AI SEO Statistics: Steel (2026-07 edition)
40 questions · 120/120 expected AI responses · 3 models · measured 2026-07-06
The question bank
The questions we tested — a frozen buyer-intent benchmark for steel.
The question set was curated from a predefined buyer-intent taxonomy and held constant for this edition. Each model received the same wording. These are the prompts behind every percentage on this page.
Show all 40 questions
Model by model
18.5% question-level model disagreement.
This rate is the average pairwise disagreement between binary behavior codes across questions and behaviors. It is not the gap between the highest and lowest aggregated model percentages.
Behavior prevalence across 40 steel benchmark questions, 2026-07 edition. Last column: equal-model mean.
| Behavior | ChatGPT | Claude | Gemini | Equal-model mean |
|---|---|---|---|---|
| Recommends hiring a professional | 42.5% | 40% | 17.5% | 33.3% |
| Suggests DIY first | 27.5% | 17.5% | 17.5% | 20.8% |
| Names specific providers | 17.5% | 12.5% | 15% | 15% |
| Gives price or cost info | 10% | 15% | 17.5% | 14.2% |
| Tells to check reviews | 0% | 0% | 0% | 0% |
| Tells to verify credentials | 25% | 25% | 10% | 20% |
| Mentions case studies / portfolio | 0% | 0% | 0% | 0% |
| Mentions local proximity | 12.5% | 15% | 7.5% | 11.7% |
| Gives selection criteria | 37.5% | 47.5% | 30% | 38.3% |
| Warns about red flags | 5% | 5% | 10% | 6.7% |
| Asks a clarifying question | 55% | 62.5% | 0% | 39.2% |
| Recommends multiple quotes | 7.5% | 7.5% | 0% | 5% |
Question-level agreement
How often all measured models received the same binary code.
Agreement is calculated question by question for each behavior. A high value can coexist with a low behavior prevalence; it means the models usually agreed on whether the behavior appeared.
All-model binary agreement by behavior across 40 benchmark questions.
| Behavior | All-model agreement |
|---|---|
| Recommends hiring a professional | 57.5% |
| Suggests DIY first | 72.5% |
| Names specific providers | 62.5% |
| Gives price or cost info | 80% |
| Tells to check reviews | 100% |
| Tells to verify credentials | 65% |
| Mentions case studies / portfolio | 100% |
| Mentions local proximity | 85% |
| Gives selection criteria | 52.5% |
| Warns about red flags | 90% |
| Asks a clarifying question | 15% |
| Recommends multiple quotes | 87.5% |
By model
How each assistant handled Steel questions.
Reading the 120 answers model by model shows how differently the three assistants treat the same steel questions. On the most consequential behavior — whether to send the buyer to a professional at all — the rate ranged from 42.5% (ChatGPT) down to 17.5% (Gemini), a 25-point gap on an identical question set.
Across the 40 steel answers it produced, ChatGPT recommended hiring a professional in 42.5% of them and suggested a DIY approach first 27.5% of the time. It named a specific provider in 17.5% of answers (about 0.5 distinct providers per answer) and included price or cost information 10% of the time. ChatGPT asked a clarifying question before answering in 55% of cases, warned about red flags or scams in 5%, and told the buyer to verify credentials in 25%, averaging 595 words per answer. On the remaining cues it told the buyer to check reviews in 0%, pointed to case studies or a portfolio in 0%, and framed the choice around local proximity in 12.5%; a selection-criteria checklist appeared in 37.5% of its answers and a recommendation to gather multiple quotes in 7.5%.
Across the 40 steel answers it produced, Claude recommended hiring a professional in 40% of them and suggested a DIY approach first 17.5% of the time. It named a specific provider in 12.5% of answers (about 0.5 distinct providers per answer) and included price or cost information 15% of the time. Claude asked a clarifying question before answering in 62.5% of cases, warned about red flags or scams in 5%, and told the buyer to verify credentials in 25%, averaging 308 words per answer. On the remaining cues it told the buyer to check reviews in 0%, pointed to case studies or a portfolio in 0%, and framed the choice around local proximity in 15%; a selection-criteria checklist appeared in 47.5% of its answers and a recommendation to gather multiple quotes in 7.5%.
Across the 40 steel answers it produced, Gemini recommended hiring a professional in 17.5% of them and suggested a DIY approach first 17.5% of the time. It named a specific provider in 15% of answers (about 0.5 distinct providers per answer) and included price or cost information 17.5% of the time. Gemini asked a clarifying question before answering in 0% of cases, warned about red flags or scams in 10%, and told the buyer to verify credentials in 10%, averaging 241 words per answer. On the remaining cues it told the buyer to check reviews in 0%, pointed to case studies or a portfolio in 0%, and framed the choice around local proximity in 7.5%; a selection-criteria checklist appeared in 30% of its answers and a recommendation to gather multiple quotes in 0%.
Taken together, ChatGPT is the assistant most likely to route a steel buyer to a professional (42.5%) and Gemini the least (17.5%). ChatGPT produced the longest answers, at 595 words on average. Specific providers were named most often by ChatGPT (17.5%) — even there, roughly one answer in 6 carried a name.
Where they disagree
The behaviors where the choice of model changes the answer.
Question-level model disagreement is 18.5% — the average pairwise rate at which models received different binary codes across questions and behaviors. The observed rate spreads below are a separate measure showing where which assistant a steel buyer happens to ask matters most:
- Asks a clarifying question: from 0% (Gemini) to 62.5% (Claude) — a 63-point spread.
- Recommends hiring a professional: from 17.5% (Gemini) to 42.5% (ChatGPT) — a 25-point spread.
- Gives selection criteria: from 30% (Gemini) to 47.5% (Claude) — a 18-point spread.
