How to keep AI content on brand

1. Step 1

   Write the voice down as rules, not adjectives

   "Friendly and professional" describes almost every company, so it does not steer a writer, human or AI. Write rules a draft can be checked against: who the writing addresses, whether it says we or you, how long sentences run, how formal it is, how much jargon is fine, and whether it uses humour, contractions or exclamation marks. Give each rule one example that follows it and one that breaks it.

2. Step 2

   Write down the positioning and the proof you can stand behind

   Say in a sentence or two who you serve, what you do for them and why you are different. Then list the proof points a draft may lean on: the results, credentials and facts about how you work that you can show evidence for. A writer with no proof points will reach for generic claims or invent specific ones.

3. Step 3

   List the words and claims to avoid, and the words to prefer

   Name the terms you never use, such as buzzwords and words that read as hype, and the terms you do use for your product, your audience and your category. Add the claims you must never make: guarantees, superlatives, comparisons with named rivals and anything legal or compliance has not cleared. A written list can be checked mechanically; a feeling cannot.

4. Step 4

   Give the writer the brand context before it drafts

   Put the voice rules, positioning, proof points and avoid-lists in front of the writer at the start of every job, as part of the brief, instead of correcting the tone afterwards. Pasting a style guide into one chat helps once. The document has to travel with every draft, and it has to be the same version each time.

5. Step 5

   Check brand fit on every draft, against the same document

   Read each draft against the written rules rather than a general impression. Does it sound like the voice rules, match the positioning, lean only on listed proof points, and avoid the listed words and claims? Check every draft, not a sample: off-brand text reads fine on its own and only shows up against the rules.

6. Step 6

   When a draft drifts, fix the cause as well as the sentence

   Name what drifted: the tone, a banned word, a claim without proof or the wrong positioning. If your rules did not cover it, add the rule and an example. Then send the draft back with that specific instruction and the passage it applies to, and check the whole draft again, because fixing one paragraph can break another.

7. Step 7

   Decide who reviews, and what they can overrule

   Name one person who knows the brand and has the authority to stop a piece, and where your team is large enough make it someone other than the person who wrote the prompt. Decide in advance which failures they can waive and which always block publishing. Review the brand document itself on a schedule too: an out-of-date rule produces confidently wrong drafts.

This page is about brand. The mechanics of review, meaning which version a person approves and what happens when it is edited, are on the AI content approval workflow, and checking claims against sources is a separate job: how to fact-check AI content.

• Voice: person, formality, sentence length, jargon level and humour, each with an example that follows the rule and one that breaks it.
• Positioning: who you serve, what you do for them and why you are different, in one or two sentences.
• Audience: the readers you write for and what they already know.
• Proof points: the results, credentials and facts you can back with evidence, and nothing you cannot.
• Words: the terms to prefer and the terms to avoid.
• Claims never to make: guarantees, superlatives, comparisons with named rivals and anything legal has not cleared.
• Calls to action: the next steps you want readers to take, in your wording.
• Examples: a few real passages that sound exactly like you, and a few that do not.

1. Move 1

   Name what drifted

   Tone, a banned word, an unbacked claim and the wrong positioning each need a different fix. "Off brand" on its own is not an instruction.

2. Move 2

   Fix the rule if it was missing or vague

   If the document did not say it, the writer could not have followed it. Add the rule and an example so the next draft starts in the right place.

3. Move 3

   Send back one specific instruction

   Quote the passage, name the rule it breaks and say what to do instead. Ask for the whole draft to be checked again, not only that paragraph.

4. Move 4

   Check the whole revised draft

   A revision can fix one paragraph and break the voice of the next, or add a claim you cannot prove. Run the same checks on the revised version as on the first.

5. Move 5

   Stop looping and write it yourself

   If repeated revisions do not fix it, the brief or the rules are probably the problem. Have a person write that section.

