Marketing with AI agents: what to hand over and what to keep

AI agents can now do much of the repetitive work in marketing: checking your site every day, drafting articles and posts, watching what people search for and ask about. For a founder or a team of two, that's real time back.

It also creates new ways to get things wrong in public: a confident but false claim, a bland post under your name, a reply that gets you banned from a community. The difference is rarely the tool. It's what you let the agent do without you.

This guide covers that split. It explains what an agent is, which jobs to hand over and which to keep, a weekly rhythm for working with one, the guardrails that stop the obvious failures, and how to tell whether it's working.

What an AI marketing agent is

A chatbot waits for you. You write a prompt, it answers, and you do something with the answer. An agent gets a goal, access to the tools and data it needs, and a schedule. It runs the steps itself and brings you finished work to review.

	A chatbot you prompt	An AI agent
Who starts the work	You, each time	The agent, on a schedule or a trigger
What it knows about you	What you paste in	Your website, documents and connected tools
What it produces	An answer in a chat	A draft, a fix or a report, ready to review
Where the work goes	You copy it somewhere	Into a queue for approval, then to the right place
Main risk	Wasted time	Acting in public under your name

The last row matters most. Because an agent can act, the question is less "can it write?" and more "what may it do without me?"

A useful way to think about it: a fast new hire who has read everything on your website but has never met a customer. It doesn't get bored or skip the dull checks. It doesn't know what you know, and it can't be accountable for what goes out under your name.

Jobs that suit agents

The best jobs for an agent repeat often, follow a clear standard and are cheap to check.

Daily audits. Checking titles, meta descriptions, broken pages, the sitemap, robots.txt and page speed is the same work every day. People stop doing it; an agent doesn't. Our technical SEO audit checklist lists what to check.
First drafts. Articles, social posts and page rewrites built from your own material: product docs, call notes, changelogs, past posts. A good first draft turns a blank page into an editing job.
Monitoring. Watching which searches you nearly win in Google Search Console, what AI assistants say when buyers ask about your category, and where people ask questions you can answer.
Repurposing. Turning one article into a LinkedIn post, an X thread and a newsletter paragraph, each in the right shape.
Reporting. Pulling numbers from Search Console and Google Analytics into a short summary of what moved.

What these have in common: a bad output costs little, because a person reviews it before anything happens.

Jobs that stay human

Some work depends on judgement, relationships or accountability. Keep it.

Job	Why it stays with you
Positioning: who you're for and who you're not	It's a bet on your market. An agent can suggest options; it shouldn't choose
Final approval of anything public	Your name and your brand are on it
Relationships with customers, partners and journalists	Trust is personal, and automated outreach wears it down
Replying in communities as yourself	Platform rules and community norms expect a real person (see below)
Claims: numbers, comparisons, customer results, legal or health statements	Models can state wrong facts with confidence
Deciding what to stop doing	It needs context about goals, money and time that an agent doesn't have

An agent can still help with each of these: research for positioning, a draft you rewrite, a list of claims to check. It shouldn't make the call.

A weekly operating rhythm

The aim is a routine where the agent does the repetitive work every day and you spend one fixed, short slot reviewing it. Here's one that suits a small team. Move the days to fit yours.

Day	Agent does	You do
Monday	Summarises last week's Search Console and analytics numbers; lists searches you nearly win and pages that dropped	Pick two or three priorities for the week
Tuesday	Drafts an article on the top priority from your own material	Edit, add your own experience, check every claim, approve
Wednesday	Drafts social posts from the article; finds community threads where you could help	Rewrite the posts in your words; reply to threads yourself
Thursday	Re-runs the site audit and drafts fixes	Review the fixes, or pass them to whoever ships code
Friday	Checks how AI assistants answer your buyers' questions	Read the results; update your brand voice and writing instructions if drafts keep missing
Every day	Monitors, audits and queues drafts	Clear the review queue in one sitting

Two habits make this work. Review in one block rather than all day. And when a draft is wrong in the same way twice, fix the instructions, not just the draft.

Guardrails

Approval before anything is published

Nothing should go public without a person approving it, including small things such as page titles and image alt text. Google's guidance on generative AI content says to "manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing", and says this applies to titles, meta descriptions, structured data and alt text too.

Writing instructions and a brand voice document

Without guidance, an agent writes like the average of the internet. Keep two short documents current: writing instructions (how to write) and a brand voice document (how you sound). For example:

Brand voice: Acme

Who we write for: finance leads at companies with 20 to 200 staff.
Tone: plain, direct, a little dry. Short sentences.
Words we use: invoice, approve, close the month, finance team.
Words we avoid: seamless, synergy, best-in-class, cutting-edge.
Never: invent customers, numbers or quotes. Never name competitors.
Always: say what the product can't do when it matters.
Example we like: "Approvals take two clicks. Exceptions still need a person."

A factual review

Google explains why this matters: "generative models don't retrieve facts, but predict a likely sequence of words based on their training data", so their output "may contain inaccuracies". Check every number, date, quote, comparison and product claim against a source before you approve. If you can't find a source, cut the claim.

