AI Already Picked Its Shortlist — Before It Searched

An AI assistant decides which brands are worth looking up before it searches. Brands the model already knew were looked up far more often than the ones it didn't — 55.7% of the time against 17.4%. Vigo's team measured it across nine industries, and the same pattern showed up in every one.

A Study of 13,281 Hidden Searches Found the Same Pattern in All Nine Industries

Francisco Vigo is the co-founder and CEO of GeoSurge, an AI visibility company, and his guest piece in Marketing Dive says most 2027 AI visibility spending is aimed at the wrong stage of the process.

His team ran a test. First they wrote down which brands the model could name with the web turned off, across nine industries. Then, over 12 days, they put 66 real buyer questions to the model 60 times each, and logged roughly 4,000 answers plus 13,281 searches the model ran in the background.

Here is the number that should stop you. Brands the model already knew were looked up far more often than the ones it didn't — 55.7% of the time against 17.4%. For the five brands it remembered best, that rate hit 67%.

The same pattern showed up in every one of the nine industries they tested.

Close to 69% of the searches were plain category searches with no brand named. Of the 31% that did name a brand, almost two-thirds pointed at one of the five the model remembered best.

Vigo also shows where memory matters most. In automotive, 82% of the searches that named a brand pointed at one of the model's top five. In finance, 77%. In fitness and wellness it was closer to 50%, because models there looked past what they already knew far more often.

Each industry in the study leans on only six to 12 prompts, so he says to read those splits as indicative.

Now here is the part that should make an owner sit up. My read: the model is not choosing between you and the guy across town when your customer asks. It already chose.

The Assistant Picks the Shortlist Before It Ever Searches

At first, brands feared an assistant's knowledge was locked at the moment it was trained. Live web search looked like the answer, since a model can look up anything it does not already know. Vigo calls that RAG, or Retrieval-Augmented Generation.

But the model does not start from a blank page. It picks what to look up first, and it picks from what it already remembers.

Modern AI search doesn't run one query either. Ask about the best option for a small company and the model breaks it into several searches at once, some generic and some naming specific companies. Google calls this query fan-out.

That is where the shortlist gets built. The queries a model writes come out of the same associations it learned during training. As Vigo puts it, "It can't search for a brand it never learned."

Which means the fight is not happening where most owners think it is.

What This Means If You Sell HVAC or Roofing (My Read)

My examples here are HVAC and roofing, which the source does not name. That is why step one below is to go check. But here is how I would think about it.

Picture an HVAC company in a mid-size market. It is the second week of July, a homeowner's AC dies, and she asks an assistant who to call for AC repair near her. The model fans that out into a batch of searches. Some are generic — AC repair near me, emergency HVAC service. Some name companies. If your company never got written into what the model remembers, you are not in the named batch. You are fighting for the generic scraps against whoever the model already trusts.

Picture a roofing contractor after a hailstorm. Same thing, worse timing. Storm damage means a hundred homeowners asking the same question in the same week. The contractor whose name the model already remembers gets pulled into the named searches. The one who does not gets whatever is left.

Here is the part that should worry you most. Vigo writes that brands the model remembers get looked up, the ones that get looked up get cited, and those citations become part of what the next round of models learns. "You can't win at query time what you lost at training time," he writes.

For a challenger, that loop gets harder to break each cycle. For the shop that is already the biggest name in town, the loop works in its favor.

I have spent a long time on this side of the desk. Over $75 million in online sales, built two companies to eight figures, and built multiple 8-figure sales teams. The thing that kills small operators is never the ad budget. It is being invisible at the moment the decision gets made. This is that moment, moved upstream.

How to Work This Into Your Week

None of this needs a new department. Here is what I would do, in order.

1. Run the questions your customers actually ask. Ten minutes. Open the assistant you use and type what a homeowner would type: who should I call for AC repair in your town, best roofing company near me for storm damage. Write down whether you show up, and who does. Do it three times over a week and compare.

2. Diagnose before you spend. Vigo's first recommendation is to diagnose before you spend. If your category is search-led, clean, easy-to-retrieve content can move the needle fairly fast. If it is memory-led, the same work tops out much lower. Know which game you are playing before you write a check.

3. Fund earned media like it is infrastructure, not fluff. What a model remembers comes from what it trained on: years of press, analyst reports, reviews, forum threads, and a steady association between a brand and its category. That means reviews, local press, and being talked about where your customers already hang out online. It builds between training runs, or it never builds at all.

4. Measure what the model remembers, not just what it cites. Vigo's line is that citation counts only show the scoreboard, while "the game is decided upstream." He also notes that tools that used to run hundreds of dollars a month now come in free versions. Ask what you are actually measuring before you trust any vendor's number.

5. Go after the generic queries, because that is where the volume is. Roughly 69% of the searches in the study named no brand at all. That is retrievability work — clean pages, clear service pages, real answers to real questions — and Vigo says it still works, even in memory-led categories.

6. Say your category out loud on your own site. Vigo describes it as a steady association between a brand and its category. If your homepage talks about comfort solutions and never says HVAC repair or roof replacement, you are making the model guess.

7. Fix the page a phone visitor lands on. If the page is slow or the form is long, the lead you just earned dies in the driveway.

Ready-to-run prompts, one per trade:

You are a homeowner in [city] with no air conditioning in July. Which three HVAC companies would you call, in order, and what do you know about each one?

I need a roof replaced after hail damage in [city]. Which local roofing contractors do you know about, and what do you know about each?

Write the first 50 words of a roofing company homepage so an AI assistant can tell in one pass that we do roof replacement and storm damage repair in [city].

Run the first two weekly. The answers are your scoreboard.

Mistakes I See Owners Make

• Chasing citations before memory. If you are not in the model's memory, winning a citation here and there is a rounding error. Fix the upstream thing first.
• Treating reviews as a nice-to-have. Reviews are part of the corpus now. They are not a vanity metric anymore.
• Assuming your category works like someone else's. Automotive sat at 82%. Fitness and wellness sat near 50%. Those are different games with different budgets.
• Reading one answer and panicking. The study used six to 12 prompts per industry and Vigo calls the splits indicative. Check more than once.