Give the AI real job facts, not a request to "write about our welding". Capture a one-page job sheet and captioned photos on the day, draft the page from those facts only, check every claim, and get written client consent before publishing. End the page with a quote form asking for drawings, material, quantity and date.
Most AI case studies go wrong at the input stage. Asked to "write a case study about our welding", a chatbot produces confident filler: tolerances you never worked to, standards you do not hold, a client quote nobody gave. A page built from a short job sheet and a dozen captioned photos reads differently, because every sentence traces back to something you wrote down.
The running example, with illustrative figures, is a three-person shop that made 40 black steel bases for a furniture maker's oak tables; the same job carries through the job sheet, the prompt, the draft and the corrections. Homepage and service-page copy is a separate job, covered in AI-written website content for a welding or fabrication shop. AI website builders can produce a home page and a list of services in minutes, but they know nothing about your jobs. The case study pages are where your own evidence goes.
These pages suit shops that sell custom work, where every buyer's real question is "have you made something like this?". If most of your hours go on repeat cutting and welding for one regular customer, your website basics and Google profile may repay the time sooner.
What a buyer of custom fabrication needs a job page to prove
Someone looking for a fabricator for a one-off frame, a small batch of brackets or a repair has one worry: have these people made something like mine, and did it turn out right? A product page cannot answer that. A job page can, if it shows the brief, the specification, the method and the result in that order, with photographs that match the words.
Different buyers look for different evidence, which is why a general "our work" gallery rarely brings the enquiries you want. The table pairs five common kinds of custom work with the evidence each buyer is hunting for.
| Kind of job (example buyer) | What the buyer worries about | Evidence the page should hold |
|---|---|---|
| Repeat batches (a furniture maker ordering table bases) | Parts that vary from one to the next; late batches | Quantity, the fixture used, the dimension you checked and the result, delivery dates for each batch |
| Stainless work in food or drink areas (a specialist food importer, a craft brewery supplier) | Cleanable joints, the right material, a consistent finish | Material grade as shown on the supplier's certificate, how the welds were finished, close-ups you captioned yourself |
| Frames and platforms people stand on or walk under | Whether anyone competent checked the design and the welds | Who supplied or approved the drawing, what inspection took place and who did it. Nothing about "safe" or "compliant" unless a record says so |
| Shopfit and display pieces (an online clothing store's pop-up) | Look, finish and install time | Finished photos in place, the finish used, how long the install took, any changes made on site |
| Repairs and modifications | Whether the fix will fail again | What failed, the likely cause, what you changed, any check afterwards |
Pick the row that matches the work you want more of and build your first page for it. The evidence column tells you what to capture on the job, which is the next step.
The job sheet to fill in on the day
The job sheet is the one input that decides whether the AI's draft is usable. It takes ten minutes at the end of a job, while the details are fresh, and it doubles as your spec notes for the file. Below is the filled-in sheet for the table-base job.
| Field | Example entry | Why it is there |
|---|---|---|
| Client type | Small furniture maker; name not agreed for publication | The AI must not use a name until consent is recorded |
| Brief, in the client's words | "40 black steel bases for 1,800 mm oak tops. They must not rock on a flat floor. Two batches before our autumn show." | Keeps the page about the buyer's problem |
| Item and quantity | 40 table bases, each two end frames joined by a stretcher | Numbers a buyer can compare with their own need |
| Material and sizes | 40 x 40 x 3 mm mild steel square tube legs, 50 x 25 x 3 mm stretcher, 6 mm top plates, levelling feet | Taken from the cutting list, not memory |
| Process | MIG welding in a fixture that held each end frame square | Stops the AI guessing "TIG" |
| Critical dimension | Foot plane flat within 1 mm, checked on a flat bench with feeler gauges | The one tolerance that mattered, and how it was checked |
| Finish | Welds ground smooth; powder-coated matt black by our coating subcontractor | Subcontracted steps are named as such |
| Inspection | Welder checked every weld by eye; dimension check on all 40; one stretcher joint re-welded | What really happened, including rework |
| Timeline | 13 working days from order to the first 20; 9 working days to the second 20 | Lead time is often what the buyer wants to know |
| Problem and fix | First fixture let end frames twist by about 2 mm; added a second clamp and a stop | Shows judgement; keep it only if the client agrees |
| Result facts | 40 of 40 passed the flat-bench check; both batches delivered on the agreed dates | Facts only, no adjectives |
| Photo list | 9 photos, each with a one-line caption | Captions stay in your words |
If you would rather capture this by talking, record a two-minute voice note while you walk round the finished job and have the AI sort it into these fields. The method is the same one used to write case studies from customer interviews, with your own job notes in place of an interview transcript. Whichever way you capture it, the sheet is checked by you before it goes anywhere near a prompt.
