AI Marketing for Architects: Case Studies, Website Copy and LinkedIn

Give the AI material only you have (site-visit voice notes, the design statement, the brief, the client's own words) and a list of phrases to ban, then edit for the decisions it flattens. Generic copy comes from generic inputs; specific inputs and a firm editing pass produce copy that sounds like your practice.

Architects have a particular problem. The best marketing material, the reasoning behind each design decision, lives in your head and in meetings, not in any document an AI can read. Ask a chatbot for copy about "a modern rear extension" and you get the paragraph every practice's website already has. AI marketing for architects works when the reasoning goes in first. The routine below captures it in about 25 minutes of talking, then uses AI to shape it into three formats: a case study, service-page copy and LinkedIn posts.

Follow me on Instagram@sagnikteaches

Why AI copy about buildings reads the same everywhere

A language model writes what's most likely to follow, and the most likely sentence about a house extension is the one many practices have already published. The usual adjectives arrive on cue. The tell isn't the individual word; it's that nothing in the paragraph could be checked against the building. Here is a typical unprompted result, shown as an example of what to avoid:

Connect on LinkedInSagnik Bhattacharya

Nestled on a quiet street, this bespoke rear extension seamlessly
blends contemporary design with the home's original character,
creating a light-filled living space that is perfect for modern
family life.

Now the same project written from the architect's own notes (an illustrative rear extension, used throughout):

Subscribe on YouTube@sagnikteaches

The garden faces north, so daylight was the whole problem. We lowered the kitchen floor by 250 mm to gain ceiling height without raising the roofline, put a rooflight over the island instead of a glazed gable, and used one steel beam where the clients had expected two columns. The extension is 4.2 metres deep. Any deeper and the new room would have lost the daylight that made it worth building.

Run a quick test on any paragraph: delete the project name and ask whether it could describe another project. If it could, rewrite it. A second test is to count the checkable facts, meaning dimensions, orientation, materials, decisions and constraints. The first paragraph has none. The second has six. I'd send anything under about ten checkable facts per 500 words back for another pass, and that figure is a working rule of mine, not a standard.

Build a voice sheet from three things you've already written

A voice sheet is a one-page set of rules describing how your practice writes. It goes into every prompt, so the AI starts from your habits and not from its own defaults. Gather three samples: a design statement from a planning or tender submission, an email in which you explained a decision to a client, and a past post or talk. Then ask for rules, not a summary.

Below are three pieces of my writing. Describe my voice as 8 to 10
rules another writer could follow: person and tense, typical sentence
length, how I explain a decision, how I handle constraints and
uncertainty, words I use, and words I never use. Quote a short example
for each rule. Do not flatter me; describe only what is in the samples.

[paste sample 1] [paste sample 2] [paste sample 3]

A filled-in result (illustrative) after editing, with a ban list added by hand:

VOICE SHEET - [practice]
1. First person plural ("we"). Past tense for decisions, present tense
   for how the building works.
2. Sentences average about 15 words. None over 30.
3. Every decision comes with its reason ("because...").
4. Name materials and dimensions. No adjective a client wouldn't say
   aloud.
5. State constraints plainly ("the budget ruled out...").
6. Never use: nestled, bespoke, timeless, stunning, light-filled,
   thoughtfully designed, curated, seamless, sleek, "passionate about".
7. No award, saving or performance claim unless I paste the source.

Keep the sheet where you can reuse it unchanged: a Project in ChatGPT or Claude, or a Gemini Gem (becoming a skill from November 2026). The ban list in rule six is the most useful line. Add a word the moment you catch the AI using it, and the next draft arrives without it.

For the basics of home, about and services pages, writing your website copy with AI covers the standard structure. The architect-specific part is getting your reasoning on to the page.

Capture what the AI can't know: a 25-minute voice note

After handover, or after a site visit, record yourself talking through the project for about 25 minutes. Dictate on your phone and keep to your own voice: don't record client or contractor conversations without their agreement. Answer eight questions in order, and don't tidy as you speak.

1. What did the client ask for, in their own words?
2. What was the real problem underneath the brief?
3. Which constraint shaped the design most?
4. Name three decisions and the reason for each.
5. What changed on site, and why?
6. What would you do differently?
7. What does the building do that photographs can't show: the light at four o'clock, how it feels to cross the room?
8. What did the client say when they first walked in?

