Run your marketing
on autopilot.

Most sites are sitting on traffic they have already earned. Positions 5 to 20 in Search Console are the cheapest wins available, because Google has already decided the page is relevant. It just has not decided it is the best answer yet.

That distinction matters more than it sounds. A page at position 14 does not need a new topic, a new angle, or a rebuild. It needs the specific reason it is losing to the four pages above it removed. That reason is usually one of four things, and all four are cheap to fix.

Start where Google already agrees with you

Export the last 90 days of queries, filter to average position 5 to 20, then sort by impressions. What you are looking for is a page with demand and no click-through: proof of interest, and proof that something on the page is not landing.

Ignore difficulty scores at this stage. Difficulty is a prediction about pages you have not written. Position is a measurement of a page you already have. You do not need the prediction when you have the measurement.

Five rows, four different problems. This is why bulk rewriting underperforms: the pages are not failing for the same reason, so the same fix cannot work on all of them.

The four reasons a page stalls at the bottom of page one

• The title does not match the intent. The query wants a list and the title offers a philosophy. This is the cheapest fix and the only one that moves click-through without moving position.
• The answer is buried. Someone searching a question wants the answer above the fold. If the first 80 words introduce your company instead, the page loses to one that does not.
• A format is missing. If four of the five pages above you have a comparison table and you do not, you are asking the reader to do work the others already did.
• Nothing points at it. A page with no internal links is a page the rest of your site does not vouch for.

Practical rule: fix the title first, the first 80 words second, the missing format third, and internal links last. That is cheapest to most expensive, and it is also fastest to slowest to show up in the data.

What to change, in order

1. Rewrite the title tag to match the query, not the page. If the query is plural and comparative, the title should be plural and comparative. Give this two weeks before judging it.
2. Move the answer up. Put the direct answer in the first paragraph, then earn the rest of the read. Search behaviour rewards this and so do readers.
3. Add the one section the competing pages all have. Open the top five results and note what they share that you lack. Usually it is a table, a set of screenshots, or a pricing breakdown.
4. Link from your three strongest pages. Slowest of the four to take effect, so do it last, but do it deliberately rather than adding links at random.

Search Console, filtered to positions 5 to 20 and sorted by impressions. Everything above the fold here is a page that already ranks and already has demand.

How long it takes, realistically

Title changes show up within one to three weeks. Content changes take longer because the page has to be recrawled and reassessed. Internal linking is the slowest and the most compounding: it does very little in month one and quite a lot in month four.

The mistake teams make is judging all four on the same timeline, concluding nothing worked, and rewriting the page from scratch. That resets whatever authority the URL had accumulated.

Written by BrightBean from the brand brief, then fact-checked against the Search Console export it cites. Published to northstarfield.co/blog with the internal links already placed.

Repurposing fails when it is a copy-paste job. One recording becomes eight posts that all say the same thing in slightly different words, the audience notices by the third, and the team quietly stops doing it.

It works when the cutting happens once, at the source, and each channel gets the format it actually rewards. That is a different job from reformatting, and it starts before anyone opens an editor.

Find the six ideas, not the thirty posts

A one-hour recording is not thirty posts. It is roughly six ideas, each of which has a natural shape on three or four channels. Plan for thirty and you will pad twenty-four of them. Plan for six and the thirty become a formatting job.

Read the transcript looking for three specific moments:

• A contradiction. Any point where the speaker says the opposite of the received wisdom in your category. These become the openers, because they earn the second line.
• A number. Anything measured, counted, or compared. Numbers survive being pulled out of context, which is exactly what you are about to do to them.
• A story. A specific customer, a specific week, a specific mistake. Stories carry video; abstractions do not.

Everything else in the transcript is connective tissue. It exists to get the speaker from one idea to the next and it will not survive being extracted.

One recording is not one post you can slice thirty ways. It is six ideas, each with a natural shape on three or four channels.

Then let each channel take the shape it wants

This is where most repurposing goes wrong. The same 200 words get posted everywhere, and every channel punishes it for a different reason. LinkedIn wants the argument, X wants the claim, TikTok wants the moment, Instagram wants the framework.

Thirty-one pieces from one hour. None of them is the same piece twice, because each one was cut for what its channel rewards rather than resized for what fits.

The transcript pass. Six ideas marked, everything between them left alone.

Schedule the follow-up before you publish the first one

A repurposing run that lands over three weeks outperforms the same posts fired off in three days, and the reason is boring: the second week is when people who missed the original start searching for it.

A cadence that holds up:

1. Week one, the contradictions. The strongest openers go out while the recording is fresh and the people who attended are still talking about it.
2. Week two, the numbers. These travel furthest because they get quoted. Give them their own week so they are not competing with each other.
3. Week three, the stories and the article. The long-form piece lands last and links back to everything that came before it, which is what makes the whole run compound instead of evaporate.

What to measure, and what to ignore

Reach on individual posts is close to useless here. Two things actually tell you whether the run worked:

• Reply rate on the contradictions. If the openers are not producing argument, they were not contradictions, they were observations.
• Search traffic to the article in weeks three to six. This is the part that pays for the whole exercise, and it is invisible if you stop measuring after the social posts stop.