Photos and captions the AI can work from
Take about nine photos while the job is in the shop: the cut material, the fixture set-up, the tack-up, a weld close-up before finishing, the dimension check, the finished part, the packed or delivered job, the job in use if the client allows it, and one shot with a tape or a hand for scale. Caption each in a few words as you take it. Send them to the AI in small batches, so a caption cannot drift onto the wrong picture, and do not edit the welds or filter the colours.
| Shot | Caption you write | Alt text the AI drafts (you check it) |
|---|---|---|
| 2 | Fixture holding an end frame square before tack-up | Steel end frame clamped square in a welding fixture on a workshop bench |
| 4 | Weld close-up at the stretcher joint before grinding | Close-up of a welded joint between a stretcher and a square steel leg, before grinding |
| 5 | Flatness check with feeler gauges on the flat bench | Steel table base standing on a flat bench while a feeler gauge is checked under one foot |
| 6 | Finished base after powder coating | Matt black powder-coated steel table base with levelling feet |
Alt text is the short description a screen reader speaks and a search engine reads. Google's guidance asks for useful, information-rich text and says to avoid filling the alt attribute with keywords; it also recommends short, descriptive file names. The image guidance in Google Search Central sets out its own examples, running from bad (missing or keyword-stuffed alt text) to better and best. Compare a stuffed line such as "welding welder custom fabrication steel table base welder" with "Steel table base clamped in a welding jig with the stretcher tack-welded". The second one is the only one a visitor would want read aloud. Name the files the same way: table-base-fixture-tack-up.jpg, not IMG_4471.jpg.
Three habits keep photos safe to publish. Look at the background of every shot for faces, other customers' parts, vehicle plates and site signs. Remember that many phones can store the shooting location inside the picture file, so export without it or strip it before upload. And use real photographs only: an AI-generated picture of "finished work" in a case study is a false claim about what you built.
A prompt that stops the AI inventing, and what came back
The prompt below tells the model what it may use, what to do when something is missing, and which words are off limits. Paste the job sheet and the photo captions under it. If a drawing carries a client's confidential design, leave the drawing out; the job sheet is enough.
You are drafting a case study page for a welding and fabrication shop's website.
Use ONLY the facts in the JOB SHEET and the PHOTO LIST below.
Rules:
- Do not add numbers, standards, certificates, tolerances, materials or client details that are not in the job sheet. Where something useful is missing, write [CHECK: what is missing] instead of guessing.
- No superlatives ("best", "precision-engineered", "world-class") and no claims about safety, strength or compliance unless the job sheet states them.
- Write as the shop ("we"), in plain English with the spelling "colour", "metre" and "aluminium". Short sentences.
- Describe the photos only from my captions. Never comment on weld quality.
- Call the client "a furniture maker" unless the job sheet says they agreed to be named.
Output:
1. Page title (under 65 characters) and a two-sentence summary
2. The brief (2-3 sentences)
3. Specification list: material, quantity, process, critical dimension, finish, lead time
4. What we did (4-6 short steps in order)
5. Result (facts only)
6. For each photo: a caption and one sentence of alt text
7. A closing line inviting a quote request, with no price
JOB SHEET:
[paste the filled-in sheet]
PHOTO LIST (number, my caption):
[paste the captions]
Run against the table-base job sheet, a response might come back like this, with work still to do.
40 black steel table bases, checked flat and delivered in two batches. A furniture maker asked for 40 steel bases for 1,800 mm oak tops that would not rock on a flat floor. We welded them in a fixture, checked every base on a flat bench and delivered 20 after 13 working days and 20 more nine working days later.
