Feed the AI your own project paperwork (brief, programme, snag list, handover photos and a client quote), have it extract a fact sheet first, then draft a one-page case study from that sheet only. You check every figure against a document and get the client's written sign-off before it goes anywhere near a tender.
That order matters because the risk with AI marketing tools for small businesses isn't dull prose, it's confident invention: a draft that says you finished early, or came in under budget, when the programme says otherwise. A tender reader who spots one wrong date starts doubting the rest of the page, so the method below puts the checking where mistakes get caught.
What a tender reader wants to see on one page
Pre-qualification forms and invitations to tender usually ask for named examples of similar work, and whoever scores them has a pile to get through. A page that answers their questions in the first ten lines gets used. One that opens with a paragraph about the firm's passion gets skimmed. So build each page around the questions a buyer is silently asking.
| Section | The buyer's silent question | What goes in |
|---|---|---|
| Fact line at the top | Is this similar to our job? | Client type, sector, floor area, programme length, working conditions, and a contract value band only if the client agrees |
| The brief | Do they understand what we need? | Two or three sentences, in the client's terms |
| The constraint | Can they cope with a job like ours? | The one thing that made it hard: trading throughout, night work, late long-lead items, tight access |
| What we built | What exactly did they deliver? | Scope as a short list: joinery, fixtures, flooring, lighting, signage |
| Programme | Do they finish when they say? | Planned and actual dates, stated plainly. If it ran late, say why and what you did |
| In the client's words | Would they hire them again? | One verbatim quote, with permission |
| Photos | Is it as good as they say? | Before, during, after and two or three details |
| Credits | Who else was involved? | Designer, main contractor, key suppliers |
To show the method end to end I'll follow one illustrative job: a 140 square metre independent bookshop refit for a six-person shopfitting firm.
Stage 1: build the project file before you open an AI tool
AI can only be as accurate as what you give it, so gather the documents first. For most jobs the list is short: the tender brief or scope summary, the current programme revision (marked as such) with the as-built dates, the variation log, the snag list (also called a punch list) with closing dates, the client's handover email, the joinery drawing schedule for quantities, and the photo folder. Do this once per job, ideally in the week of handover while people still remember the details.
Before anything leaves the building, run the permission check. Look at the contract for a confidentiality or publicity clause. Ask whether the client has a launch embargo or brand rules about how their stores can be shown. Scan the photos for people, stockroom shelving, safes, cash offices and security equipment. Any of these can end a client relationship faster than a late delivery.
The bookshop's file, in short: a 140 square metre ground-floor refit on a six-week programme (planned finish Friday of week six, shop reopening the next day), with the shop trading throughout. Three installers and two apprentices worked on site at the peak. The joinery drawing schedule lists 22 shelving bays, about 38 metres of timber shelving and one cash desk, and the scope summary adds a new track lighting layout. Joinery installs ran in the evenings for three weeks so the shop could open each morning. The snag list had 31 items, all signed off four days after handover. The client's handover email thanked the team for keeping the shop open.
Documents hold the facts but not the story. Spend ten minutes with the site manager on the day of handover and record the conversation on a phone, with their agreement and with only your own people in the room. Five questions are enough. What was the hardest thing on this job, and what did you do about it? What did you change on site that wasn't in the drawings? What did the client praise, in their own words? What would you do differently on a similar job? Which photo shows the work you're proudest of? Add the transcript to the project file as one more document, and the "constraint" section of the page stops being generic because the AI has something real to work from.
Stage 2: have the AI extract a fact sheet, not write a story
This is the step most people skip, and it's the one that stops the invented details. Paste the documents (or upload them, if your plan allows files) and ask for an extraction only. A Project in ChatGPT or Claude, or a Gemini Gem (becoming a skill from November 2026), is a good place to keep the prompt so you reuse it unchanged.
You are helping a six-person shopfitting firm write a case study.
Use ONLY the documents below. Do not use outside knowledge and do not guess.
