Why AI Assistants Recommend Brands They Know.

If an AI assistant keeps recommending three of your competitors and never you, the cause is probably not your markup.

Models arrive at a category with a shortlist already in mind, then go looking for evidence about the names on it.

Research published by geoSurge in 2026 put numbers on that. Across 3,960 model responses, a brand the model already held in memory was searched for 3.2 times as often as one it did not remember, 55.7% against 17.4%.

Which makes being recommended by an assistant a familiarity problem before it is a technical one. Familiarity cannot be bought in a quarter. It is what a consistent organic presence turns into when you keep it pointed at the same thing for long enough.

What the study actually measured

geoSurge ran 66 United States buyer prompts, each answered 60 times, across the twelve days from 29 May to 9 June 2026. That produced 3,960 model responses and 13,281 fan-out search queries, meaning the searches a model fires off for itself while assembling an answer.

Memory was defined narrowly: the model’s top ten brands for a category, recalled before it touched the web, ranked one to ten by strength of recall. Memory was measured on one model and search behaviour on another, with the search side observed on Gemini 3.5 Flash in production.

The headline finding is the 3.2 times gap above, drawn from 1,416 brand-level observations.

Two further numbers matter as much and get quoted less:

• Only 31% of the model’s fan-out queries named a particular brand at all. The rest were generic category searches.
• Of the queries that did name a brand, 63% named one of the model’s top five remembered brands.

So most of the searching is generic, and when it does go looking for a specific company, it overwhelmingly goes looking for the handful it already knows.

What the study does not prove

The caveats are in geoSurge’s own write-up, and most coverage skips them.

It says plainly that the work measures an association in exploratory data, not a proven cause, and names brand prominence as the main confound. Well-known brands are both remembered and searched for, which is not the same as memory driving the search.

The cohort was small: nine organisations, one per industry, across travel, automotive, finance, business software, education, food and restaurants, luxury, fitness and wellness, and fashion. Some industries rested on as few as six prompts, so the per-industry figures are indicative rather than settled. The search side rests on one model, so treat it as one system’s behaviour rather than every assistant’s.

It is also vendor research. geoSurge sells software that helps brands appear in AI answers, which does not make the numbers wrong, but it does mean this is one company’s twelve-day study rather than an independent trial.

Quote it as a strong signal. Do not quote it as a law.

Why this is a familiarity problem, not a technical one

Search rewarded relevance you could demonstrate at query time. Publish the better page, answer the question more directly, get ranked for it.

An assistant works in the opposite order. It starts with recall, then searches to confirm or fill in around what it already believes about your category.

If you are not in that recall set, you are not losing on the quality of your page. You are losing before the query is typed.

That is a brand problem wearing a technical costume, and brand familiarity is built the way it always has been: the same name, attached to the same category, in public, repeatedly. It is the same logic as topical consistency beating posting volume in ordinary organic terms.

What most brands try first, and why it stalls

The instinct is to treat this as a new channel with new tricks. The usual first moves:

• Rewriting the homepage with the phrases you want the model to repeat back
• Adding schema markup and hoping recall follows
• Publishing a burst of category content, then stopping after six weeks
• Subscribing to a dashboard that scores your AI visibility without changing anything that produces it
• Chasing a slot on whatever “best X” listicle ranks this month

Some of that is reasonable housekeeping. None of it touches the actual variable, which is whether the model already associates your name with your category before anyone opens a chat window.

Step 1: Settle on one name and one category sentence

Decide two things and write them down before anything else changes.

The name. One spelling, one capitalisation, one decision about trading name versus legal entity. If you appear as three variants of yourself across the press, your profiles and your listings, you are asking a model to merge three weak signals rather than count one strong one.

The category sentence. One plain sentence saying what you are and who it is for, in the words buyers actually type, not the words your positioning deck uses. “A social operations partner for brands publishing daily” is a category sentence. “We help brands tell their story” is not.

If two people in your business would write that sentence differently, you do not have one yet.

Step 2: Make that sentence true everywhere you control

Push it into every surface you own, without creative variation. Variation is the enemy here.

• Website homepage and about page
• Every social profile bio, on every platform, including the dormant ones
• Company pages on professional networks, marketplaces and app stores
• Founder and senior team profiles
• The bio you send to podcast and event organisers
• Any directory or listing you are already in

This is dull work, and it has the best return of anything in the plan, because it is the only part that is entirely yours. It is also the first thing worth checking in a social audit, since drift here is usually invisible from the inside.

Step 3: Publish so the category is unmistakable

An association forms from repetition across many sources. So the published work has to keep landing on the same category rather than wandering across whatever was interesting that week.

That is not an argument for posting more. It is an argument for posting about a narrower set of subjects, more consistently, with the category words appearing naturally in titles, headlines and descriptions.

It also gives a model something to find when it does search, and that path is real. The study includes a case where a brand the model had never recalled was still pulled into its searches, on the strength of live web content. Memory is a heavy thumb on the scale, not a locked door.

Step 4: Earn the mentions you cannot write yourself

The signals that build category memory are the ones other people produce. Coverage, industry write-ups, partner pages, community threads, reviews, and the answers real people give when someone asks for a recommendation.

geoSurge’s own reading is that memory is earned over time through category authority, mentions, coverage and consistent association with a category, rather than won at the moment of the query.

There is no honest shortcut here. Sponsored placements and paid coverage buy you a mention that reads as an advert to a human, and there is no reason to think they produce the independent, repeated association the study describes. The version that works is slower: be genuinely useful in public, be quotable, and answer the question in the place the question gets asked.

Where your social feed fits

The honest answer is that nobody outside the labs can tell you exactly which social content enters a training corpus or a retrieval index. Anyone who tells you they know is guessing.

What can be said is narrower and still useful. Profiles are public pages carrying your name and your category. Video titles, descriptions and captions are text. Community threads recommending you are text. That is where the public record of what you are gets written, and your social operation is writing it daily whether anyone is steering it or not.

So the reasonable position is to treat every bio, caption and description as part of your public description rather than as a post. Worth doing even if no model ever reads one, because humans do.

What good looks like

• One name, spelled and capitalised identically in every public place
• One category sentence, used near enough word for word across your own surfaces
• A publishing pattern that keeps returning to the same three or four subjects
• A steady trickle of third-party mentions you did not write and did not pay for
• A named owner for all of it, and a quarterly check that it has not drifted

None of it is fast. All of it compounds, which is what a familiarity problem needs.

How NBK thinks about it

We would not sell you an AI visibility project, because there is nothing in it yet that is separable from the ordinary work of being consistently and publicly present in your category.

The pattern we would look for first is drift: a brand describing itself several different ways across its own properties, and publishing across too many subjects to be firmly associated with any of them. That is a system problem, not a content problem, and it is the same argument as running social media like a system rather than a calendar.

The uncomfortable part is that the work is the same work. Clear positioning, a narrow subject range, a cadence you can hold, and enough public output that other people start describing you the way you describe yourself. Assistants have simply raised the price of doing it badly.

Next step

If your brand is described three different ways across your own channels, and you are not sure which version the internet has settled on, that is the place to start rather than a new tool.

If your social output feels busy but not effective, an NBK audit is a good way to find the constraint in the system.