• Write the voice down as rules, not adjectives. A site can have one brand profile, written on the site’s Brand profile tab, and its voice is stored as a voice descriptor: a JSON object such as tone, formality and person. The descriptor reaches the writer as the stored JSON text, keys and punctuation included, and is cut off after a fixed length, so keep it short and put the important rules in short values. Where ingestion is enabled for the workspace, the app’s re-ingest can draft the profile from the site’s homepage URL, and over MCP brand_ingest (brand:write scope) takes the URLs you list, at least one. A fixed cap limits how many are fetched and the extraction reads fewer than that, and pages with almost no text are skipped, so list only the pages that show the brand best, with no sitemap. brand_get (brand:read scope) reads the profile back, without its style examples.
• Write down the positioning and the proof you can stand behind. The profile also stores the positioning and the proof points. The brief is written from the research plus the positioning and proof points, and its instructions ask for an angle that reflects those proof points.
• List the words and claims to avoid, and the words to prefer. The profile has lists of forbidden terms and forbidden claims. The writer is given both and told to honor them. After drafting, the text is searched for each one, ignoring case, and every match raises the Risk score, which can fail its hard gate.
• Give the writer the brand context before it drafts. The brief gets the positioning and proof points. Every draft and revision gets those plus the forbidden claims, the forbidden terms and the voice. A site with no profile is drafted in a neutral expert voice with fewer first-hand claims, and its Brand fit score then has no positioning or proof points to judge against.
• Check brand fit on every draft, against the same document. Every version uprank scores gets a Brand fit score from a model that is given the brand’s positioning and proof points and the article text. A Brand fit score of at least 80 is one of the hard gates, and a missed gate shows in the Quality tab as a blocker with the score and the minimum. A version that has not cleared every hard gate cannot be approved, pushed to Strapi or recorded as a live publication request.
• When a draft drifts, fix the cause as well as the sentence. When a version fails a gate, the pipeline revises it before it reaches you. It takes the explanation for each failing or weak dimension, Brand fit when it is one of them, writes a new version with those explanations as revision instructions, re-checks the claims and scores the new version, for a bounded number of passes that stop early once every gate passes. A version that still fails goes to review with its blockers listed. Over MCP, article_revise takes free-text instructions and a list of dimensions to focus on, such as Brand fit, and creates a new version that is scored again.
• Decide who reviews, and what they can overrule. Saving the brand profile and approving a version both need the Editor role or higher, so a Viewer can do neither, and each profile save is recorded in the audit log. Approval only becomes available once the version has cleared every hard gate, Brand fit among them, so the gate checks first and a person decides second. uprank does not require a second reviewer.

The full list of gates is on quality gates, and the product view is an AI writer that doesn’t hallucinate sources.

• Judge brand fit the way an editor does. The score is one model’s opinion, formed from the positioning, the proof points and the article text, and for a long article it reads only the first part. It is not given the voice descriptor or the forbidden lists. Forbidden terms and claims are matched separately for the Risk score. Nothing checks a draft against your voice rules, so an article can pass Brand fit and still sound wrong, or fail it and sound fine. Read the draft.
• Use every field of the profile. Audiences, products and services, preferred calls to action, preferred terms, style examples and competitor names are not part of the instructions the brief, the draft or the scorers receive today. You edit them in the app; over MCP, brand_get returns all of them except style examples. Put what matters into the voice descriptor, the positioning and the proof points, which are.
• Catch a reworded forbidden claim. Forbidden terms and claims are found by matching their text, wherever it appears, so a paraphrase passes and a term inside a longer word can match. A single matched term on its own may not fail the Risk gate, while a matched claim adds more, so treat the lists as a check that raises risk, not as a ban. Review for meaning.
• Learn from your edits. The profile changes only when someone saves it or re-ingests, and saving it does not score articles that are already written. Score a version again to see the effect of a change.
• Do better than your own pages. Ingestion builds the profile from the pages it is given, so a thin or out-of-date site gives a thin or out-of-date profile. A re-ingest replaces every field, including edits made by hand, so re-ingest first and edit after.
• Replace the reviewer. Brand fit is a matter of judgement, and a person still decides whether a piece sounds like you.

How do you keep AI-written content on brand?

Write the brand down as rules rather than adjectives: voice, positioning, proof points you can back, and the words and claims to avoid. Give that document to the writer before each draft, check every draft against it, and send back drift with a specific instruction. Correcting the tone after the fact does not scale, and a writer given no brand context falls back to a generic voice. In uprank a site can have a brand profile, and its briefs and drafts are written from it.

What should AI brand voice guidelines include?

Voice rules with an example that follows each one and one that breaks it, the positioning in a sentence or two, the audience, proof points you can show evidence for, the words to prefer and avoid, the claims never to make, calls to action in your wording, and a few real passages that sound like you. Prefer rules and examples to adjectives, and keep it short enough that a writer can hold it in mind.

Can AI learn my brand voice from my website?

It can draft a starting point, where ingestion is enabled for the workspace. In the app it reads the site’s homepage URL, and over MCP you list the pages. uprank asks a model to extract the profile from those pages and saves it. The model is told to use only what the pages say, which does not guarantee it, and your pages may not show the voice you want from now on, so read the result and edit it. A re-ingest replaces every field, so edit after the last ingest.

How does uprank check that a draft is on brand?

Each scored version gets a Brand fit score from a model that is given the brand’s positioning and proof points and the article text, and a Brand fit score of at least 80 is one of 13 hard gates. Forbidden terms and claims are checked separately by text match and raise the Risk score, which can fail its own hard gate. When the pipeline writes a draft that fails, it revises it a bounded number of times and then holds it for a person; scoring a version on its own does not revise it. A failing version cannot be approved, pushed to Strapi or recorded as a live publication request until it passes.

• AI content approval workflow: review before you publish: who approves which version, and what sends an edit back through the checks.
• Quality gates: every dimension, weight and hard gate.
• How to fact-check AI content: check each claim against its source before you publish.
• AI writer that doesn’t hallucinate sources
• How to add citations to AI-generated content
• AI blog writer vs AI chatbot for blog posts
• FAQ: product, agents, tokens and publishing.
• Agent setup guides: connect a coding agent over MCP.