The transcript is the raw material. Here is an extract from the rear extension's:

...so the brief was really "somewhere we can all be in one room without
being on top of each other". Underneath that they were eating standing
up in a galley and the garden was a step down they never used. North
facing, so daylight is the whole problem. We went 4.2 deep, any deeper
and you lose the light at the back. Dropped the floor, we wanted three
hundred but the footing was shallower than the survey said, so it's
two-fifty. Rooflight over the island instead of a glazed gable because
the gable gave them glare at the table in the evening. One steel
instead of two columns; the engineer wasn't keen, then drew it in an
afternoon...

Before any drafting, ask the AI to interrogate it: "Ask me five follow-up questions about anything unclear, missing or vague. Don't draft yet." A good set (illustrative) includes: "You said the gable gave glare. At what time of day, and from where?", "What did the engineer need to change to draw the single beam?" and "What did the client do in the old galley that the new layout fixes?" Answer each in a sentence. Those answers are where the specifics come from.

One voice note to one case study, with the numbers

Here is the whole chain for the same rear extension, a family of four, about 24 square metres of new floor (4.2 metres deep by 5.6 wide). The voice note runs 25 minutes and produces a transcript of roughly 3,400 words. The AI extracts a fact sheet of about 40 items, each tagged with the line it came from. Ten minutes of checking removes two it misheard, and another ten go on the follow-up questions. From the fact sheet and the voice sheet it drafts about 550 words in ten minutes. You edit for 50 minutes, ending at 480 words.

The first draft contains five checkable facts. After your edit it has fourteen, because you've put back the reasons: why 250 millimetres and not 300, why a rooflight and not a gable, why one beam. The structure below keeps each section tied to its source.

A typical weak sentence from the first draft: "The design maximises natural light through a carefully positioned rooflight." Your edit replaces it with the reason from the transcript: "We put a rooflight over the island rather than a glazed gable, because the gable gave the clients glare at the table around five o'clock in winter." The second is longer, and it's the one a prospective client is more likely to remember, because it shows you saw their problem. Similar tutorials that go deeper on the interview side include writing case studies from customer interviews with AI.

Service pages: write from the questions clients actually ask

Don't start a service page with a blank prompt. Collect 20 real questions from enquiry emails and first calls, and give them to the AI to cluster. The clusters tell you what the page must answer, in the order people ask, which is usually: what happens, how long does it take, what will it cost me, what do I need to do. Then build from your own records, not from the AI's general knowledge.

A solid service page for extensions and alterations has these parts: who it's for; what happens at each stage; what we need from you; how long each stage takes in our last ten projects; how our fee works (the structure, with figures only if you publish them); what we don't do; projects like yours; and the questions people ask. For the durations, pull dates from your own project records and state them as a range with the source: "from our last ten extensions". Never let the AI supply a typical timeline.

Approvals deserve care. Requirements differ by place and change over time, so a sentence such as "we check which approvals apply to your project at the first meeting" is better than a rule the AI half-remembers. For the full set of project and service page sections, see architect website project descriptions and service pages written with AI.

Here are 20 questions clients have asked us (below) and our voice sheet.
1. Group the questions into 5 to 7 themes and rank them by how often
   they appear.
2. Propose a page outline for "Extensions and alterations" that answers
   the themes in that order.
3. Write the section "What happens at each stage" in our voice. Leave
   [DURATION FROM RECORDS] wherever a time or cost is needed. Do not
   supply typical durations, fees or legal requirements.

Home and about pages: two short jobs that carry the practice

A home page needs one sentence saying what you do and for whom, and an about page needs the people. Ask the AI for ten versions of the home-page sentence, giving it your voice sheet, your list of services and three real client questions, then pick one and cut it again. A before and after shows the gap. The AI's default: "We are an award-winning architectural practice dedicated to creating inspiring, sustainable spaces." The version built from the practice's own facts: "We design extensions and alterations for family houses, and we start with the problem underneath the brief." The second tells a visitor whether they're in the right place, and it contains nothing the practice can't back up.

For the about page, record a five-minute voice note on why you do this work and what kind of job you'd happily do again, and ask for 200 words from it. Then keep only the sentences that nobody else could have said. People hire architects they trust, so name who they'll deal with, what each person does on a job, and who answers the phone.

