HubSpot Retired the Blog Research Agent. Keep the Human Check.

HubSpot's own documentation page for the Blog Research Agent now opens with one line: "On September 7th, 2026, the blog research agent is being sunset. After that date, you will not be able to install this agent."

That date has passed. If you were about to install it, you cannot. If you already had it running, HubSpot's page does not say what happens to your installation, so ask your rep instead of assuming.

Which leaves the half that was always hard. Generating a draft was never difficult. Deciding whether it is good enough to carry your business's name is, and that decision does not retire with the tool. So: what the agent did, what it drew on, what a generator like it spends out of your credit pool, and the review step that transfers to whatever you use next.

What the Blog Research Agent actually did

It lived inside the blog tool instead of in a separate AI product. HubSpot's own description of it is one sentence: "With the blog research agent, you can research topics, identify sources, and generate optimized blog posts."

It needed a Content Hub or Smart CRM subscription at Professional or Enterprise. The full path, while it lasted:

1. Content > Blog, then the Suggestions tab on the Blog posts page
2. Open the Select an agent dropdown and choose Blog Research Agent
3. Click Generate content
4. Read what appears under Latest Agent Suggestions
5. Open in editor on a card to edit and publish, or click the X on the card, type why you are dismissing it, and click Submit

One click produced whole draft posts, which then sat in a queue for someone to approve. That is a publishing tool wearing a research tool's name, and it is why the approve step mattered more than the generate step.

What it drew on, which nobody ever published

This is the question an ops or revenue leader should ask first, and the honest answer is uncomfortable: HubSpot's page for the agent never named its sources. Not the search data, the competitor set, or which of your own pages it had read.

An agent whose inputs you cannot list is an agent whose output you cannot check. You can read a draft and judge whether it sounds right. You cannot trace a claim back to where it came from, so every factual sentence in it is unverified until a person verifies it.

Compare that with the agents where HubSpot does document the inputs. Breeze can draw on your knowledge base, your website and blog pages, uploaded files and your CRM records, and HubSpot gives you knowledge vaults to control which of those each agent is allowed to read. That is why deploying a HubSpot AI agent without breaking your data starts with cleaning those four things before you configure anything. When you know the inputs, a bad answer is a data problem you can fix once. When you do not, it is a proofreading job forever.

What a generator like this spends, out of a pool you share

HubSpot's page for the agent named no price. It did not need to, because the money is upstream: Breeze's content generating agents run on HubSpot credits per action, and credits behave in three ways that catch people out.

• They are bought in bulk, at roughly $10 per 1,000, and a little under $9 per 1,000 on an annual commitment.
• They do not roll over. Unused monthly credits expire.
• Every team draws from the same pool. Support, sales and marketing all spend from one bucket.

So a generator producing drafts nobody publishes is not free. It is spending the credits your Customer Agent needs to resolve conversations. We did that arithmetic properly, including where the bill hides, in what HubSpot's AI agents actually cost to run.

The Australian frame that content marketing advice leaves out

Here is the part that decides whether AI content survives a risk conversation in an Australian business.

On 21 October 2025 the National AI Centre published its Guidance for AI Adoption, which replaces the 2024 Voluntary AI Safety Standard. Six essential practices:

1. Decide who is accountable
2. Understand impacts and plan accordingly
3. Measure and manage risks
4. Share essential information
5. Test and monitor
6. Maintain human control

There is a deeper version written for governance and technical staff. Three of the six bear directly on publishing AI-assisted content, and they are the three content teams skip: decide who is accountable, test and monitor, maintain human control.

This is not early-adopter territory. The ABS surveyed nearly 7,000 businesses and found that 12 per cent of all Australian businesses used AI in 2024-25. Among medium-sized businesses it was 22 per cent, up from 3 per cent three years earlier. We set those numbers against the vendor surveys claiming 68 per cent in what the AI adoption numbers actually say.

The review step, in five checks

Five checks, in this order, before anything AI-assisted goes live.