Platform rules

Each platform sets its own rules on automation, and they change. Read the current versions before you connect any tool. This is what they say at the time of writing:

Platform	What the rules say	What it means for agents
Reddit	Reddit's rules ask you to "participate authentically in communities where you have a personal interest, and do not spam", and not to "intentionally mislead others"	An agent can find threads and draft notes; you post from your own account, say you work on the product, and follow each subreddit's rules
Hacker News	The guidelines say "please don't put generated text in HN posts. Write your text yourself" and "please don't automate posting"; for comments, "don't post generated text or AI-edited text"	Use agents for research and preparation; write and post everything yourself
X	X's developer guidelines require the official API, prohibit automated liking and auto-replies to anyone mentioning a keyword, need X's prior approval for AI-generated replies, and require automated accounts to carry a label	Posting your own approved posts through an official integration is the cautious path; keep replies and likes manual
LinkedIn	The User Agreement prohibits using "bots or other unauthorized automated methods" to "create, comment on, like, share, or re-share posts, or otherwise drive inauthentic engagement"	Use only authorised integrations, and keep comments, likes and messages manual

Our guides to Reddit marketing and launching on Hacker News cover each community in more depth.

How to tell whether it's working

Measure outcomes, not output. Thirty drafts a month mean nothing if nobody new finds you.

Question	Where to look	What to watch
Are more people finding us on Google?	Search Console Performance report, with the non-branded query filter	Clicks and impressions for searches that don't include your name, month on month
Did the pages we worked on improve?	Search Console, filtered by page, comparing before and after	Position, CTR (click-through rate) and clicks
Do visitors do anything?	Google Analytics	Visits from search, and key events such as sign-ups
Do AI assistants name us?	Ask ChatGPT, Gemini, Perplexity and Claude your buyers' questions; Search Console's generative AI performance report for Google	Whether you're named, and who is named instead
Is quality holding?	Your review queue	How much you change in each draft, and how many you reject

Google says the non-branded filter isn't available for sites with few impressions; a regex filter that excludes your brand name does the same job. Its generative AI performance report shows impressions in AI Overviews and AI Mode. For other assistants, keep a simple log of the same questions asked each month. Our guides to finding the searches you nearly win and generative engine optimisation go into detail.

Be patient with search. Search Console data normally takes 2 to 3 days to arrive, and Google says recrawling a changed page can take from a few days to a few weeks. Judge content work over months, not days.

Common failure modes

Publishing unedited drafts

The most common mistake is approving drafts without reading them properly, because they look finished. Read every draft as if a sceptical customer will. Check the facts, cut anything generic, and add something only you know: a customer conversation, a mistake you made, a number from your own data.

Volume over quality

Google says "appropriate use of AI or automation is not against our guidelines". But it also says that "using generative AI tools or other similar tools to generate many pages without adding value for users may violate Google's spam policy on scaled content abuse". That policy, added in March 2024, applies "no matter how it's created". Google's people-first content guidance lists "Are you using extensive automation to produce content on many topics?" as a warning sign. Publish fewer, better pieces.

Sounding like everyone else

Without a brand voice document and real examples, drafts drift towards the same safe phrasing everyone else uses. Readers notice. Give the agent examples you like and words you never use, and edit hard in the first weeks.

Automating the human parts

Automated replies, likes and comments are exactly what the platform rules above prohibit or restrict. The agent finds the conversation; you take part in it.

Set and forget

Your positioning changes, your product changes, your customers change. If the documents the agent reads don't, it keeps telling last quarter's story. Review them once a month.

How the BeforeQuery AI Marketer helps

Here's what a day can look like with the BeforeQuery AI Marketer on a paid plan.

In the morning, the daily rundown arrives in Slack, Telegram or WhatsApp, where you can also chat with your AI Marketer. Chat answers questions; it doesn't take actions. The SEO Agent has audited your homepage and written two fixes, each with steps and a copy-ready snippet, and the Coding Agent can ship fixes like these as a GitHub pull request you review and merge. The GEO Agent has asked AI assistants up to 15 buyer questions and recorded whether you were named, which competitors were named and which sources were cited.

In the dashboard, the Writer Agent has an article drafted from your own material. The X Agent and LinkedIn Agent each have a post in your voice; LinkedIn posts go to your personal profile. The Reddit Agent has found a thread where you can genuinely help and drafted a reply, which you post yourself.

You approve, edit or reject each item in the dashboard. Nothing is published without your approval, and you can publish now or pick a time. Drafts follow your company documents and writing instructions, so editing those is how you change future drafts. You can start free.

Frequently asked questions

Will Google penalise content written with AI?

Google says appropriate use of AI or automation isn't against its guidelines. What it acts on is content made mainly to manipulate rankings, however it was produced, and its scaled content abuse policy covers generating many pages without adding value for users, including with generative AI. Publish fewer pieces, check every fact, and add experience or information only you have.

Can an AI agent post on Reddit or Hacker News for me?

It shouldn't. Reddit's rules ask you to participate authentically in communities where you have a personal interest and not to spam. Hacker News asks people not to put generated text in posts or comments and not to automate posting. An agent can find threads and help you prepare, but write and post the words yourself, from your own account.

Can I post AI-drafted content to X and LinkedIn automatically?

Read each platform's current rules first. X's developer guidelines require the official API, ban automated liking and keyword-triggered auto-replies, and need prior approval for AI-generated replies. LinkedIn's User Agreement bans bots and other unauthorised automated methods for creating, commenting on, liking or sharing posts. Approving each post yourself and publishing through an official integration is the cautious approach, with replies, likes and comments kept manual.

Should I tell readers that AI helped write something?

Google suggests asking "Who, How and Why" about your content, including whether the use of automation is clear to visitors through disclosures or in other ways, and adding a disclosure where readers would reasonably expect one. In communities, the norms are stricter: post as yourself, say you work on the product, and follow each community's rules.

What should I never hand over to an agent?

Positioning, final approval of anything public, relationships, replying in communities as yourself, and judgement on claims such as numbers, comparisons and customer results. An agent can research and draft for all of them, but the decision belongs to a person who is accountable for the result.