A case study for a project that isn't built yet

You won't always have a finished, photographed building when you need marketing. A project at the approvals stage can still make a useful page if it's honest about its status. Keep the same structure and swap two sections: replace "What changed on site" with "Where the project stands", and use two drawings the client has allowed you to publish, such as a plan and a section, in place of photographs. Put the status in the fact strip, for example "Approvals granted; start on site planned for the spring", but only if that is true and the client is happy for it to be public.

A concept visual can sit alongside if the caption says "concept image". When the building is finished, return to the same page, add the photographs and the site story, and keep one page that grows rather than two that overlap. The voice note for the second pass takes ten minutes, because you already have the first half.

LinkedIn posts from a site visit: one idea each

A site visit is a post factory if you capture it properly: three photos and a 60-second voice note. Ask the AI for three possible angles: a decision, a detail and a lesson. You choose one, it drafts roughly 120 words, and you cut. Say the note from the day the steel went in runs: "Steel's in, and it's the first time you can feel the space. Standing on the old floor line you're looking at the roof edge; standing on the new one you're looking at the garden. We drew a 300 drop and the footing was shallower than the survey, so it's 250. Floor to beam is 2.6 metres. The rooflight goes in next week."

Angles it might offer: "What 250 millimetres does to a room", "Why one steel beat two columns", and "The footing the survey missed". Take the first. A result after your edit (illustrative):

Standing on the old floor line, you're looking at the roof edge. Standing on the new one, 250 mm lower, you're looking at the garden.

We saw that on site this morning, with the steel just in. We had drawn a 300 mm drop; the footing was shallower than the survey showed, so we settled on 250. Floor to underside of beam is now 2.6 metres.

It's less than we hoped for, but the view from the island is the point of the room. The rooflight goes in next week, over the island rather than a glazed gable, because the gable gave the clients glare at the table in winter evenings.

That's about 570 characters, well inside the 3,000 LinkedIn allows. What I'd strip from a typical AI draft of this: an opening line of "Excited to share", a closing question about "dream spaces", and a string of hashtags. The first two lines carry the point because that's what appears before the "see more" cut. LinkedIn's own help text says posts can contain up to 3,000 characters and points longer pieces to articles. For an owner's voice and style on the platform, see LinkedIn posts with AI without sounding like AI.

You can also turn the finished case study into a document post. LinkedIn accepts PDF, PowerPoint and Word files up to 100 MB and 300 pages, wants PDFs with layers flattened, shows videos and animations as static images, and asks you to add a title. Six to eight pages is plenty: ask the AI to split the case study into page-sized chunks, one idea per page, then check that each page makes sense when seen alone.

The editing pass every draft gets

Run the same six checks on every draft, in this order. On a 500-word piece the whole pass takes 20 to 30 minutes, and it's where the practice's voice comes back.

1. Delete the project name. Could the paragraph describe another project? If so, rewrite it.
2. Count the checkable facts. Under about ten per 500 words goes back for another pass.
3. Search for your ban list and for any adjective a client wouldn't say aloud.
4. Check every number, name and date against the document that holds it.
5. Read it aloud once, and replace any sentence you'd be embarrassed to say to a client.
6. Ask who would mind if it were wrong, whether that's the client, the engineer or the photographer, and send them the paragraph that concerns them.

Permission, copyright and confidentiality before publishing

Clients have a say in how their home appears, photographers usually own the copyright in their pictures unless a contract says otherwise, and consultants and contractors expect credit. Before you publish, fill in a permission table for the project. This one is for the same extension:

Two more points. First, keep client-identifying documents out of consumer AI tools unless you have checked the terms and your client contract. Your practice's AI policy should say what may be pasted where, and an AI policy for architects gives clauses you can adapt. Second, the legal side of copyright and client data is covered in copyright and client data risks of AI in architecture practices. Ask a solicitor for anything specific to your contracts.

Claims to verify before anything goes live

Architecture marketing attracts claims AI will happily write: awards, performance figures, savings, sustainability. Each one needs a source you can produce on request.