1. Name a person, per post. The first essential practice is accountability, and most businesses satisfy it at the programme level: marketing owns AI content. That is a department, and a department does not check a number. One name against each published piece, recorded, so that six months later you can ask a specific person where a number came from.

2. Check every number, date and legal claim against a source you can link. Fluent drafts fail hardest on exactly these three. If you cannot put a link next to a statistic, the sentence goes without the statistic. That rule alone would have caught most of what has needed correcting on our own site.

3. Check the Australian specifics. A model trained mostly on American text will write "utilize", quote US privacy law and use a financial year ending 31 December. Our list, every time: Australian spelling, Australian dollars, a financial year ending 30 June, and the Privacy Act and the Spam Act 2003 in place of GDPR and CAN-SPAM.

4. Check you have not published it already. A generator working from your own site will happily produce a fourth post on a topic you have covered three times. Search your own blog for the primary keyword first, and improve the existing post instead of adding a competitor to it.

5. Keep the record. Test and monitor is a practice, and practices leave records. Log which posts were AI-assisted, who approved each one, and what happened to them after 90 days. Without that log you cannot answer whether the agent was worth its credits, which is the only question your CFO will actually ask.

Why one AI-assisted post ranks and the next one does nothing

The usual line is that in a saturated market, publishing more content will not get you more traffic. True, and useless, because it never says what does. Here is the mechanism.

A post earns a position when it answers a question with something the other results on that page do not have. In practice only about four things qualify: a number you measured yourself, a method a reader can follow, a mistake you made, or a local specific. Generic drafts fail because they restate the consensus, and the consensus is already ranked.

Which is precisely what a generative agent is built to produce. Ask it for a post on a topic and you get the fluent average of everything written about that topic, which is what positions four through thirty already are.

The machine drafts the parts that are the same for everybody. A person adds the parts that are only true for you. Skip that second part and you have paid credits for a page nobody will ever reach.

The research work that did not retire

Mapping a topic to what the searcher is actually doing still matters, and no tool does it for you:

• Learning: "what is", "how to", "examples of"
• Comparing: "best tools for", "X vs Y", "templates"
• Buying: pricing, implementation, agency partner, reviews

Then prioritise on four things: relevance to your ideal customer profile, what the searcher wants and which format answers it, the internal expertise and proof you can put in the piece, and whether you can realistically rank for it at all. HubSpot's buyer intent tools cover the first of those from your own traffic.

Competitive gaps are the part genuinely worth automating, and you can build that in HubSpot without an agent: how to build a competitor analytics kit in HubSpot. Answer engines read the same page differently again, which we covered in AEO vs GEO vs SEO.

What we would do this week

1. Check whether anything in your portal is still generating suggestions you are not reading. Content > Blog > Suggestions. An unread queue is credits leaving the building.
2. Write down the inputs any replacement is allowed to read, before you switch it on. If the vendor cannot tell you, that is your answer about how much checking the output needs.
3. Name the accountable person per post and put it in the same place you track the post.
4. Put the credit pool on somebody's dashboard. One shared bucket that does not roll over is a budget, and budgets get watched.

If you are choosing between a native agent and something you build yourself, what Agent Hub and Agent Builder actually launched and why it matters is the decision piece. Agent Builder carries approval steps, so the review above is enforced by the tool instead of by somebody's memory.

Where that leaves you

We are a HubSpot Diamond Partner and we build on Claude, which mostly means we have published AI-assisted drafts we later had to correct, on our own site, before doing it on anyone else's. Our AI Engineering work is largely this one decision: what the machine does on its own, and what a person signs.

A question to leave you with. If a draft published itself on your site tonight with a wrong number in it, how long before anyone noticed, and whose name would be on it?

Sources

• HubSpot, set up and use the blog research agent
• National AI Centre, Guidance for AI adoption: foundations
• National AI Centre, Guidance for AI adoption: implementation guidance
• ABS, business adoption of artificial intelligence accelerates in 2024-25