AI Visibility for Construction Firms: Prompts, Wrong Facts and a 30-Day Plan

Commercial buyers now use AI assistants to build a shortlist before they issue an RFP. Their questions are specific about scope, place and risk. Here are realistic examples of what they type:

• "Which general contractors in Phoenix build ground-up medical office buildings?"
• "Best design-build firms for tilt-up warehouses in North Texas"
• "Is [firm name] licensed as a Class A general contractor in Virginia?"
• "Which construction companies near Denver can bond a $25 million public school project?"
• "Union contractors in Chicago with healthcare renovation experience"
• "Who built the new library in [town], and do they do municipal work in other states?"
• "Compare [firm A] and [firm B] for a multifamily project in Atlanta"

Each prompt mixes a project type, a region and a qualification. That tells you what to publish. An engine can only repeat a fact it can find, so the firm that states "we build medical office and outpatient facilities in Arizona and Nevada" in plain words gives the engine a sentence it can quote. A gallery of unlabelled project photos does not.

One caution: the prompt a buyer types is not the query the engine runs. OpenAI says ChatGPT search rewrites a question into one or more targeted queries. Treat example prompts as a way to test coverage, not as keywords to stuff into pages.

Different buyers use different assistants, and each engine draws on a different index. As of October 2026, the ones worth checking separately are:

• ChatGPT. Its search feature is governed by the OAI-SearchBot crawler. Answers given without search come from trained knowledge with a fixed cutoff, so a 2026 rebrand or new license may be missing.
• Google AI Overviews and AI Mode. Google says a page must be indexed and eligible to show with a snippet to appear as a supporting link. There is no separate AI path.
• Microsoft Copilot. It grounds on the Bing index. In Otterly's 2026 experiment, Copilot accounted for 2,718 of the 3,934 citations tracked for one test site, after Google had dropped that site.
• Perplexity and Claude. Both run their own search crawlers (PerplexityBot, Claude-SearchBot).

The lesson from Otterly's test is that one engine can cite you heavily while others barely do. Measure each engine on its own rather than trusting a single blended number. Answers can also differ between an engine's API and its consumer app, which this comparison of API and app answers explains.

Construction is full of near-identical names. Words like Summit, Pinnacle, Heritage and Builders repeat across states, and many firms trade under a DBA that differs from the legal entity on the license.

Say you run Summit Builders Group, a commercial GC in Ohio. A residential remodeler called Summit Builders operates in Oregon. An assistant answering "Is Summit Builders licensed in Ohio?" may blend the two: your license with their reviews, or their service area with your name. That's a hypothetical, but the pattern of entity collision is common enough to check for.

What you can do is make the distinguishing facts appear together, in text, on one fetchable page:

• Legal entity name and every DBA
• Headquarters city and state
• License numbers and classifications, by state
• Year founded and any former names, mergers or acquisitions

Then make sure directories show the same combination. When the sources an engine reads agree on name plus city plus license, there is less room for a merge. If you also do residential work, the home services and contractors guide covers the consumer side.

No published study counts which facts AI gets wrong about construction firms specifically. These are the ones that change often, vary by state, or matter most to an owner's risk review, so check them first:

• Project types. Engines may describe you by your most photographed project, not your core market.
• Service regions. A firm that expanded from one metro to three may still be described by its original footprint.
• License classes and states. A wrong classification or a missing state can knock you off a shortlist.
• Bonding capacity. If you publish a single-project and aggregate figure, state it as a number with a date. If you don't publish it, check that no answer invents one.
• Ownership and leadership. Acquisitions and successions often lag in trained knowledge.
• Delivery methods and union status. Design-build, CM at risk, open shop or union.

When an answer is wrong, work out why before you fix it. An outdated answer from trained knowledge and an invented one need different responses, as the guide to hallucinations versus outdated answers explains. For the ChatGPT side specifically, see fixing ChatGPT errors in trained knowledge versus search.