What we did. We cut 40 x 40 x 3 mm square tube to the cutting list and built a fixture to hold each end frame square. We MIG welded the end frames and stretchers, ground the welds smooth and sent the bases for matt black powder coating. [CHECK: the job sheet gives no weld size or wire type, so none is stated.]
Result. All 40 bases passed the flat-bench check, with the foot plane within 1 mm. One stretcher joint was re-welded after the visual check. [CHECK: the client has not agreed to be named.]
That is close to usable, and four edits still matter. The title says "checked flat", which is true, but a headline that says "sit flat" would overstate a bench test. The re-weld is honest and shows your inspection works, so keep it only if the client agrees. Both CHECK flags need an answer: delete the weld-detail note, and keep "a furniture maker" until consent arrives. Finally, convert "nine working days later" into the delivery dates from your records so a buyer can see the pace.
Now compare a lazy prompt. Ask "write a case study about a welding job making steel table bases" and a common failure is plausible extras: "precision-engineered to aerospace tolerances", "every base flat to within 0.1 mm", "certified to the highest industry standards", a client named from a pasted email. None of it was in your notes, and a buyer who works to real tolerances will spot it at once.
One realistic slip on the same job is easy to miss. Suppose the client's email, pasted in for context, says they "like TIG on visible joints". The draft then calls the process TIG, while the photo caption beside it says MIG. The mismatch only shows when the page is read with the photos open next to it. Make that your last check: read the text with the pictures beside it, caption by caption.
Claims to keep, soften or cut before publishing
Every sentence in a case study is a promise a buyer may rely on. Use one test: could you show the client a record for this on a Tuesday afternoon? If yes, keep it. If it is true but vague, soften it to what the record shows. If you have no record, cut it. The table sets out the usual suspects.
| Claim type | What an AI draft may say | Test | Action |
|---|---|---|---|
| Dimensions and tolerances | "Within 0.1 mm", "perfectly square" | Is the figure on your inspection note? | Use the recorded figure and how you measured it |
| Process | "TIG welded", "robotic welding" | Does the job sheet say so? | Correct to the job sheet |
| Standards and certificates | "Certified to the highest standards" | Can you show a current certificate that covers this work? | Name the exact certificate or cut the line |
| Testing | "Load-tested", "pressure-tested" | Is there a test record, and who ran it? | State the test and who did it, or cut it |
| Safety and compliance | "Safe for public use", "fully compliant" | Did a competent person sign it off in writing? | Cut unless a document says exactly that |
| Speed | "Fast turnaround" | Do your order and delivery dates show it? | Give the dates or working days |
| Client quote | An invented testimonial | Did the client write and approve these words? | Use only their approved words |
| Weld quality from a photo | "A clean, strong weld" | Did a qualified person inspect it? | Remove; the AI cannot inspect a joint |
The records behind these tests are worth keeping tidy anyway. If you do not yet keep procedure records and material certificates in one place, this look at AI and weld records shows what to file and what AI can and cannot do with it. A claim that you hold a certificate or follow a named procedure should be written from the document itself, copied exactly, and never from memory or from the AI's guess at what a shop like yours would hold.
Consent: ask before the photos leave your phone
Consent covers more than a name. Ask about the client's business name and logo, photos taken on their premises, anyone visible in a shot, other firms' products in the background, and any quote you want to use. Check your contract for a confidentiality clause before you ask. The person who ordered the work may not own the site or the product, so ask whoever can actually agree. For jobs on secure or sensitive sites, describe the work generically and leave the site out of it altogether.
A short email is enough for most small jobs. Here is the wording for the table-base client; replace the bracketed parts.
Subject: Can we feature your table bases on our website?
Hi [first name],
Thanks again for the table base order. We would like to publish a short case study about the 40 steel bases we made for you, with photos from our workshop and of the finished bases.
What we would show (please strike out anything you would rather we left out):
- Your business name and logo
- Photos of the bases in our workshop
- Photos of the bases at your premises
- How many we made and how long it took
We would not show prices, your address, anyone's face, or anything you ask us to leave out. You can read the page before it goes live, and you can ask us to change or remove it at any time by replying to this email.
