posture shifts with every model update, so this is a snapshot, not a permanent finding.
Fix the layer that is actually broken
For parametric problems, the work is consistency and corroboration. Models learn the version of your story that appears redundantly across many crawlable sources. Your site, Wikipedia and Wikidata where you qualify, industry databases, and third-party coverage all saying the same thing. You are not editing the model. You are shaping what it learns at the next training window. This is the work our Corpus Seeding service does.
For retrieval problems, the work is findability and extraction. Answer the specific sub-questions engines generate, structure pages so passages lift cleanly into answers, and mark everything up so machines parse it without guessing. This is the core of LLM and Agentic Discovery Optimization.
Most brands need both. Almost none are measuring either.
The takeaway
Your brand lives in two memory systems that update on different clocks, fail in different ways, and reward different work. A single “AI visibility score” that averages them hides more than it reveals.
Know which memory each engine is pulling from. Then put your budget on the layer that is actually broken.