Carpenters in ChatGPT and Gemini answers: what the shop's own site has to say

What does a homeowner ask an assistant about carpenters?

A specific job, in a specific town, with one condition attached. Nobody asks for "a carpenter"; they ask for the thing they want built and the place they live, and the assistant answers with the names of people whose pages say they build that thing there. The questions take shapes like these:

• "who builds custom built-in bookshelves in Portland"
• "deck builder in Boise who is licensed and insured"
• "custom kitchen cabinet maker near Madison, not a chain"
• "carpenter for small repairs like a rotted door frame in Tucson"
• "how much does a carpenter charge to build a mudroom bench"
• "how long does a custom closet take from quote to install"

Every question carries a specialty word: built-ins, deck, cabinets, repairs. With web search on, the pages the assistant fetched are what it has to answer that word from. If the word for your best work appears nowhere on your site except inside a photo, your site gives no answer to the question that asked for it. In Canada and the UK the same trade is often called joinery, and a shop that serves both vocabularies has to use both, once, in text.

What your site has to state

The six questions above map to six statements. Each belongs on the page where a person would look for it, in text, under a heading that says what the section answers. The table gives the statement and its form.

The captions are the step carpenters skip, and the one that matters most for this trade, because the portfolio is where the evidence lives. "Walnut built-in with hidden cable run, Sellwood, Portland" is a sentence an assistant can read and quote. The same photo with the file name IMG_4471.jpg and nothing else gives it nothing to read. That is what citability means for a carpenter: the work described in words, on the page, next to the picture.

Which schema type fits a carpenter?

There is no Carpenter type in the schema.org vocabulary. The type that fits is HomeAndConstructionBusiness, one of the LocalBusiness subtypes, and it tells a machine the kind of business before it reads a word. The properties are the ones Google's local business documentation reads: name, address, telephone, opening hours, geo and priceRange, plus areaServed for the towns and a description that names the specialties. A template:

{
  "@context": "https://schema.org",
  "@type": "HomeAndConstructionBusiness",
  "name": "Your Carpentry Shop",
  "description": "Custom built-ins, kitchen cabinets, decks and interior trim for homes in Boise, Meridian, Nampa and Eagle.",
  "url": "https://www.example.com/",
  "telephone": "+1-555-555-0100",
  "address": {
    "@type": "PostalAddress",
    "streetAddress": "100 Example Street",
    "addressLocality": "Boise",
    "addressRegion": "ID",
    "postalCode": "83702",
    "addressCountry": "US"
  },
  "areaServed": ["Boise", "Meridian", "Nampa", "Eagle"],
  "openingHoursSpecification": [{
    "@type": "OpeningHoursSpecification",
    "dayOfWeek": ["Monday", "Tuesday", "Wednesday", "Thursday", "Friday"],
    "opens": "08:00",
    "closes": "17:00"
  }],
  "priceRange": "$$$"
}

Write it as JSON-LD and mark up only what the page shows, which are the two rules in Google's structured data introduction. The description in the markup should be a sentence that also appears on the homepage. Markup that names four towns the site never mentions is a claim with nothing behind it.

What the audit found for carpenters

Carpenters had the best markup coverage of any trade with a hundred or more sites in our audit of 9,483 small business websites (40 trades, 90 metros, September 2026): 15.8% had no schema.org markup at all, against 16.1% for movers and 33.6% for real estate agents at the other end. Whoever builds carpenters' sites is adding markup.

The rest of the picture is less kind, and it is the picture for the whole sample rather than for carpenters alone. Most of the markup that existed described the website, not the business; LocalBusiness or a subtype appeared on roughly a quarter of all sites. 61.5% had no question-and-answer content of any kind. 64.7% showed no review signal a crawler could read. 3.0% rendered their content only in JavaScript, so a crawler that runs no scripts received an empty page.

The pages were not short. The median homepage ran to 880 words, and only 3.9% were under 150. If a carpenter's words tell the story of the shop, that is fine to have, and it answers none of the six questions. Length was not the problem in any trade; the form of what was said was.

The fix order for a carpentry shop

1. One name, address and phone, on every page. Identical in the footer, the about page and the contact page. Add the towns you work in as a written list. Nothing after this counts until a machine can tell whose site it is reading.
2. A page per specialty, named plainly. Built-ins, cabinets, decks, trim, repairs: each with the word in the heading, a paragraph saying what you build and how, and the range of cost and time for it.
3. Captions on every portfolio photo. What it is, the material, the town, as visible text under the image, with the same words in the alt attribute. Do the twenty best photos first.
4. License and insurance as facts. The number, the issuer and the coverage, in the footer and on the about page, in the same words.
5. HomeAndConstructionBusiness markup. The template above, filled in from the footer so the two agree.
6. The six questions, answered under their own headings. Each answer two or three sentences, the way you would say it on the phone.
7. Reviews on the page as text. Two or three real ones, first name and platform, and a rating in markup only where the rating is shown with its source.

The order is the order of dependence: a caption on a site with three addresses is evidence for nobody in particular, and markup that describes a text the site does not carry is a contradiction rather than a confirmation.

How to know whether it worked

Write the six questions above in your own towns and specialties, put each to ChatGPT, Claude, Gemini and Perplexity three times, and record whether you were named, whether you were recommended, whether you were first, and who else was. Repeat monthly on the same questions. The measurement we run for clients uses twelve questions, each asked once per reading of one assistant, Claude, before the fixes, once they are live and every month with monitoring, and the questions have to stay fixed or the readings stop comparing.

One caution on reading the result. In our probe, one buyer question per business put to Claude once named 14.9% of the businesses, and we call that directional rather than a rate because one run of one question is a coin flip. Three runs a month for six months is a trend. What the trend will show is whether the specialties you named and the captions you wrote are now being read, and that is all any measurement can show; no one outside the assistants can promise the answer.