Drafted by BrightBean from the webinar transcript, split across channels according to what each one has historically rewarded on this account. Published to northstarfield.co/blog via Ghost.

An agent with no memory rewrites your positioning every session. It is confident, it is fluent, and it is starting from nothing every single time, which means the third draft contradicts the first and nobody notices until a customer does.

The fix is not a longer prompt. Prompts are read once and thrown away. What you want is a durable store the agent reads before it writes and writes back to only with permission.

Give the agent something durable to read

The tempting design is to paste the brand guide into the system prompt. It works until the guide is longer than the context you can afford, and it fails silently before that: the model sees the text but has no way to tell you which part it used.

Exposing the same knowledge as a tool changes both properties. The agent pays for the tokens it actually needs, and you get a log of what it looked at before it drafted.

$ brightbean mcp status
  brand    connected   voice, audience, 218 learnings
  social   connected   11 channels, 3 pending review
  blog     connected   northstarfield.co

> brand.search({ q: "how we talk about pricing" })
  3 matches
  voice.pricing          confidence 0.91
  objections.tiers       confidence 0.84
  proof.pricing-page-ab  confidence 0.77

Three matches, each scored, each traceable. When the draft comes back wrong you can ask which of the three it leaned on, which is a question you cannot ask of a system prompt.

Write back behind a gate

The obvious next step is to let the agent update your brand context whenever it learns something. In practice that is how one bad session poisons your voice for a month, because the agent is equally confident about the thing it got right and the thing it inferred from a single outlier.

Queue the write instead. Surface it as a proposed learning with the evidence attached, and let a human accept or revert it.

The expensive failure is not a bad draft. It is an agent that silently learns the wrong thing and applies it to everything it writes next.

What breaks in production

Three failure modes account for most of the trouble, and none of them are model quality.

• Tool sprawl. Every additional tool costs the agent tokens to consider and gives it another way to pick wrong. Fewer, broader tools beat many narrow ones.
• Unbounded search results. An agent that gets forty matches back will use the first three and pretend it read all forty. Cap the result set and score it.
• No idempotency on writes. Agents retry. If scheduling the same post twice creates two posts, you will find out on a Monday morning.

A learning in the review queue: what changed, what evidence produced it, and one click to revert.

A sane default permission set

1. Start read-only. Let the agent draft for a week with no write access at all. You will learn more from what it gets wrong than from any evaluation harness.
2. Open the learning queue next. Proposals only. Watch what it wants to write back before you let it.
3. Grant scheduling last, and scope it. One channel, a review step, and a cap on posts per day.

Drafted by BrightBean against approved MCP server documentation, with the terminal output taken from a real status call. Published to northstarfield.co/blog from a Git repository.

A scheduler moves content you already wrote. The category worth paying for now is the one that writes it, learns from what happened, and changes the next draft accordingly. They look similar in a feature list and behave nothing alike on a Monday morning.

The useful question is not which has more integrations. It is whether your bottleneck is distribution or creation, because the answer decides which of the two is a waste of money for you.

What the two categories actually do

That last row is the honest one. Both categories fail, they just fail differently. A scheduler fails quietly when the person who writes goes on holiday. A system fails when nobody reads what it produced for six weeks and the voice slides somewhere you would not have chosen.

How to tell which one you need

Count the hours rather than the features. Ask the person who currently publishes where the week goes.

• If the writing is done and the pain is the posting, a scheduler is the honest answer and it is cheaper. Buy the boring one.
• If the calendar is empty because nobody had time to write, a scheduler will not fill it. It will show you the gap on a nicer grid.
• If you have writers but they spend their week reformatting, you have a repurposing problem, which is a third thing and cheaper to solve than either.

The question is never which tool has more integrations. It is whether your bottleneck is distribution or creation.

The questions worth asking on a demo

Feature lists are written to survive comparison. These five questions are harder to answer with a slide.

1. Show me a draft this produced six months into an account. Anyone can demo the first draft. The interesting question is whether draft two hundred still sounds like the brand.
2. What happens when a post does unusually well? If the answer is a notification, the loop is not closed. Something should change in what gets written next.
3. Who can publish without asking? If drafting and publishing share one permission, you will find out why that is a problem eventually.
4. Where does the blog fit? Tools that treat long-form as an afterthought usually mean two subscriptions and two calendars.
5. What does it cost when we double the team? Per-seat pricing on a tool used by five people is a different product from the same tool used by twenty.

The same week in both categories. One is a grid you fill; the other arrives filled and waits for approval.

The case for buying neither

If you publish twice a week to one channel and it is working, both categories are overhead. A shared document and a recurring calendar hold will beat either until the volume genuinely hurts. Tooling should follow a bottleneck, not create one.

Written by BrightBean from a category brief, with the comparison table generated from its own feature notes and checked line by line before publishing. Published to northstarfield.co/blog via Webflow.