When engines cite other sites instead of yours, being named correctly on those sites is the route into the answer. That's an association in the evidence, not a promised effect. For construction firms, the sources owners and engines both lean on are:

• State contractor license lookups. These are the authority on classification and status. Confirm the legal name matches your website.
• Public bid portals and plan rooms. Award records tie your name to project types and owners.
• Trade association member directories. They list markets, regions and delivery methods.
• Google Business Profile and Bing Places. Address, category and service area feed local answers.
• Owner and architect project pages. A hospital or school district naming you as GC is a strong independent mention.
• Local business press and award lists. These often describe ranking, size and specialty.

Work through each one and log mismatches in name, city, license number and project types. Fix your own listings directly. For listings you don't control, contact the publisher with the correct facts and a link to your source page.

The strongest evidence in AI visibility is about access, and it works by removing blocks. Run these checks before you touch content.

robots.txt, per bot

Rules apply to each crawler separately. OpenAI states that each of its settings is independent: blocking GPTBot stops training use, while OAI-SearchBot controls ChatGPT search. A blanket AI block can remove you from ChatGPT search without anyone noticing. A minimal allow looks like this:

User-agent: OAI-SearchBot
Allow: /

User-agent: PerplexityBot
Allow: /

User-agent: Claude-SearchBot
Allow: /

Firewall and CDN

A firewall or CDN can block a crawler that robots.txt allows, often with a 403. From 2026-09-15, Cloudflare blocks Training and Agent crawlers by default on ad-displaying pages for new domains and free-tier customers, while Search stays allowed. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.

Rendering and snippet controls

Vercel found that ChatGPT and Claude crawlers fetch JavaScript files but don't execute them. If your project list or license table loads only after JavaScript runs, those crawlers may see nothing. Also check for stray noindex, nosnippet or data-nosnippet tags, which keep content out of AI answers.

1. Days 1 to 5: baseline. Write five to eight prompts like the examples above for your real markets. Run them in ChatGPT, Google, Perplexity and Copilot, logged out. Record whether you're named and every wrong fact. The free checker on this site runs up to three buyer questions as a snapshot; the free brand mention check explains what it covers.
2. Days 6 to 10: access. Review robots.txt per bot, test firewall responses, compare raw HTML with the rendered page, and check Google Search Console and Bing Webmaster Tools separately for index status.
3. Days 11 to 18: one facts page. Publish a plain-text page with legal name, DBAs, HQ, license numbers by state, project types, regions, delivery methods and, if you share it, bonding capacity with a date. Make sure it's in the HTML before JavaScript runs.
4. Days 19 to 26: third-party cleanup. Correct license lookups, profiles and association directories so they match the facts page. Request fixes on owner and press pages.
5. Days 27 to 30: re-run and compare. Repeat the same prompts. Search-backed answers may change; trained-knowledge answers usually won't until a new model ships. Log what moved and what didn't.

For how tools flag errors at scale, read how AI visibility tools detect factual errors.

Can I ask OpenAI or Google to fix a wrong license or region in an answer? You can report a bad answer through each product's feedback tools, but there is no published process that guarantees a correction. The practical route is fixing the sources the engines read: your own facts page and the directories they cite. The guide on [asking AI companies to correct brand information](/questions/ask-ai-company-correct-brand-info) covers what each channel can and cannot do.

Does adding schema markup help a contractor get cited in AI answers? The evidence doesn't support that. Google says structured data isn't required for generative AI search, and an Ahrefs matched-control study found AI Overview citations fell 4.6% after pages added schema. Existing markup does no harm. It's just not a lever for AI answers. Plain-text facts on a fetchable page and correct access settings matter more.

Why does ChatGPT describe our firm as it was two years ago? When ChatGPT answers without searching, it uses trained knowledge frozen at a cutoff. As of October 2026, OpenAI lists a February 16, 2026 cutoff for the GPT-5.6 family and August 31, 2025 for the Instant path. Anything newer, such as an acquisition or new state license, only reaches answers through search, so check that OAI-SearchBot can reach your site.

Should we block GPTBot to protect our project photos and bids? That's a business choice, and it won't remove you from ChatGPT search. OpenAI says GPTBot governs training use only and OAI-SearchBot governs ChatGPT search, and each setting is independent. You can disallow GPTBot while allowing OAI-SearchBot. Don't assume blocking one bot covers the other in either direction.