Documents: (1) scope summary, (2) programme revision F with planned
and actual dates, (3) snag list with closing dates, (4) the client's
handover email, (5) photo log (numbered, one line each).
Produce a fact sheet with these headings: Client type; Sector; Floor area;
Programme (planned and actual start and finish); Working constraints;
Scope (list); Materials and finishes (only those named in the documents);
Numbers (every quantity, count or duration, each with the document it
came from); People and credits; Client quote (verbatim, with source and
date); What the documents do NOT tell us.
If a fact appears in only one document, mark it (single source). If two
documents disagree, show both and flag the conflict. Do not use praise
words such as "stunning", "state-of-the-art" or "award-winning".Here is the kind of sheet that comes back (illustrative), trimmed to the lines that matter:
Client type: independent bookshop (retail). Floor area: 140 sq m.
Programme: planned finish Fri week 6 (programme rev F). Shop
reopened Sat week 6 (handover email).
Working constraints: shop trading throughout; joinery installed
6 pm to midnight for three weeks (programme rev F, single source).
Numbers: 22 shelving bays (drawing schedule); about 38 m of timber
shelving (drawing schedule); 31 snag items, closed in 4 days (snag list).
Client quote: "Keeping the shop open the whole time made the
difference." (handover email)
CONFLICT: programme rev F gives handover as Fri 12 June; the snag list
is headed "handover inspection 15 June".
The documents do NOT tell us: contract value, final cost against
budget, safety record.Notice the last two lines. The conflict flag sends you to the signed handover certificate: perhaps 12 June was the handover and 15 June the follow-up inspection, and the page should state only the date the certificate shows. The "do not tell us" list is your guard against invention: if contract value or budget performance isn't on that list because you haven't supplied it, the draft must not mention it.
Stage 3: draft the one-pager from the fact sheet only
Start a fresh chat, or a fresh message, and give the AI only the fact sheet, not the raw documents. That keeps it from re-reading something you've already settled.
Using only the fact sheet below, write a one-page case study for our
website, 350 to 450 words, in plain English with our usual spelling.
Structure: a one-line summary (client type, floor area, programme);
"The brief";
"The constraint that shaped the job"; "What we built"; "How the
programme went" (planned against actual dates, no adjectives);
"In the client's words" (verbatim quote only); "Credits".
Do not add any fact that is not in the fact sheet. If you want to
add something, list it under "Questions for the firm" at the end.
[paste the fact sheet]A typical first draft (illustrative) includes sentences such as: "Our award-winning team delivered a state-of-the-art refit, completed two days ahead of schedule and under budget." Every claim in it is a problem. The fact sheet holds no award, nobody said "state-of-the-art", the programme says the shop reopened on the planned date, and no document mentions budget. The AI tends to invent these because case studies it has seen are full of them.
The fix is to cut the sentence and write the true one: "The shop reopened on the planned date, Saturday of week six, after six weeks of work done mostly in the evenings so it could keep trading." It's plainer and a buyer can verify it, which is why it persuades. Keep the AI's structure and your own facts, and read the result aloud once to hear any phrase that doesn't sound like you.
A late job is the real test of honesty. Suppose a fashion-store refit slipped three working days because the glass for the changing rooms arrived a week late. Say exactly that: "Planned finish Friday of week five; actual finish the following Wednesday. The glass delivery moved by a week; we re-sequenced lighting and flooring ahead of it and recovered four of the seven days." A plainly stated delay with a recovery action tends to read as more credible than a record that looks too clean, so tell the AI up front: "If the actual finish is later than planned, say so, give the cause from the documents and say what we did about it."