• Awards and shortlistings. Name the award, the year and the category, or cut it. A draft that says "award-winning" about a project that was only entered has made an error you will be asked about.
• Energy and environmental performance. Quote the measure and where the figure comes from: a calculation, a test result, a certificate. Some places now restrict generic environmental claims such as "eco-friendly" in consumer marketing, so name the specific measure and its number instead.
• Cost and programme. Publish figures only with the client's approval and only from the final account or programme. An AI will not know them and may invent them.
• Quotes. Verbatim, approved and dated. A quote the AI tidied is a quote you've changed.
• Concept images. A render isn't a building. Label it as a concept, put it in a separate post from built work, and never place it beside a photograph of a different outcome.

A leaked name, a borrowed award and an averaged timeline

Each is plausible, and the cheap defence is in the checks above.

• The leaked client name. A client's surname appears in a pasted email, and the AI weaves it into the draft. It is caught in the final read, but the PDF export still carries the project's working title, which includes the surname, as its file name and document title. Replace names with [client] before pasting, and check file names and document properties before anything is sent.
• The borrowed award. The practice's awards page mentions a different project, and the AI attaches "shortlisted for a national award" to this one. The claims list catches it, because the award has no source against this project.
• The averaged timeline. The AI writes "from brief to completion in twelve months" by averaging durations from several projects in the files it was given. Durations come from the project's own records or they come out.

A monthly rhythm: 3.5 hours, one case study, four posts

The routine fits a small practice because the heavy work happens once per project. In week one, record the voice note (25 minutes), have the AI extract the facts and check them (25 minutes) and choose photos (10), which is 60 minutes. In week two, draft and edit the case study (60 minutes) and chase permission (15), a total of 75. In week three, write four LinkedIn posts from notes you already have, at 15 minutes each, which is 60. In week four, spend 15 minutes looking at what happened. That is 210 minutes, or three and a half hours.

Refresh one service page a quarter, using new client questions and your latest project durations; that adds about 90 minutes in the quarter. Measure with what you can see: whether enquiries mention a project or a post, whether tender teams ask about a case study, and which posts prompted replies from people who commission buildings. After a few months you'll know which format earns the conversation, and the voice sheet will have a longer ban list and a sharper description of how your practice writes.

Questions architects ask about AI marketing

Should I tell people AI helped write my marketing?

Rules differ by place and by profession, so check your professional body's guidance on advertising. The safe habits are to treat the final text as yours, because you are answerable for it, to label concept renders and any AI-generated images clearly, and never to present a generated image as built work. Accuracy matters more than the tool.

Can I paste client drawings or briefs into ChatGPT?

Check your client contract and your practice's AI policy first. Consumer ChatGPT and Claude plans let users switch off use of their chats for model training in privacy settings, while business plans do not train on business content by default. Even so, strip names and addresses, prefer written notes to drawings, and use an approved business plan for anything confidential.

How often should an architect post on LinkedIn?

One considered post a week beats daily filler, and that is my suggestion, not a platform rule. Four posts a month from four site visits or decisions is a realistic rhythm for a small practice, and each can come from a one-minute voice note. LinkedIn allows posts of up to 3,000 characters, but most of yours will be far shorter.

Does Google penalise AI-written website pages?

Google's spam policy targets scaled content abuse: many pages generated mainly to manipulate rankings, and it says unoriginal, low-value content at scale is the problem however it is created. A handful of pages built from your own project facts, voice notes and client questions is a different thing. Make each page say something only your practice could say.

What if the client won't allow photos?

Write about the design problem instead of the building. With permission, use diagrams, section sketches or a detail drawing, and describe the project generically: a family house, a north-facing garden, a tight budget. You can also use a project of your own, or an unbuilt concept clearly labelled as a concept.

Further reads

• AI Marketing for a Construction Company: Case Studies, Bids and LinkedIn — The contractor's version: case studies, bids and LinkedIn.
• Architecture RFP Response: Reuse Past Proposals With AI, Without Copying — Reuse past proposals in bids without copying.
• AI Tools for Architects 2026: A Toolkit From Brief to Handover — AI tools across project stages, from brief to handover.
• Claude vs ChatGPT vs Gemini for Architects: Which to Pay For — Which assistant to pay for in a small practice.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: LinkedIn Help ('Post and share updates' for the 3,000-character limit; 'Upload and share documents' for the 100 MB and 300-page limits); Google Search Central spam policies (scaled content abuse); OpenAI and Anthropic plan and privacy pages (model-training settings).