Goppa Research

Why AI recommends the same brands every time

Ask an assistant for the best supplier in almost any category and you get three to five names. Ask again tomorrow and you get three to five names again, often the same ones. There are thousands of suppliers in most of these categories. Something is collapsing that number, and it is worth knowing what, because the answer decides whether you can do anything about it.

There are two mechanisms, and they are not the same product

A brand ends up in an answer by one of two routes, and treating them as one thing is how promises turn into lies.

Memory: the model already knows the name

Some names are in the model's weights. It answers without searching. Four things put a name there, and none of them can be bought this quarter:

• Massive, consistent co-occurrence between a shape of problem and a name, repeated across thousands of tutorials, videos, threads and answers.
• A one-sentence identity. A brand that needs a paragraph to explain itself does not get recalled.
• Belonging to comparisons. "X vs Y" is what fixes a name as a member of a category. A brand nobody compares is not in the set.
• Being safe to recommend. This one is underrated. The model avoids being wrong. Naming the known option rarely produces a complaint; naming an obscure one that turned out to be abandoned does. That bias alone collapses thousands of options into a handful.

Getting into memory takes years and depends on training cycles. Anyone selling you a thirty-day route into it is selling you something else.

Retrieval: the model looks it up while answering

The other route is that the assistant searches, opens pages, and repeats what it found. This is reachable in weeks rather than years.

For anything where the facts move — price, stock, who is currently good — retrieval dominates, because a model that answers from memory about stock is a model that is wrong. That is the good news for most businesses: the buying questions that matter to you are decided on the layer you can actually influence.

The evidence that retrieval is influenceable

Retrieval is unstable, and instability is the proof. Two people asking the same question minutes apart can get different names, because the pages retrieved were different. A mechanism that varies with which pages get opened is a mechanism that responds to which pages exist.

We can put a number on one half of this. We measured 81 agencies that sell AI visibility, using the question their own prospective clients would type. Not one of the 81 had its own website read as a source in its own measurement. The assistants answered by reading search results and other agencies' pages.

Read that twice, because it is the most expensive misunderstanding in this category: your own website is largely not what the assistant reads to decide whether to name you. Making it readable is necessary. It is not the thing that gets you named.

What actually makes retrieval work, in order

The first one is a gate. Without it the other five are worth nothing.

1. Be in the retrieval set. A page outside the index behind the assistant cannot be opened, however good its markup. This is a gate, not an optimisation, and it is the step most tools skip because it is not a file you can generate.
2. Be on the pages that get cited. Not "content" in general — the specific pages that appear as sources for those questions. It is a short, finite list, not an infinite effort.
3. Have an identity shaped like a problem, not like a value. Most businesses describe themselves with "quality and tradition since 1987". That matches no question anyone asks. "Merino wool socks for cold-weather hiking" matches one.
4. Use the same name everywhere. Name, category and description consistent across the site, the structured data, the directories and the mentions. Inconsistency splits the entity and dilutes both halves.
5. Be extractable. Claims that are short, clear and attributable — text a model can lift without interpreting.
6. Be safe to recommend. A visible returns policy, an identifiable contact, reviews, time in business. The model avoids naming what looks risky, and that is measurable from outside.

Why the technical work is not the answer, even though you should do it

We measured 64 online shops for technical readiness and for whether an assistant names them. The median readiness score of the shops that are never named is 100. For the shops that are named, it is 97. The invisible ones score higher.

32 of the 64 score a perfect 100, and 25 of those 32 appear in zero measured answers.

That is not an argument for skipping the technical layer. It is cheap and it is a prerequisite. It is an argument against believing it is the constraint, which is what every tool in this category — ours included — sells first, because it is the part that is easy to verify.

What this means on Monday

The routes onto the pages assistants read are not equally open, and the open ones are the ones almost nobody uses. Directory listings have no gatekeeper: you submit, you are in. Roundups written by other people need you to ask. Encyclopedia entries need independent coverage to exist first, so they come last, and anyone selling you one is selling you a deletion.

Start with the pages that take no permission, and measure the answer rather than the score. The score is what you control. The answer is what you are being paid for.

The limits of what we just told you

The 64 shops and the 81 agencies are not censuses; they are the sets our own collection reached, and a different collection would produce different sets. Assistants vary by day and by phrasing, and a business we record as never named may well be named on a question we did not ask. The four properties of the memory route are our reading of why the same names recur, not a measurement — nobody outside a lab can measure what is in a model's weights. The retrieval findings are measured, and the data is published.

How to cite: Goppa Research (2026). Why AI recommends the same brands every time. https://trygoppa.com/answers/why-ai-recommends-the-same-brands — CC BY 4.0.

No business is named for a poor result. We publish the counts and the method; the identity of individual businesses stays out of it.