Stage 4: check every claim against a document
Print the draft, or keep it beside the fact sheet, and tick each factual statement against the document it came from. A table makes the habit mechanical. Here is the bookshop draft after its check:
| Claim in the draft | Source document | Status |
|---|---|---|
| Shop kept trading throughout the refit | Handover email; programme note on trading hours | Keep |
| 22 shelving bays, about 38 metres of timber shelving | Joinery drawing schedule | Keep |
| Reopened on the planned date | Programme rev F; client's handover email | Keep, once the certificate settles the handover date |
| 31 snag items closed within four days | Snag list, closing dates | Keep |
| Completed two days early | None; programme shows planned finish | Remove |
| Under budget | None | Remove |
| Award-winning team | None | Remove |
The rule is blunt: a claim with no document goes. If a figure matters, such as contract value or final account against budget, find it, get permission to publish it, and add it to the fact sheet first. Never let the draft be the first place a number appears.
For a wider checklist of claims that AI-written pages tend to get wrong, see the website content claims check. It's useful here because case study pages are exactly where bold, unprovable claims creep in.
Stage 5: choose, caption and describe the photos
Eight to ten photos is enough. A buyer wants proof of scale, quality and tidiness, not volume. A good set for the bookshop would be a wide before shot and a wide after shot from the same spot, one shot mid-works showing floor and stock protection, a close view of a joinery junction, the cash desk, the lighting layout, and one picture of the team with their consent.
AI is useful for the captions and alt text. ChatGPT takes images of up to 20 MB each, and Claude takes up to 20 files per chat with each image up to 10 MB, so a set of eight fits in one go on either. Number the photos in your prompt rather than relying on file names, and describe the ones that need context:
Photos 1 to 8 are attached in order. For each, write alt text of
20 words or fewer that describes what is visible, in plain English,
no "image of", no brand names unless I give them.
Then write a caption of one sentence for photos 1, 2, 5 and 7 that
says what a tender reader should notice.
Photo 3 is mid-works, floor protection down, shop open beyond the screen.
Photo 7 is the cash desk, solid oak top, built in our workshop.A usable result (illustrative) is: "Photo 5. Timber shelving bays curve along the rear wall, with a rolling ladder and new track lighting overhead." Check it against the picture. If the alt text claims "oak" and the shelving is ash, fix it. Alt text that is accurate, short and specific serves screen reader users and search engines alike.
Keep edits boring: crop, straighten, correct brightness. Do not remove a bin, replace a sky or let a tool tidy the room digitally. Google's guidance for Business Profile photos says an image should represent reality, with no significant alterations or excessive use of filters or AI, and a client who visits the finished shop will see any difference. Move the bin before you take the picture.
The bookshop page as it stands before sign-off
After the claims check, the draft has lost three sentences and gained one honest line about the dates. This is the version that goes to the client:
Independent bookshop refit: timber shelving installed while trading. Retail, 140 square metres, six-week programme, shop open throughout.
The brief. The owner wanted the ground floor rebuilt around fixed timber shelving and a single cash desk, without closing to customers.
The constraint that shaped the job. Fitting 22 shelving bays is noisy, dusty work, so we installed it between 6 pm and midnight for three weeks and the shop could open every morning.
What we built. 22 shelving bays, about 38 metres of timber shelving in total, one cash desk and a new track lighting layout. Three installers and two apprentices worked on site at the peak.
How the programme went. Planned finish: Friday of week six. Handover signed: Friday of week six. The shop reopened the next morning. The snag list had 31 items, all closed four days after handover.
In the client's words. "Keeping the shop open the whole time made the difference."
Credits. Shop design by [designer]. Lighting supplied by [supplier]. Photography by [photographer].
It runs to about 170 words. The prompt asked for 350 to 450, and the gap is the point: with only verified facts available, the page is shorter than the AI's first attempt, and that is fine. A buyer prefers 170 words they can trust to 400 they have to discount. If the page feels thin, the fix is another fact in the file (the debrief is the usual source), not another adjective.
Stage 6: get the client's sign-off in one email
Ask once, list what you need, and attach the draft. A vague "OK to feature you?" gets a vague yes that won't cover the PDF in your tender packs.