- Tells the buyer to verify credentials: from 10% (Gemini) to 25% (ChatGPT) — a 15-point spread.
- Suggests a DIY approach first: from 17.5% (Claude) to 27.5% (ChatGPT) — a 10-point spread.
The widest single gap — asks a clarifying question, 63 points — means a steel buyer can receive materially different guidance on the same question depending only on which assistant they happen to open, so any visibility strategy built on a single model's behavior describes only part of the steel market.
Where they agree
The points of near-consensus in Steel.
On other behaviors the three models move almost in lockstep — the points of near-consensus for steel, where all three landed within a few points of each other:
- Tells the buyer to check reviews: 0% across all three models.
- Mentions case studies or portfolio: 0% across all three models.
- Names a specific provider: 12.5%–17.5% across all three (a 5-point spread).
- Warns about red flags or scams: 5%–10% across all three (a 5-point spread).
Measured question by question, the three assistants coded a response the same way most consistently on "tells the buyer to check reviews" (identical coding in 100% of questions) and least consistently on "asks a clarifying question" (15%).
Every behavior, measured
All twelve coded behaviors for Steel, averaged across the three models.
The behaviors AI models reproduce most often for steel are asks a clarifying question (39.2% on average), gives selection criteria (38.3%) and recommends hiring a professional (33.3%); the rarest are mentions case studies or portfolio (0%), tells the buyer to check reviews (0%) and recommends multiple quotes (5%). Each figure below is the share of a model's 40 answers in which the behavior appeared at least once, averaged across the 3 models with the full per-model range in parentheses:
- Asks a clarifying question: 39.2% on average (ChatGPT 55%, Claude 62.5%, Gemini 0%) — a 63-point spread.
- Gives selection criteria: 38.3% on average (ChatGPT 37.5%, Claude 47.5%, Gemini 30%) — a 18-point spread.
- Recommends hiring a professional: 33.3% on average (ChatGPT 42.5%, Claude 40%, Gemini 17.5%) — a 25-point spread.
- Suggests a DIY approach first: 20.8% on average (ChatGPT 27.5%, Claude 17.5%, Gemini 17.5%) — a 10-point spread.
- Tells the buyer to verify credentials: 20% on average (ChatGPT 25%, Claude 25%, Gemini 10%) — a 15-point spread.
- Names a specific provider: 15% on average (ChatGPT 17.5%, Claude 12.5%, Gemini 15%) — a 5-point spread.
- Gives price or cost information: 14.2% on average (ChatGPT 10%, Claude 15%, Gemini 17.5%) — a 8-point spread.
- Mentions local proximity: 11.7% on average (ChatGPT 12.5%, Claude 15%, Gemini 7.5%) — a 8-point spread.
- Warns about red flags or scams: 6.7% on average (ChatGPT 5%, Claude 5%, Gemini 10%) — a 5-point spread.
- Recommends multiple quotes: 5% on average (ChatGPT 7.5%, Claude 7.5%, Gemini 0%) — a 8-point spread.
- Tells the buyer to check reviews: 0% on average (ChatGPT 0%, Claude 0%, Gemini 0%).
- Mentions case studies or portfolio: 0% on average (ChatGPT 0%, Claude 0%, Gemini 0%).
Trust signals
How well the models protect the steel buyer.
Beyond whether to hire, the rubric codes how carefully each assistant protects the steel buyer once a decision is made. Telling the buyer to check reviews or ratings appeared in 0% of answers on average. Verifying credentials or certifications appeared in 20%. Warning about red flags or scams appeared in 6.7%.
On structuring the decision, a selection-criteria checklist showed up in 38.3% of answers on average and a recommendation to gather multiple quotes in 5%. The single least-reproduced protective signal for steel is "tells the buyer to check reviews" at 0% on average — the clearest opening for content that supplies it, since the models are not yet reliably surfacing that guidance on their own.
Referral behavior
Do AI models name Steel providers?
For service providers the decisive question is whether these systems name anyone at all. Across 120 steel answers, a specific provider was named in 15% of responses on average — roughly 0.5 distinct providers per answer. In practice the assistants behave far more as an explanatory layer than as a referral engine for steel: visibility comes from being the reasoning a model reproduces, not from being the named recommendation.
The question set
What these 40 Steel questions cover.
The 40 questions behind every percentage on this page form a frozen steel (manufacturing / industrial B2B; buyer hiring decisions for this specific service) buyer-intent benchmark. Each was put to all 3 models once, with identical wording, so the rates above describe how the assistants handled this exact steel question set — not a general prior or a hand-picked subset. The full list is shown earlier on this page; the coded percentages are what those specific questions produced.
How to read this
A note on the numbers.
A percentage here is the share of a model's 40 answers in which the behavior appeared at least once — not a confidence score. Because each model answered every question exactly once on 2026-07-06, the figures describe this specific steel question set and snapshot rather than a general prior. The full protocol and coding rubric are documented in the study methodology.
Methodology
A controlled snapshot, documented end to end.
40 frozen benchmark questions, one expected response per model per question (ChatGPT API (gpt-5-mini), Claude API (claude-sonnet-5), Gemini API (gemini-3-flash-preview)), collected 2026-07-06 and coded against a fixed 12-behavior rubric. The pipeline validates the schema, recomputes aggregates and reports consistency issues. AI outputs vary with model version, location and time, so the figures describe this edition's exact sample and measurement window. Read the full methodology →
Citation
Cite this edition.
Authority Specialist. “AI SEO Statistics: Steel (2026-07 edition).” AuthoritySpecialist.com. https://authorityspecialist.com/research/ai-seo-statistics/services/manufacturing/steel