Insights•6 min read

AI Has the Wrong Category for You: What Our Scans Showed

July 8, 2026 by NXTG.ai

GEOAI VisibilityOperator Series

On June 2, 2026, I pointed our own GEO scanner at our own products. The runner, dogfood-scan.ts, asks ChatGPT with web search enabled the buying questions a real customer would ask, four generated prompts per domain, then records who gets cited and in what order. All four NXTG products scored 0 out of 100. Forge: 0. Dx3: 0. Faultline: 0. NextGen AI: 0.

I already published that scoreboard. The zeros were the headline, and frankly, the zeros were expected for products this young with no third-party footprint.

The finding that actually changed our roadmap took a second pass over the raw output. Next to every score, the scanner logs the competitor set: the companies the assistant recommended instead of us. When I read those lists product by product, the problem stopped being "we rank low in our category." The problem became: AI has filed every one of our products in the wrong category. You cannot win a buying question for a category the model does not place you in. And most of the "vs" pages we had queued up would have made it worse.

The Wrong Shelf

Here is what the competitor sets said, product by product.

Forge is an AI delivery and agent-orchestration platform. Its job is taking an AI prototype to production. The assistant cited it against Taskade, Canva, and ToolHorizon. Task managers and a design tool. The model read Forge as generic productivity software.

Dx3 is a knowledge and intelligence platform built on pgvector and Apache AGE. The assistant cited Kustomer and Tidio. Those are customer-support chatbots. Wrong aisle entirely.

Faultline is data pipeline observability. It drew Integrate.io and Qlik, which are data-integration vendors. The closest of the four, but still adjacent: integration moves data, observability tells you when the movement broke.

NextGen AI, the studio umbrella over all of it, drew ApiScout, AI Magazine, and RankedSuite. Three citations with no coherent category between them. The model has no idea what shelf the org belongs on, because an org umbrella has no single buyer category.

Four products, four scans, zero cases where the model compared us to what we actually compete with. That is not a ranking problem. A ranking problem means you are in the race and losing. This is an entry problem: we were never entered.

How I Found It

The scanner's score is a summary. The diagnosis lives in the annotations, specifically in who got cited and in what order.

For "best X" buying questions, ChatGPT with web search cites in a rough, observable order: first, third-party roundups and listicles it already trusts (TechRadar, G2, Capterra, Wikipedia), then authoritative comparison content, then the vendor's own pages, and the vendor's pages only when they are structured and category-clear. A score of 0 is not a penalty. It is absence. A new product with no third-party footprint and an ambiguous homepage is invisible at every one of those three tiers.

The moment the wrong-category pattern snapped into focus was reading the Forge set. Taskade next to Canva next to a tool directory is not "who beats Forge at orchestration." It is the model guessing a category from a thin homepage and answering a different question than the one we sell into.

That raised a measurement caveat I had to be honest about: generateBuyingPrompts derives its four prompts per domain from what the model can infer about you. If your homepage does not state your category, you may be getting scored against a hallucinated one. The score could move before a single new page ships, just by fixing the sentence that tells the model what you are.

What I Changed

1. Reordered the roadmap: category definition before comparison. The queued "Forge vs Taskade" style pages went from first to second. Job #1 is now one category-definition page per product, Forge first, because a comparison page that accepts the wrong frame teaches the model the frame is right.

2. Specced the Forge category page as a reframe, not a rebuttal. "What is an AI agent-orchestration platform" defines the space and explicitly contrasts it with task managers. It intercepts the Taskade citation by answering the category question, not by arguing rank within the wrong category.

3. Re-pointed the comparison targets at the real competitors. Dx3 gets compared against IBM watsonx and Google Vertex AI, not against the support chatbots the scan surfaced. Faultline gets an honest Integrate.io and Qlik comparison that states the observability-versus-integration distinction up front, since that mis-cite was merely adjacent.

4. Put applicationCategory into the JSON-LD spec for every product page. SoftwareApplication schema with an explicit category string is the cheapest machine-readable category-teaching signal we own. The spec went to the build team with the copy.

5. Fixed the homepage category signal first. One plain sentence per product naming the category, above the fold. Cheapest change on the list, and per the measurement caveat, the only one that could move the score before anything new gets indexed.

6. Filed a scanner change: log the generated prompts alongside the score. Next scan, we verify we are being measured on the right question before we trust the number.

The score has not moved yet, and I did not expect it to. Re-scans run weekly, and GEO lift rides content indexing plus third-party citations, which move in weeks, not days. What moved immediately was the work order: roughly a dozen planned pages got re-sequenced in one afternoon, and two of them got killed outright.

The Principle

Check who AI cites alongside you before you write a single "vs" page.

The competitor set is the diagnostic. The score is only the symptom. When an assistant answers a buying question, it is answering from a category model, and if that model has you on the wrong shelf, every move you make inside the wrong frame reinforces it. A "Forge vs Taskade" page written straight would have been us formally enrolling in the productivity-app category, in writing, on our own domain, with structured data attached.

The order of operations that follows: define the category, position within it, then compare. Skip step one and steps two and three compound the error.

This generalizes past our four products. If you run marketing for a client, the competitor set is the first artifact to pull when they ask why AI never mentions them. It answers a question the score cannot: not "how visible are we" but "visible as what."

If you want the same diagnostic on your own brand, run the free scan on your domain. When the report comes back, skip past the score. Read the competitor set. If the companies listed next to you are not the companies you actually compete with, you have a category problem, not a ranking problem, and now you know which page to write first.

This is part of the Operator Series and week two of the GEO dogfood thread. Previous: "We Built an AI-Visibility Tool and Scored 0/100 on Ourselves".

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