Subject: Can we feature [project] on our website? Draft attached
Hi [first name],
Thank you again for [project]. We'd like to write it up as a short
case study for our website and for tender documents. A draft and
eight photos are attached.
To publish we need your approval on:
1. The facts and figures in the draft (please correct anything).
2. Your quote: "[quote]".
3. Naming: [company name], or "a [sector] business" if you would
rather stay anonymous.
4. The photos. We have left out [stockroom, security equipment,
staff faces]. Tell us if anything else should go.
5. Using the page and a one-page PDF in tender submissions to
other clients.
Reply "approved" to confirm, or send changes. We will not publish
until we hear from you, and you can ask us to change or remove it
later.
Thanks,
[your name]Keep the reply with a log line: date, who approved, what the approval covers, the version they saw, and a date to revisit. If the client changes a figure, update the fact sheet, not just the draft, so the next version doesn't reintroduce the old number. The freelancer's version of the same approval step, with more example wording, is in getting client approval on an AI-drafted case study.
Stage 7: publish the page, the PDF and the tender-pack version
Publish three versions of the same facts. The web page is the full story with all the photos. A one-page PDF in a fixed layout, with the fact line in a band across the top, goes into tender packs and email follow-ups. A short LinkedIn post points people to the page. Canva is a quick way to make the PDF: its Free plan includes up to 20 uses of its Standard or Premium AI tools a month, and Pro is $144 a year.
The third version is the tender-pack edition. Add a short box at the top headed "Relevant to this tender" and rewrite it each time using the invitation's own criteria. Suppose an invitation asks for three projects from the last five years, each over 100 square metres, delivered while the premises were occupied, with a client reference. Paste the criteria and a table of your library, and ask which pages fit. Against a library of six pages, the answer might look like this:
| Case study | Floor area | Occupied during works? | Fits the criteria? |
|---|---|---|---|
| Independent bookshop | 140 square metres | Yes, trading throughout | Yes |
| Pharmacy refit | 110 square metres | Yes, trading throughout | Yes |
| Fashion store | 220 square metres | No, closed for the works | No |
| Restaurant fit-out | 180 square metres | No, closed for the works | No |
| Coffee shop | 95 square metres | Yes, trading throughout | No, under 100 square metres |
| Office reception | 60 square metres | Yes, staff in the building | No, under 100 square metres |
Only two of the six qualify. The honest response is to submit those two, say plainly that the nearest miss is the 95 square metre coffee shop, and explain why it is relevant, rather than stretch the criteria. The table also shows a gap in the library: you have no occupied job over 100 square metres beyond these two, which tells you what to document next. Have the AI draft the relevance lines, then read each against its case study to be sure it claims nothing the page doesn't show.
On the website, give each page a title that names the sector and the work, such as "Independent bookshop refit: timber shelving installed while trading". Write each from its own facts. Google's spam policies list generating many pages with generative AI without adding value for users as scaled content abuse, so one real page per real project is the safe pattern, and cloning a case study with different nouns is not.
If you already do bigger bids, construction tender writing with AI shows how to check a response against the scoring scheme, and AI marketing for a construction company covers the equivalent for larger contractors.
Three hours for the first case study, then less
The first case study takes about three hours. Gathering the project file takes 45 minutes, the fact sheet and its check 30 minutes, the draft and edit 40 minutes, photos and alt text 30 minutes, the sign-off email 10 minutes, and publishing the page and PDF 25 minutes. That is 180 minutes. Once the file habit is in place and the prompts are saved, the sixth should come down to roughly 90 to 120 minutes (my estimate), because gathering and publishing are the only long parts.
The software bill is small. A chat assistant is free on its basic plan or about $20 a month at list price on ChatGPT Plus or Claude Pro (Claude is $17 a month on annual billing). Canva Pro adds $12 a month on its $144 yearly price. A full toolkit therefore costs between nothing and about $32 a month, and the paid plans only earn their place if you're writing more than a couple of pages a month. Plan details change, so check the vendors' pages before you pay.
