APOG Systems

AI visibility — and the part nobody quotes

More people are asking an AI assistant for a plumber instead of scrolling Google. If it never names you, you don't exist in that conversation. That's the pitch you've probably already heard. Here's what the research actually says, including the part that gets left out of every sales deck we've seen.

The study everyone cites

The reference work in this field is “GEO: Generative Engine Optimization” by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, presented at KDD 2024. It tested 10,000 queries across 25 domains — the only controlled experiment of its kind. Its framing of the problem is exactly right:

Given the black-box and fast-moving nature of generative engines, content creators have little to no control over when and how their content is displayed. Aggarwal et al., “GEO: Generative Engine Optimization,” KDD 2024

And it's the source of the number you'll see quoted everywhere:

Through rigorous evaluation, we demonstrate that GEO can boost visibility by up to 40% in generative engine responses. Aggarwal et al., “GEO: Generative Engine Optimization,” KDD 2024

How that 40% was actually produced

We read the code the authors released with the paper. The top-performing methods work by instructing the language model to make things up. Verbatim from the study's own source — src/geo_functions.py, which you can open and read yourself:

So what the study proves is that answer engines reward text that is shaped like evidence — whether or not it is evidence. That's a finding about a weakness in the machines, not a marketing best practice.

We won't do that on your website. Inventing statistics or citations on a contractor's site is deceptive advertising, and it stays deceptive whether or not it works.

What that means for what we sell

We do the honest version: real quotes, real numbers, real links to real sources. And here is the part a vendor is not supposed to tell you — the honest version has never been measured. The paper didn't test it. Nobody has. It plausibly captures the same effect, because the engines appear to respond to the shape of the text, but plausibly is the honest word and we're going to keep using it.

There's more. In the one part of that study run against a real production engine (Perplexity), adding cited sources scored 19.0 against a 24.7 baseline — it performed worse than doing nothing. Any vendor quoting you a guaranteed lift from “adding citations” either hasn't read the tables or is counting on you not to.

Being read is not the same as being trained on

Two different things get called “AI crawling,” and vendors blur them together because the blur sells better.

Retrieval is a live lookup. Someone asks an assistant for an electrician in Houston and it goes and reads pages right then, to answer that question. That is where a citation with a link to your business comes from, and it is the only one of the two that can put a call on your phone this month.

Training is your text being absorbed into a model's weights. It surfaces months or years later as vague recall, with no link and nothing a reader can click or check. And a model that memorized your hours in 2025 will still be repeating them in 2027.

These use different crawlers, and you control them separately. So when someone tells you your business needs to be “in the AI,” the question worth asking back is: retrieved, or trained on? For a contractor it is retrieved, every time. Nobody needs to train a model on your website for an assistant to name you today.

Two things follow, and neither is what the sales pitch leads with.

Your website is not the main lever. When the question is local, the assistants lean on your Google Business Profile and the big directories rather than your homepage copy. That is why most of what we do points off your site.

But a site that cannot be fetched cannot be retrieved at all. That is the unglamorous half: pages that render on the server, a phone number and address sitting in the actual HTML instead of being assembled by JavaScript after the fact, and nothing in your robots file quietly waving the search crawlers off. Most sites we look at fail at least one of those, and it is almost always a setting somebody switched on years ago and never looked at again.

What the evidence does support

Directories are where the citations come from

Foundation Inc and AirOps analyzed more than 28 million AI responses to local queries across ChatGPT, Gemini, Perplexity and Google AI Mode in Q4 2025. Yelp drew 512,680 citations — about 3.4× the next directory, and made up 72.5% of the competitive directory set on Google AI Mode. Between 91% and 97% of those citations came from unbranded searches — people asking for a plumber, not for a company by name.

That is the strongest evidence in this whole field, and it points off your website entirely. Which is why we work on your profiles and listings, not on your homepage copy.

But don't overestimate how often AI is involved yet

AI Overviews appear on 17.7% of home-service searches, and on only 12.2–12.4% when someone includes a city or “near me” (WebFX, 237,990 queries, July 2025). The high-intent moment — someone ready to call — is still the map pack. We fix that first and we'd tell you to hire someone else if they told you otherwise.

Being visible off your own site matters more than tuning it

Ahrefs studied 75,000 brands in May 2026 and found YouTube mentions the single strongest correlate of AI brand visibility (ρ ≈ 0.737) — stronger than branded search volume or site size. One honest caveat: their sample was established brands by construction, so a Houston contractor sits outside the population they measured. It's a direction, not a promise.

What we actually do

Prices are on the pricing page. Call (713) 487-6870 or email gus@apogsystems.com and we'll walk your listing with you.

Get a free AI visibility check

Tell us who you are. We'll check whether Google and ChatGPT name your company, then Gus calls you with what we found.

Free. No contract. We'll only use this to call you back — see our privacy policy.