Insights•6 min read

AI Visibility for Agencies: 24 Scanned, 22 Scored Zero

July 9, 2026 by NXTG.ai

GEOAI VisibilityOperator Series

I ran 24 real US and UK marketing agencies through geo.nxtg.ai, our live scanner, and 22 of the 24 scored 0 out of 100 — invisible in ChatGPT for the exact category questions their own clients ask. I was scanning other people's agencies that day, not our own products: we were building an outreach list for our GEO product, and the fastest way to open a conversation with an agency is to show them their own AI visibility score. For each one the scanner asks ChatGPT with web search enabled four category-pinned buying questions, the kind a real client actually types, then records who gets cited and in what order. I expected a spread: some agencies strong, some weak, a handful of zeros. Instead, 96 questions in, the scoreboard read 22 zeros out of 24. My first thought was that the scanner had broken.

What Was Happening

Here is what the run actually produced. Twenty-four unique agencies, hand-picked as outreach targets across local SEO, ecommerce, web dev, B2B, white-label, and a few shops that sell AEO and GEO services directly. Four buying questions each, 96 in total, every question pinned to that agency's own stated category. Across all 96, the agency being scanned was cited a grand total of 3 times. Twenty-two of the 24 scored 0 out of 100: absent from every category question I had built specifically for them to win.

The citations did not vanish. ChatGPT cited plenty of domains, just not the agencies. Across the run, 105 distinct other domains got cited instead. And the shape of that list is the real finding: 87 of those 105 domains, 83 percent, were cited exactly once. No incumbent owns these categories. The winners are a long, flat tail of one-time mentions.

Then the part I did not expect. In agency categories, software beat agencies. For "Shopify agency" questions the top citation was Shopify's own app store. For "Shopify SEO agency," the answers were mostly plugins and audit tools, like yoast.com, with a single actual agency named. For "HubSpot marketing agency," the top citation was hubspot.com itself. Buyers were asking for a firm to hire. The machine kept handing them a tool to install.

How I Found It

A scanner returning all zeros is exactly what a broken scanner returns. So before I trusted the finding, I tried to disprove it.

Two things ruled out breakage. First, two of the 24 agencies did score. The pipeline was clearly capable of returning a non-zero result on the same run, same prompts, same day. Second, I re-scanned. The zero-scores reproduced on an independent re-run: an agency absent from its category on the first pass was still absent on the second.

The partial scores were a different story, and worth a paragraph of honesty. The two agencies that did score returned different citation counts run to run. One banked a partial score on the first pass and a noticeably lower one on an independent re-scan the same day. ChatGPT web search is stochastic: the exact citation count for a partially visible domain is not stable. So I stopped trusting the partial numbers, and I am not quoting them here. Absence reproduces. Partial presence does not. That asymmetry is itself the useful signal: a 0 is a hard, repeatable fact, and a 30 is a mood.

What I Changed

1. Stopped treating the zeros as a scanner bug and started treating them as the dataset. The reflex when 22 of 24 come back empty is to debug the tool. I inverted it: the two non-zeros proved the tool worked, so the zeros were the reality I had scanned for.

2. Split every score into "absent" and "present," and threw out the present magnitudes. Because partial counts are stochastic, I only report the binary now: cited in category, or not. The zero-versus-nonzero line is the only line that reproduces, so it is the only line I let drive a decision.

3. Started logging the competitor set, not just the agency's score. The 105-domain scatter and the software-beats-agency pattern were invisible until I read who got cited instead. A bare score tells you that you lost. The citation list tells you who you lost to, and that changes the entire pitch.

4. Rewrote the outreach opener around the vacuum, not the loss. "You score 0" reads like an insult. "Almost nobody in your category scores at all yet, here is the open window" reads like a tip. Same data, opposite conversation.

The result: a scan that started as lead generation became the clearest map I have of an open category. Every zero is a seat nobody is sitting in yet.

The Principle

Here is what generalizes past the 24 agencies we scanned. Your reputation in the human world does not transfer into AI answers by default. Not your rankings, not your awards, not your spot on somebody's "top 10 agencies" list, not even a platform's Premier Partner badge. Those are signals built for humans reading web pages and directories. The model answering a buying question is running retrieval against a different index, and on the evidence of this run it has not indexed your reputation at all. Even agencies that literally sell AI-search optimization scored 0 out of 100 in their own categories.

That sounds like bad news. It is closer to a land grab. When 83 percent of the winners are cited exactly once, no incumbent has locked the category down. AI visibility is a separate asset class from web reputation, it is currently cheap, and it is measurable before it becomes expensive. The order of operations matters: you cannot own a category in AI answers until you measure where you actually stand in it, and most operators have never run the measurement.

One caveat I will hold to. This was 24 hand-picked agencies, four prompts each, one day, ChatGPT web search only. A strong pattern on a small, deliberate sample, not a census of the industry, and not a claim about Gemini or Perplexity, because I did not test them. But the zeros reproduced, and a reproducible zero is a fact you can act on. Measure your own category before you assume your human-world reputation is doing any work inside the machine. On the evidence here, it is not.

This is part of the Operator Series: field research from running AI-native workflows at scale. Previous: "AI Has the Wrong Category for You", on scanning our own products and finding AI had filed every one in the wrong category.

AxW is the founder of NXTG.AI. He writes about AI operations at nxtgai.substack.com.