An embargoed layout, a superseded programme, a generous caption
Three plausible slips, so you can recognise them before they cost you a client:
- The embargoed layout. A page goes live with a photo of a new store layout a week before the client's launch. The client asks for it to come down and doesn't invite the firm to tender again. The permission check in stage one exists for this, and so does the single question "has launch happened yet?" in the sign-off email.
- The superseded programme. The AI picks up the first programme revision, which showed a finish in week eight, instead of revision F, which the client agreed after bringing the opening date forward, and credits the firm with finishing "two weeks early". It reads convincingly. Marking the current revision in the project file, plus the claims table, is the defence; a tender reader who compares your page with the client's reference will find the mismatch.
- The generous caption. The AI describes timber shelving as "solid oak" because the prompt mentioned oak once. It is ash veneer on ply. Captions get less checking than paragraphs, so check them with the same rule: if the photo and the specification don't show it, the word comes out.
How to tell whether the pages are earning invitations
Don't invent a return figure; count what you can. Keep a simple tally of every tender or pre-qualification response that reuses a one-pager and how long assembling the references section took. Add "which project page did you look at?" to your enquiry form. Note when an invitation arrives after you sent a link to a case study. After a few months the tally will show which pages do the work, and which sector you have too little evidence in. That tells you what to document on your next job, which is worth more than any figure a vendor could quote you.
Case study questions shopfitters ask
Should I name the client or keep them anonymous?
Name them only with written approval. Many fit-out contracts carry confidentiality terms, and retail clients often have launch dates and brand rules. If approval isn't coming, offer an anonymous version that still carries the facts a buyer scores: sector, floor area, programme and working conditions. A 'national fashion retailer, 400 square metres, trading throughout' page is still a usable reference.
Can I use AI-enhanced or AI-generated photos?
Not for the finished work. Google's own guidance for Business Profile photos says an image should represent reality, with no significant alterations or excessive use of filters or AI, and a tender reader may visit the site. Crop, straighten and correct the brightness, and keep AI for captions and alt text. If you show a concept render, label it as one.
How many case studies do I need before this pays off?
There is no magic number. Pre-qualification forms often ask for three relevant projects, so having six or eight to choose from, spread across the sectors you want, makes each form quicker to answer. Start with the two jobs you are proudest of and whose clients are most likely to say yes, then add one after every handover.
What if the client won't give a quote?
Ask whether you can quote a sentence from their handover email, with their approval. Failing that, offer a named referee who is willing to take a call, or publish the facts without a quote. A page with accurate programme dates and photos still does its job; a quote you wrote for them doesn't.
Will Google penalise AI-drafted case studies?
Google's spam policy targets many pages generated mainly to manipulate rankings, and its text says unoriginal, low-value content at scale is the problem however it's created. One page per real project, built from that project's own facts and photos, is a different thing. Don't clone a page twenty times with the nouns swapped.
Further reads
- AI-Generated Welder Website Marketing: Case Study Pages for Custom Work — A parallel method for custom metalwork case study pages.
- AI for Construction Contractors: Bid Invitations and Follow-Up — Follow up bid invitations once your pages are live.
- Construction Photo Documentation App: What AI Adds, and Which Plan — Capture dated site photos you can use later.
- Sign Company Marketing: Case Studies and B2B Emails With AI — A job sheet, case-study prompt, consent email, local page test and three outreach emails a sign company can use this month.
- Architect Website: Project Descriptions and Service Pages Written With AI — A copy workflow for architects: project descriptions from verified facts, service pages by project type, honest area pages and an FAQ on fees and process.
- 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 spam policies (scaled content abuse); Google Business Profile Help (photo guidance); OpenAI help pages (image uploads); Anthropic vision documentation and help centre (upload limits); Canva pricing page; ChatGPT and Claude list prices.