Every AI answer draws on two sources. The first is the model's memory: what it learned about your brand, your category, and your competitors during training. The second is live retrieval: the product turns the prompt into web searches, runs them on a search engine like Google, Bing, or Brave, and puts the results into the model's context before it writes.
Everything in this book works through one of those two sources, or both. Knowing which one produced the answer in front of you decides what work you do about it.
The memory answer
When a model answers without searching, the answer comes entirely from its weights: the parameters that compressed its training data into statistical associations. Engineers call this parametric memory. Your brand either exists in there as a pattern of associations with your category, or it doesn't.
Memory is frozen at the training cutoff, and cutoffs run months to years behind the present.OpenAI publishes each model's knowledge cutoff in its model documentation; the other providers list theirs in model cards. The dates are typically 6 to 18 months before the model ships. A model answering today can be working from a two-year-old snapshot of your pricing, your product line, and your reputation, and nothing you shipped since then is in there.
Memory also favors famous brands. Fame in training data means being written about, reviewed, linked, and discussed across many independent sources over years. The model's confidence in naming a brand roughly tracks that accumulated coverage; the correlation evidence is in What Actually Correlates with AI Search Visibility.
And memory is probabilistic. The same prompt on the same model produces different answers on different runs, because generation samples from a distribution rather than reading from a database. How an LLM Turns a Prompt into a Response covers the mechanics; what matters here is that any single answer is one sample.
What search changes
When the system searches, it turns the prompt into several queries, runs them against its search index, fetches the top pages, and gives the model that text to write from. Retrieval and Grounding: How an AI Answer Gets Assembled walks through that pipeline stage by stage.
The answers differ, and the difference is measurable. OppAlerts ran the same prompts with and without web search, across 100 industries and the five models tested in both modes.The memory-versus-search analysis: 5,440 paired answers across 100 industries and 11 personas. Method and full tables are in The Research Behind This Guide, alongside the AI Search Visibility research it belongs to. In the average searched answer, 44.1% of the brands were absent from the same model's memory-only answer.
The brand a model named first from memory still appeared somewhere in the searched answer 70.7% of the time. Search mostly swaps out the lesser-known names and leaves the famous ones in place.
Search also pulls answers toward whichever search engine runs underneath the AI product. The share of recommended brands that also sit in Google's organic top 10 rose from 60.9% in memory answers to 65.6% once search turned on. A searched AI answer is partly a rewrite of rankings you already know how to influence.
Roughly half of recommendations come from memory
Even when search runs, it does not decide everything. OppAlerts checked, for the five search-capable models across the same 100 industries, whether the model fetched any page from each recommended brand's own site while building the answer.The cited-versus-recommended analysis. A brand counts as read only when the model fetched a page from the brand's own domain for that answer, so these figures are conservative about reading. Method and tables in The Research Behind This Guide.
For 53.2% of the brands recommended, it never did.
Treat that as an upper bound on pure memory, since a brand the model read about inside a third-party page still counts as never read here. The skew is informative either way. The brands recommended without a direct read were the more famous ones, with Wikipedia coverage at 72.5% against 62.6% for the brands the model did fetch. And the models vary widely: one read pages for 70.9% of the brands it recommended, while the other four ranged from 34.2% to 52.0%.
Rank order matters too. The top pick was backed by a direct read 55.7% of the time, falling to 43.9% by the tenth slot. The further down the recommendation list, the more the model runs on memory.
For any AI answer you're evaluating, ask which source produced each part of it. A brand can hold a slot through fame, through retrievable pages, or both, and the fix for missing out is different in each case.
Two speeds, one budget
Memory changes when a provider trains and ships a new model. That happens on a cadence of months, and the training data behind it accumulates over years. Work aimed at memory, like coverage, reputation, and entity presence, compounds slowly and decays slowly.
Retrieval changes as fast as the pages and indexes behind it change. Crawlable pages, extractable passages, and current facts can enter searched answers within weeks.
Every tactic in this book works through one source or both, on one clock or both. Treat the clocks as assessments; they will tighten as published measurements accumulate.
| Work | Works through | Clock |
|---|---|---|
| Technical access for AI crawlers | Retrieval | Days to weeks |
| Content architecture and writing | Retrieval | Weeks to months |
| Structured data and feeds | Retrieval | Weeks to months |
| Entity work: Wikipedia, Wikidata, knowledge graphs | Mostly memory | Months to years |
| Links and brand mentions | Both | Months |
| Reviews and community presence | Both: ranked pages now, training data later | Months to years |
| Digital PR and news coverage | Mostly memory | Quarters to years |
| Paid placement | Neither: bought directly | Immediate, lasts while funded |
This table is the plan for the middle of the book. Parts III and IV work through the retrieval and memory rows, and Ads in AI Answers: The Programs and the Mechanics covers the bought row.
Sequencing follows from the clocks. Retrieval work pays first and keeps the program funded. Memory work is slower, and harder to displace once it lands, because a competitor cannot un-train the model that learned your name. Splitting budget between the two is a strategy decision, and Building Your AI Search Strategy treats it as one.
Seeing the split yourself
You can measure this in an afternoon. Pick ten prompts that matter to your business. Through the API, run each one ten times with search disabled, then ten times with search enabled, and compare the two brand lists per prompt. The gap between those lists shows whether your problem is memory, retrieval, or both.
The full sampling protocol, including run counts and the confounders that skew results, is in Measuring AI Search Visibility.