If you are happy, a reply saying "agreed", with any changes, is all we need.
Thanks,
[your name]
[shop name]
Keep the reply with the job file and note the date. If the client wants a quote on the page, take it from their reply or ask them to write two sentences themselves; do not let the AI compose a testimonial and attach their name to it. The legal line on AI-written testimonials is worth five minutes before you publish any quote at all. This email is not legal advice, and a lawyer can look over the template once so you can reuse it.
Quote form wording that brings drawings and dates
The case study brings the visitor in; the form decides whether the enquiry is something you can price. Custom work needs facts that a plain "message" box never collects, so the form should ask for them in the buyer's language and explain why each one matters.
| Form field | Wording that works | Why you ask |
|---|---|---|
| What do you need made? | "Describe it in a sentence or two, for example '20 stainless brackets' or 'a steel frame for a 3 m stair'." | A plain-language brief to start from |
| Drawings or photos | "Upload what you have: PDF, DXF, DWG, STEP, a sketch or a photo with measurements." | Drawings turn a guess into a price; list only the formats you can open |
| Material | Mild steel, stainless steel, aluminium, not sure | Changes cost, process and finish |
| Quantity | "One-off, or how many? Will it repeat?" | Sets fixture time against batch price |
| Where it will be used | Indoors, outdoors, food or drink area, carries weight or people, near the public, not sure | Flags jobs that need an engineer's input or inspection |
| Standards or inspection | "Does a drawing, standard or inspection report have to be followed?" | Catches paperwork before you quote |
| Finish | Raw, painted, powder-coated, galvanised, brushed or polished, not sure | Subcontracted finishes add lead time |
| Date | "When do you need it, and is that date fixed?" | Separates deadlines from wishes |
| Delivery | Delivery or collection | Affects packing and price |
Compare two enquiries about the same job. The weak one: "Need a gate welding, call me." The strong one: "Need a 2.4 m double gate frame in mild steel, galvanised, hung on two 150 mm square posts; sketch attached; needed in four weeks; collection." The second can be priced or queried in minutes. The first costs you a phone call before you know whether it is a gate frame, a repair or a hobby project.
An AI summary can take the strong enquiry further by writing the three-line brief you read first. Keep the human step: a person decides whether to quote. This prompt works on a pasted enquiry.
Summarise this enquiry for a welding shop in three lines: what is wanted, material and quantity, and date. Then list what is missing before we could price it, as questions to ask the customer. Do not suggest a price. Do not promise anything on our behalf.
ENQUIRY:
[paste the form submission]
Finish with an automatic reply that sets expectations: that you price custom work from drawings, that someone will reply within your usual working days with questions or a quote, and what you need if they have not uploaded a file. When the drawing arrives, the next problem is whether AI can price a fabrication job from it. If a person answers the form's questions, store the answers in the job file so the next case study starts from facts you already hold. And if an AI tool reads submissions, say so in your privacy notice and ask a lawyer or data-protection adviser how to word it.
Where the pages live and how many to publish
Give every job its own page, linked from your services page and back again: the service page says what you do, the job page proves it. A tidy page has an answer-first summary at the top (what, for whom, how many, how long), then the brief, a specification list, the steps, the photos with captions, the result and a plain invitation to send a drawing. Put the job type and material in the page title, such as "Steel table bases in two batches | [shop name]", so the title says what the page shows.
Start with three to five pages, one for each type of custom work you want more of, then add a page when a job is worth showing. Resist the temptation to have the AI turn out a page for every town, material and weld process. Google's spam policies describe scaled content abuse as many pages produced mainly to manipulate rankings, and name using generative AI tools to make many pages without adding value as one example. Doorway abuse, meaning near-identical pages written to catch similar searches, is a separate item on the same list. A real job with real photos and a specification is the opposite of both.
Once a page is live, reuse it. A single photo and two sentences become a post on your business profile, and a longer version suits a LinkedIn post to buyers in the trade. The posting routine is covered in the Google profile tutorial linked earlier, so there is no need to write the same material again from scratch.
Awkward jobs: NDAs, rework and subcontracted finishes
Some of your best work will be hard to publish. Three situations come up often, and each has a workable answer.
