AI writing sounds generic when the agent receives a generic assignment. Asking it to "sound human" changes the surface, but it does not add facts that belong to the business.
We saw that gap inside seodraft this week. The earlier post on this topic was READY with a passing gate but had a null brief because it predated the current editorial workflow. In a separate run, the gate rejected three concrete mistakes before the draft reached human review. This article explains the process we use now: profile, evidence, brief, real links, and a gate.
- Generic AI writing starts with generic input
- Give the agent a business profile it can use
- Collect evidence before asking for the body
- Save the brief before writing the article
- Use real internal links
- Make a gate reject observable mistakes
- Review the article as the founder
Generic AI writing starts with generic input
AI writing sounds generic when the agent gets broad instructions and no information that belongs to the business. Any capable model can complete "write a guide about SEO" with accurate definitions and paragraphs that could appear on a hundred sites.
Search results for this topic lean heavily on humanizers, prompt lists, and style advice. Those tools may change sentence rhythm or vocabulary. They cannot add a measured result, a failed experiment, or a decision that the company actually made.
The old post in this workspace showed the limit. The earlier post on this topic was READY with a passing gate but had a null brief because it predated the current editorial workflow. It could satisfy surface rules without recording its angle, evidence, or reason to exist.
The goal is not to hide the agent. The goal is to give it specific material and a process that lets a founder inspect how the material was used.
Give the agent a business profile it can use
A useful profile defines the business, audience, voice, limits, and editorial goal precisely enough to reject interchangeable copy. "Professional tone" is not useful because almost any draft can claim to meet it.
A working profile answers a short set of questions:
- What does the product do, and what does it not do?
- Who is the reader, and what do they object to?
- Which words does the brand use or avoid?
- Which outcomes must the article not promise?
- What specific next step can it offer?
The seodraft profile, for example, forbids promises about rankings, traffic, or model citations. It requires SEO and GEO to be described as probabilistic processes. That constraint removes a predictable class of exaggerated copy before drafting starts.
The current profile requires at least two internal links and targets 900 to 1,500 words when the topic justifies the length. Those checks are measurable, but they remain subordinate to the answer. A short question should not be stretched to satisfy a word target.
Collect evidence before asking for the body
First-hand evidence must come before drafting because it forces the article to say something a competitor cannot copy. Prompts improve when they contain material, not when they accumulate style instructions.
Useful evidence includes:
- a measured number and its source;
- a real error and the condition that produced it;
- a decision and the tradeoff behind it;
- a named case that can be shared;
- an external source read for a specific claim.
This post uses two real runs. One showed a READY article with no saved brief. Another gate run rejected a non-post internal link, fewer internal links than required, and written-out evidence the matcher could not trace.
The last failure was especially useful. The rule had the right goal, but its matching was narrower than natural writing. That is a product finding, not a generic lesson invented to fill a section.
Save the brief before writing the article
A brief turns intent, angle, evidence, and structure into decisions that can be reviewed before the full draft exists. Without it, the agent makes every choice while writing, and a weak premise becomes hard to separate from weak prose.
A minimum brief should store:
- Search intent.
- The angle that does not repeat the search results.
- At least three first-hand evidence items.
- The H2 outline and each section's opening answer.
- Any image or table that genuinely helps.
The current workflow saves the brief before the body and requires every evidence item in the brief to reach the article. That sequence prevents a useful fact from appearing in planning and disappearing during drafting.
The Spanish keyword brief was reused from cache, so regenerating this post added USD 0.00 in DataForSEO spend. Reusing recent research is part of the system. Paying for the same search results twice does not improve the article.
Use real internal links instead of invented slugs
Internal links should come from an inventory of existing posts so every destination is valid and useful. Guessing a URL from a plausible title creates a broken link or points to content that does not exist yet.
Choose the target by editorial need. A reader who wants the technical stack can continue to the tools founders use for a product blog. A reader who wants a first-hand account can read the founder's personal agent inside seodraft.
Quantity still needs judgment. Two relevant links can extend the reader's question. Five links added only to pass a check make the article harder to follow.
Before saving the body, verify each URL and make sure its anchor text describes the destination. This is a cheap check that prevents the CMS from publishing editorial debt.
Make a gate reject observable mistakes
A gate is useful when it blocks concrete failures in structure and evidence without pretending to replace editorial judgment. It can verify that the brief exists, count internal links, and trace declared evidence into the body.
A real gate run rejected a non-post internal link, fewer internal links than required, and written-out evidence the matcher could not trace. Two problems belonged to the draft. The third produced product feedback about a rule that was too literal.
That is the difference between a gate and a decorative score. A blocking error should name the failed condition and allow a verifiable correction. An advisory can flag a long paragraph or missing image without preventing a founder from reviewing the work.
Do not turn the gate into the reader. If an agent repeats a number awkwardly just to satisfy a matcher, the draft gets worse while the score improves. The right fix may be better wording or a better rule.
Review the article as the founder
The final review must decide whether the article represents the business and whether each important claim has support. Passing the gate means no coded errors remain. It does not mean the post deserves publication.
Read the title, introduction, every H2, and the closing first. That pass reveals whether the article has an argument or only a set of sections. Then verify the evidence and links one by one.
The decision is straightforward: publish when the article answers a real question with material that belongs to the business. If a competitor could publish it by changing the product name, it is still generic.
If you want to run this process with the agent you already pay for, try seodraft. The agent writes, the gate checks, and you approve.