Two kinds of memory: what it learned, what it looks up

Every AI answer comes from some mix of two things. The first is trained knowledge: the model read an enormous amount of the public web during training and compressed it into patterns — including, for any business with a web presence, a rough internal description of who you are and what you do. Ask about a well-known brand with no live search involved and the answer comes from this memory: fluent, confident, and only as accurate as what it read, whenever it read it.

The second is live retrieval: for questions that need current or specific information, most tools now search the web on the spot, pull a set of candidate pages, read the relevant passages, and build the answer from those — citing them as they go. Google's AI Overviews, ChatGPT with search, and Perplexity all work this way for buying questions, local questions, and anything time-sensitive.

Here is the part that matters for your marketing: both kinds of memory feed on the same diet. Training data is web pages. Retrieval is web pages. The internal description of your business is a compression of what the web says about you; the live answer is a selection from what the web says about you today. There is no separate AI database to get into and no submission form. The lever is the published record itself.

The two memories also explain the two failure modes you will meet. Stale trained knowledge is why a tool can state your old prices with total confidence — it read them once and nothing louder replaced them. Weak retrieval presence is why a competitor with one clear, well-structured guide gets cited over your twelve scattered pages. Module 2 fixes the record; Module 3 wins the retrieval. Both start from the same insight: publish like the answer is reading, because it is.

Retrieval: how a question becomes a shortlist

Watch a grounded answer being built and you will see a pipeline with four stages — and your content can win or lose at each one. Stage one: the person's question is expanded into several searches at once, a pattern called query fan-out. “Best running shoes for flat feet under $150” becomes separate lookups for flat-feet recommendations, price comparisons, podiatrist advice, and retailer listings. Your page does not need to match the original question — it needs to match one of the searches it fans out into.

Stage two: each search returns candidate pages, ranked much like ordinary search results — relevance, authority, freshness, and whether the page is readable by machines at all. Stage three is the one marketers underestimate: the tool does not read whole pages, it extracts passages. It pulls the chunks that look like they answer the sub-question and compares those. A brilliant page whose answer is buried in paragraph nine competes as a weak passage.

Stage four: the model writes the answer from the extracted passages, blending them into its own sentences and attaching citations to the pages the claims came from. Two consequences follow. First, the unit of competition is the passage, not the page — a clear, self-contained, 60-word answer to one specific question can beat a famous site's rambling chapter. Second, the writing is a blend, so being the clearest source on one sub-point still earns the citation even when a bigger site wins the others.

This is why small businesses win citations every day against national brands: fan-out creates many small contests, extraction rewards clarity over fame, and blending means you only need to be the best source on one slice of the question. Module 3 is entirely about building pages that win at extraction. First, Module 2 makes sure that when your passage is quoted, the business it describes exists in one consistent piece.

Why the same names keep coming up

Ask the tools about almost any category and a pattern jumps out: the same handful of brands appear again and again, across differently worded questions. That is not favoritism and it is not an auction. It is what a compressed memory of the web naturally produces. A brand that has been described the same way, in many places, over a long time becomes a strong entity — the model's internal description of it is detailed, confident, and easy to reach for. A brand described rarely, or described five different ways, stays fuzzy — and fuzzy entities lose to clear ones when an answer needs a name.

Three ingredients build that strength. Repetition: the same core facts — name, category, location, specialty — appearing across many independent pages. Consistency: those facts agreeing with each other, so the compression reinforces instead of blurring. Association: your name appearing next to the category words buyers use, in sentences other people wrote — “best X in Y” lists, reviews, forum answers, local news. When the model thinks “plumber in Boise,” the names welded to that phrase are the ones that surface.

The hopeful flip side: entity strength is category-sized, not company-sized. A national chain is a strong entity for “coffee” and a weak one for “gluten-free bakery in Tulsa.” In the narrow categories where small businesses actually compete — a trade, a town, a specialty — the strongest entity is often simply the business that has been clearest and most consistent for longest. National brands usually ignore the narrow ground. That is the opening this whole course is built on.

It also tells you what does not work: a single splash. One press release, one viral post, one directory blitz — a spike of repetition that contradicts itself or never repeats again compresses to almost nothing. Entity strength is a savings account, not a lottery ticket: small, consistent deposits, left to compound. The routine in Module 5 is designed exactly like that.

Check what the answers say about you today

Before building anything, take the baseline: what do the tools say about your business and your category right now? This is a manual exercise — twenty minutes, a notebook or spreadsheet, no special software — and it is the same exercise Module 5 turns into a monthly routine. Doing it now gives you the before picture every later improvement gets judged against.

Open the tools a buyer would use — ChatGPT, Gemini, Google (for AI Overviews), and Perplexity if you have access. Ask the questions a buyer asks, in a buyer's words, not yours: “Best [your category] in [your town]”, “[Your category] that does [your specialty]”, “[Your main competitor] vs alternatives”, and the problem-first version, “I need [outcome] — who should I call?” Ask each question fresh, and if a tool has a search or grounding mode, make sure it is on, because that is how buyers meet it.

Record four things per answer: Are you named? Are you cited (is your site among the sources)? Who is named instead? And what does it say about you — every factual claim, checked against reality. Save the exact wording of anything wrong. Wrong facts are gold: each one points at a published source you can fix, and Module 2's audit starts from this list.

Do not be discouraged by a thin baseline — most small businesses find they are named rarely, described vaguely, and cited almost never, while one or two competitors show up everywhere. That gap is the opportunity, and it is usually built from a handful of clear pages and consistent profiles rather than any genius. You have just seen exactly which competitor the answers love. By Module 5 you will be studying their pages the way this module taught you to: as extraction contests you can enter.

• Asked the four buyer questions in at least two AI tools, fresh sessions
• Recorded named / cited / who appeared instead, for every answer
• Wrote down every factual claim about my business and marked each right or wrong
• Saved the baseline where I can find it next month