A client who cannot be named. Describe the industry and the problem, show the process photos that carry no identifying detail, and leave out the finished item in place. A fixture, a jig and a close-up of a ground weld say plenty without revealing the buyer. If the contract bars any mention of the work, you cannot publish even a generic description, so do not.
A job that went wrong first. In the example, the first fixture let the end frames twist by about 2 mm. Explaining how you caught a problem and fixed it can show that your checking works, but only if the client agrees to the story being told and the facts are exact. Keep the account short and neutral. If a failure involved the client's own design, leave it out.
Work that others finished. Powder coating, galvanising, machining and installation are often subcontracted. Say so in the specification ("powder-coated by our coating subcontractor") and do not take credit for a step you did not perform. If you want to quote a result that belongs to the subcontractor, such as coating thickness, ask them for the record and cite it as theirs.
Two hours a page, and the enquiry log that judges it
The time cost is small enough to put numbers on. Swap in your own rates and job values.
| Step | Time |
|---|---|
| Job sheet and nine captioned photos on the day | 25 minutes |
| AI draft, then claim check against records and photos | 50 minutes |
| Consent email, reply and file note | 15 minutes |
| Upload, alt text check, links to and from the service page | 30 minutes |
| Total per case study | 2 hours |
At a shop rate of $65 an hour, a page costs about $130 of time. If five pages help win one extra custom job in a year worth $4,800 at a 30% gross margin, that job returns $1,440, set against $650 of page time. The arithmetic is only a prompt to be honest about your own numbers. The real test is whether enquiries improve.
For 90 days, record three things in a simple sheet. First, how many form submissions arrive and how many include a drawing or photo. Second, how many you can price without a phone call. Third, the answer to "how did you find us?" on every quote, with the case study page named where it applies. If drawings and dates arrive more often after you rewrite the form, the form is working. If visitors reach case study pages but never reach the form, move the invitation higher up the page. Keep a note of any claim you cut during checking; that list is a useful reminder of what an AI draft can add to your pages.
Questions welders ask about case study pages
Will Google penalise a case study written with AI?
Google's spam policies target many pages made to manipulate rankings that add little value to users, and they apply 'no matter how it's created'. One case study built from real job facts, your own photos and a claim check is not that. The risk is volume: dozens of near-identical AI pages. Publish a page because the job deserves it, and keep the evidence behind every statement.
What if a client says no to being featured?
Ask what they would accept before dropping the job, because an anonymous version often works. Call them 'a furniture maker', show process photos with no premises or branding in the background, and keep the specification and result. If they refuse any mention at all, leave the job off the site and keep the notes for your own records and quotes.
How long should a case study page be?
Let the facts set the length. Most jobs need a short summary, the brief, a specification list, four to six steps and the result, which usually runs to a few hundred words alongside the photos. Adjectives add length without adding anything a buyer can use. If a job needs more, add the details buyers ask about, such as lead time, finish and how you checked the work.
Should a case study show prices?
Usually not. Show the quantity, the material, the lead time and the result, then invite a quote request. A price on a page goes out of date, ignores differences between jobs and pulls in enquiries for work that was never comparable. If you want to filter out small jobs, say so on the form, for example with a minimum order line.
Can AI tell whether a weld in a photo is good?
No. A chatbot can describe what is visible in a photo, but it cannot judge penetration, porosity or compliance with a procedure. Inspection is done by a qualified person, and the page should only state what that person recorded, such as a visual check of every weld or a named test report. Keep AI out of quality judgements.
Further reads
- AI Tools for Welding Shops: Quotes, Drawings and Job Scheduling — Which AI tools help a welding shop with quotes, drawings and scheduling.
- AI Marketing Tools for Small Businesses: A Shopfitter's Case Study Pages — How another trade turns finished jobs into case study pages.
- AI Alt Text: How to Generate and Check Image Descriptions — A fuller routine for generating and checking image descriptions.
- AI Marketing for Contractors: Reviews, Case Studies and Local Posts — Reviews, case studies and posts that support each other.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Google Search Central guidance on image alt text and file names and its spam policies on scaled content abuse and doorway abuse (read October 2026).