The 2026 Houston AI Visibility Index

When a Houston resident opens ChatGPT and asks who to call for a root canal, a flooded kitchen, or a car-accident claim, an answer comes back in seconds. It is short. It is confident. It names two or three businesses. There is no second page.

For twenty years, being found meant ranking on Google. That problem had a tail: page two still existed, position seven still got clicks. The AI answer has no tail. You are named, or you are not in the conversation. We built the Houston AI Visibility Index to measure, for the first time and at the level of a single composite score, who the machines are naming — and who they have quietly erased.

Author’s note

Every local-business owner I talk to in Houston is about to lose the front door of their business, and most of them do not know it yet. I filed five patents on the mechanics of this shift and then did the unglamorous thing: I ran the measurement. This is Volume I of a recurring index. The numbers are uncomfortable. They are also, for the operators willing to act, the clearest map of unclaimed ground I have seen since the early days of search.

— Shayne Beavan

Executive summary

We dispatched 750 buyer-intent prompts across 5 frontier assistants — ChatGPT, Claude, Gemini, Grok, and Perplexity — and analyzed all 3,750 machine-responses for three things: was a business named, was the mention carried by a citation, and were the stated facts true. We scored 150 anonymized business cohorts across 3 Houston sectors on a single 0–100 composite. Six findings stand out:

1. The median Houston business scores 31 out of 100. Visibility is the exception, not the baseline. 48% of the frame is effectively invisible — named in fewer than one in five relevant answers.
2. AI recommendation is winner-take-most. The top five cohorts in each sector capture the majority of all mentions; everyone else competes for the remainder.
3. The five assistants disagree sharply on trust. Citation rate runs from 94% (Perplexity) down to 27% (ChatGPT). The most-used assistant cites the least.
4. Hallucination is real and uneven. Models invent verifiable facts about real businesses between 1.9% (Claude) and 5.8% (Grok) of the time.
5. Structured data is the strongest available predictor of being named — schema.org completeness tracks mention rate at r = 0.71, ahead of Google Business Profile completeness and classic domain authority.
6. Machine authority is geographic. Inner-loop ZIP codes are recommended roughly 3.2× more often than equally dense suburban codes, leaving large authority dead zones unclaimed.

What the Index measures

HAVI is a composite, not a single metric, because being recommended by an AI is not a single behavior. A business has to be retrieved, trusted, cited, and described accurately. Each cohort receives a 0–100 score built from four weighted sub-indices.

The frame is Houston metro (Harris + Fort Bend counties): 50 established providers in each of 3 sectors, assembled from public directories and scored as anonymized cohorts. The prompt bank holds 250 buyer-intent questions per sector — the phrasing residents actually type, not keywords. Every prompt ran against every model during May 1–28, 2026.

Finding 1 — The Houston visibility gap

The metro-wide median composite is 31 out of 100. Read that plainly: a typical established Houston business is named in roughly a third of the AI answers where it could belong, often without a citation, and sometimes with a fabricated detail attached.

The distribution is steep. The top decile clears 72. Just below the leaders, the floor drops out: 48% of the frame scores under 20 — businesses that exist in public records, on Google, and on their own websites, and are almost entirely absent from the AI layer now sitting in front of all three.

Finding 2 — Winner-take-most

Search returned a page of ten options. The assistant returns a list of three. Across the frame, the top five cohorts in each sector capture the majority of every mention the sector receives.

This is the structural break between search and answer. A tenth-place result on Google still drew traffic. A tenth-place business in an AI’s consideration set is usually not mentioned at all. The cost of being average moved from *less traffic* to *no presence*. Home services is the most concentrated and the least machine-present sector in the study, with a median of just 24; dental, where the earliest operators have started publishing structured identity, leads at 38.

Finding 3 — The five assistants do not agree

Treating “AI” as one thing is the most expensive assumption an operator can make. The assistants differ most on the dimension buyers use to trust an answer: whether it is cited.

Perplexity cites 94% of the time and will omit a business rather than invent one — if your entity is not on a crawlable page, you are simply left out. ChatGPT, the most-used assistant in the study, cites only 27% of the time, which makes its recommendations the hardest for a buyer to verify and the hardest for a business to trace back to a source. Grok is the most willing to assert specifics it cannot support, fabricating a verifiable fact 5.8% of the time. Claude fabricates least, at 1.9%, but is also the quickest to decline to name anyone when its confidence runs low.

Finding 4 — Structured data predicts presence

We tested which signals move with mention rate across the frame. The strongest is not a classic SEO signal. It is whether a business publishes machine-readable identity: schema.org Organization markup, a consistent name/address/phone footprint, and a connected sameAs graph.

These are associations, not proof of cause, and we report them as such. But the ordering is consistent and it is actionable: the signals a retrieval system can parse directly outrank the signals built for a human skimming a results page. The AI internet weights structured data differently than the search internet did. The businesses publishing it are not winning because they are larger. They are winning because they are legible.

Finding 5 — The geography of machine authority

Machine authority clusters, and the cluster is the inner loop. ZIP codes close to the urban core concentrate AI recommendations far beyond their share of businesses, while high-population suburban codes — full of qualified providers — go thinly represented.

The dead zones are the opportunity. A roofer in 77449 or a dentist in 77072 competes against far fewer machine-legible rivals than one inside the loop. Authority there is not contested. It is unclaimed.

What this means for operators

Whether an AI assistant recommends you over a competitor in the next twelve months is, to a first approximation, a function of how completely your business is represented in machine-readable form today. That is the uncomfortable conclusion of this Index, and also the hopeful one: the gap is large, it is measurable, and it is mostly unclaimed. The businesses that close it first will hold the position the way early domain owners held theirs — not because they were biggest, but because they were there first and built to be read.

Be the answer when AI is asked. Or be the business that wasn’t.

Methodology and limitations

Frame. 150 established providers (50 per sector) across Houston metro (Harris + Fort Bend counties), drawn from public directories. Cohorts are anonymized; no score in this report is attributed to a named third-party business.

Instrument. Deep AI Solutions’ cross-model AI Visibility Scanner dispatched 250 buyer-intent prompts per sector (750 total) to the then-current flagship consumer model from each of the five providers during May 1–28, 2026, capturing 3,750 machine-responses. Each response was parsed for mention, rank, citation, sentiment, and factual accuracy against a verified record of each cohort.

Scoring. The HAVI composite weights Mention Rate (40%), Machine Trust (25%), Citation Capture (20%), and Hallucination Resistance (15%).

Limitations. Findings describe this frame, this prompt bank, and this measurement window. Correlations are associations, not causal claims. Model behavior changes between versions — consumer defaults shifted even during the window we measured. This is Volume I of a recurring Index; method and frame will be versioned as it repeats. Figures are Deep AI Solutions’ own measured results, not independent third-party audits.