AI wrote the first draft, a human checked the facts: how our content pipeline works

761 VenturesSeptember 18, 20264 min readDelray Beach, FL

A landscaper's website has 31 posts on it. The owner did not write them, but every fact in them came from the owner, and that distinction is the whole pipeline.

Short answer

Our content pipeline works like this: a model drafts a post from the owner's own notes and the live site, a claim guard checks every factual statement against a source, real project photos are attached, and the owner reads it before it publishes. The writing is the cheap part. The checking is what makes the post safe to put under a business name.

Where the facts come from

The draft does not start from a blank page and it does not start from the model's general knowledge. It starts from a pile the owner already has: project notes, the questions customers ask on the phone, the site's existing pages, the services list, the towns they work in. For a builder, that pile includes 549 project photos with dates and locations. For an advisor, it includes the firm's own explanations of what they do and what they do not do.

The model's job is to turn that pile into a readable post on one topic. It is not allowed to add facts. If the owner's notes do not say how long a kitchen remodel typically takes, the post does not say it either. It can say "it depends on the scope, and here is what changes it," which is true and useful.

The sourced claim guard

This is the part most content tools skip. After the draft is written, a second pass reads it sentence by sentence and sorts each statement into one of three bins:

  1. Sourced. The statement traces to something in the owner's pile or on the live site. It stays.
  2. Opinion or framing. "Most homeowners underestimate how much the lighting plan matters." Fine, as long as it reads as the business's view and not as a measured fact.
  3. Unsourced claim. A number, a date, a "surveys have found," a "the average project costs." It is cut or rewritten as a qualitative statement. There is no other option.

The guard is strict about numbers. If a figure is not in the source pile, it does not appear. The post can say "a growing share of our clients ask about this" but not a percentage. It can say "within a few days" but not an exact count the owner never gave us. This rule has cost us some punchy sentences. It has never cost a client an awkward conversation.

Real photos only

Stock images are banned from the pipeline. Every post is illustrated with the client's own work: the photo of the actual cabinet install, the actual back yard, the actual crew. Photos carry a caption with the project type and the town, which is also the kind of thing search engines and AI answer engines read. A real photo of a pergola in West Palm Beach says more than a page of copy about pergolas.

When a client has no photos for a topic, the post waits. We would rather publish one fewer post than publish one with a borrowed picture.

The owner reads it

Every post lands in the owner's inbox or Slack with the draft, the photos, and a short list of anything the claim guard changed. The owner reads it the way they would read a letter going out under their name, because that is what it is. They fix the tone, correct a detail only they would know, or send it back. Nothing publishes until they tap approve.

At first this feels like homework. After a few weeks it takes a few minutes, because the drafts have absorbed the owner's corrections and the pile has grown. The review step is also where the best material comes from: an owner reading a draft about drainage will often add the one story about a job that makes the post worth reading.

Weekly, on a schedule

Posts go out on a fixed day each week. The advisor site has 53 posts built this way; the landscaper site has 31. The cadence matters more than any single post, both for search and for the habit. An owner who knows a draft arrives every Tuesday reads it on Tuesday. A site that publishes steadily is one that search engines and AI answer engines return to.

Why the review is the product

Anyone can generate a thousand words about kitchen remodels. The thing a business actually needs is a thousand words it can stand behind. The draft is the cheap part of our pipeline. The claim guard, the real photos, and the owner's read are the expensive part, and they are the reason a post can carry the company's name without anyone holding their breath.

This is the content system behind every site we run, from a cabinet shop to a transit company: drafted by a model, checked against sources, illustrated with real work, and approved by the person whose name is on the door.

Questions people ask

Does the AI make up facts in the posts?

It is not allowed to. Drafts are built only from the owner's notes and the live site, and a second pass checks every statement against a source. Any number or claim that cannot be traced is cut or rewritten as a qualitative statement.

Who approves a post before it goes live?

The owner, every time. The draft, the photos, and a note of what the claim guard changed arrive in their inbox or Slack, and nothing publishes until they approve it. Their corrections also improve the next draft.

Why do you refuse to use stock photos?

Because a real photo of the client's own work, captioned with the project type and town, says more to a reader and to a search engine than a borrowed image ever could. If there are no photos for a topic, the post waits.

Share
Written by the 761 Ventures team

Operators in Delray Beach, Florida who build websites, CRMs, AI assistants and automation for businesses like yours, then write down what worked.

Want this done for your business?

Two minute intake. A real person reads every one and replies within a business day.

Start a project
OlderOpen models versus hosted models for small business softwareNewerShould your business put an AI assistant on its website?

Keep reading