Mark
Back to Blog

Three Inputs AI Can't Supply: Your Content, Your Taste, Your Review

Every brand can use the same AI. What sets content apart is what you feed it, the standard you hold it to, and the person who signs off. A practical guide.

Three lines on a dark blue background: Your content. Your taste. Your review.

A language model can write a plausible post on almost any topic. What it cannot do is know what happened on your last customer call, which phrases you would never use, or whether a claim is safe to print. Those come from you. When two competitors use the same model and supply none of them, they tend to publish similar posts.

Three inputs decide whether AI-assisted content is distinct: your content (the first-hand material only you hold), your taste (the standard you apply to every draft), and your review (a named person who decides what ships). The research below explains why the model alone is not enough, and the later sections cover how to set them up whether or not you use Mark.

Why does AI content from different brands read the same?

AI content from different brands reads the same because the models behind it behave alike. In a 2026 PNAS Nexus study, Emily Wenger and Yoed Kenett found that responses from large language models "mirror other LLM responses far more than humans do other humans," even after controlling for response structure. They also found it holds across different models, not just one.

A 2026 meta-analysis by Alwin de Rooij and Michael Mose Biskjaer pooled 19 studies and 61 effect sizes. The authors found a small but statistically significant homogenization effect when people create with generative AI (d = 0.334). The authors note that individual performance can improve while diversity across users shrinks.

Both studies used creativity tasks, not marketing copy, so they do not measure your blog. The findings do match a pattern most owners have already noticed: ask the same tool for the same kind of post and you get the same kind of post.

What counts as "your content," and why can't the model invent it?

Your content is the material that exists only because your business exists. A model can only rearrange what it has read, so anything specific to you has to come from you.

  • First-hand experience. Notes from customer calls, jobs you finished, what went wrong last quarter. Without it, the draft describes the category, not your work.
  • Original data. Your audits, survey results, prices and lead times. Without it, the draft quotes the same public statistics as everyone else.
  • Real language. How customers describe the problem, in their words. Without it, the draft uses generic marketing phrases.
  • Sourced claims. Certifications, specifications and guarantees, each tied to a document. Without them, the draft invents a claim or leaves it out.

Google's guidance asks the same questions. Its self-assessment asks whether the content provides "original information, reporting, research, or analysis" and whether it demonstrates "first-hand expertise and a depth of knowledge."

The same material also helps with AI answers. In the GEO study by Pranjal Aggarwal and colleagues, adding citations, quotations and statistics raised a source's visibility in generated answers by 30 to 40 percent on the paper's position-adjusted word count metric. That figure measures share of one generated answer in a benchmark, not clicks or revenue. The practical point is narrower: those three additions need real sources, real quotes and real numbers. A model asked to add them can make some up unless you supply them.

What is taste, and how do you write it down?

Taste is the set of decisions about what your brand says, what it refuses to say, and what good looks like. It is usually unwritten, which is why a model cannot follow it. Writing it down turns a feeling ("that doesn't sound like us") into rules a draft can be checked against.

Start with five short lists:

  1. Who you are writing to, in one sentence, with what they want and what makes them hesitate.
  2. Words and moves you avoid. Hype words, stock openers, promises you never make.
  3. Three posts you would be proud to have written, so the standard is shown, not described.
  4. Things you never claim, such as rankings, guarantees or savings you cannot document.
  5. Your default call to action, so every post ends the way you would end it.

Keep it short enough that a new hire would read it. A thirty-page brand book does not get applied to a draft at 4 p.m. on a Thursday. Five lists that people actually check do.

Google's own questions are a reasonable taste test, too. One of them asks whether this is the sort of page you would want to bookmark, share with a friend or recommend.

Why does review have to be a person, and what do they check?

Review is where your content and taste meet the draft, and only someone accountable for the result can do it. Google's guidance says it is "critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing." Its people-first content page, describing what quality raters assess, says "using generative AI to produce large amounts of text without manual oversight or curation represents little to no effort."

A useful review is four questions, asked in this order:

  1. Is there anything in this that only we could have written? If not, send it back for more of your material.
  2. Does every claim have a document behind it? If not, cut it or find the source.
  3. Does it break any of our five lists? If so, fix the draft, and ask whether the list needs a new line.
  4. Would the named owner put their name on it? If not, it does not ship.

The last part of question 3 matters most over time. A correction that stays in the reviewer's head is spent once. A correction added to your taste notes improves every later draft.

Do the three inputs help in AI answers too?

The three inputs plausibly help in AI answers too, though nobody has measured the effect for small-business content. Answer engines assemble responses from sources, and Google's quality questions ask whether a page provides "substantial value when compared to other pages in search results." If your page says what ten others say, there is little reason to cite yours. If it holds a number, a quote or a claim only you can source, there is a reason. That is an inference from the guidance and the GEO study above, not a tested result.

A 30-minute starter kit

You do not need new software to begin.

  1. Collect (10 minutes). Put three things in one folder: your last five customer emails or call notes, your current price and terms, and any audit or report you own.
  2. Write the five taste lists (10 minutes). Use the prompts above and keep each to a few lines.
  3. Name the reviewer (5 minutes). One person, with authority to say no, and the four questions where they will see them.
  4. Close the loop (5 minutes). Set a recurring reminder to move repeated corrections into the taste lists.

Then give a model your folder and your lists with the first draft request, and review what comes back.

How does Mark build these inputs into the workflow?

Mark was designed around the same three inputs, and its docs describe each one.

  • Content. Under Strategy, the Brand page has an Entities and evidence group where you list approved experts and first-party evidence such as an audit or dataset. If you list none, drafts say the evidence is missing instead of making some up. Claims & sources holds each fact a post may state, with the quote from the source that proves it.
  • Taste. The Brand page also holds positioning, differentiators, tone, do and don't lists, example posts and call-to-action defaults. Personas, pillars and angles sit under Messaging, and you can edit all of them. Through Mark's MCP tools you can also list phrases Mark should block or warn on.
  • Review. By default a person approves every post, and nothing publishes until then. A draft that cites an unconfirmed claim is flagged, and the flag blocks approval until someone confirms the source. A person confirms each Brand group by saving it, and an agent's change stays "Not confirmed" until they do.

One caveat, because it matters: a channel can be set to an Auto approval policy, which approves posts that have no claim flags without a reviewer. Required is the safer setting if review is the point. The details are in the Mark docs, and you can see how it works at use-mark.com.

What this approach does not fix

  • Real material is required. A business with no customer notes, data or opinions has nothing to supply yet. The first job is to collect them, usually by talking to customers.
  • Review adds time. Review costs minutes per post. Whether that pays back depends on how costly a generic or wrong post is for you.
  • Taste can be wrong. Written rules can harden into habits. Revisit them when results or customers tell you something new.
  • We have not measured the ranking effect. The studies above show sameness and show what Google and the GEO paper reward. None of them shows that supplying these three inputs will lift your traffic.

Sources

  1. Wenger, E. and Kenett, Y. N., "Large language models are homogeneously creative," PNAS Nexus 5(3), pgag042 (24 March 2026)
  2. de Rooij, A. and Biskjaer, M. M., "Generative AI Makes Creative Output More Homogeneous," ECCE 2026 (ACM)
  3. Aggarwal, P. et al., "GEO: Generative Engine Optimization," KDD 2024 (arXiv 2311.09735)
  4. Google Search Central, "Creating helpful, reliable, people-first content"
  5. Google Search Central, "Google Search's guidance on using generative AI content on your website"
  6. Mark docs, "Set up your brand, personas, pillars, markets, and claims" and "Review and approval"