At Ahrefs, we publish many data-driven posts.
Publishing them is enjoyable. Straightforward. They usually get a ton of search visitors too.
However such posts, like “High Google Searches” or “Most Requested Questions on Google”, are solely price studying if the numbers are present.
Google is aware of that too, which explains the spike in search visitors each time we replace the posts.
So somebody has to maintain them recent.
Ideally, each month, somebody (often the writer) needed to pull recent information from Ahrefs (or the API), filter out the junk and format the tables. The posts with {custom} charts had been worse: spec the brand new chart, hand it to design, wait, assessment it, ship it again for tweaks. Solely then paste all the things into WordPress with out breaking the format, replace the dates, and republish.
One submit is okay. Bearable tedium. However 14 posts? 20? The extra you publish, the extra it turns into a slog. I may lose an entire afternoon and don’t have anything to indicate for it besides a submit that mentioned the identical factor as final month with barely completely different numbers.
It’s probably the most tedious jobs on the content material workforce.
So, we made a compromise. We refreshed them each quarter. (And to be trustworthy, there are some posts we by no means even received to.)
Quick ahead to at the moment. We don’t try this anymore. Letaido does it for us. It’s been working quietly for 2 months now. Altogether, it’s saving us at the very least 20 hours monthly. Not solely can we now replace them each month, we are able to publish extra of such posts, and replace them recurrently too.
It’s a real win/win: far much less drudgery for us, and brisker, extra correct numbers for the reader.
Drop me a few of that fireside emoji, sure please.
Automating content material advertising and marketing like that is apparently retro to confess in 2026, with Gartner saying greater than 40% of agentic AI initiatives shall be scrapped by the tip of 2027.
With the quantity of LinkedIn bragging and far of “AI agent” demos being merely performative, I can perceive the disillusionment. Happily, this one works.
Nevertheless it works exactly as a result of it’s boring. It doesn’t write our articles. It merely does the tedious half, which is an enormous a part of content material advertising and marketing.
I name it the Information Refresh Hub. It’s a instrument that lives in our Letaido workspace.
As soon as a month it pulls recent information for all 14 datasets — key phrase volumes and questions from Key phrases Explorer, AI citations from Model Radar — cleans every one by its personal guidelines, and saves the outcomes so I can see precisely what it stored and what it threw out. Then it builds a WordPress draft with the brand new tables in place and emails me to say it’s prepared.
I need to be trustworthy about how unglamorous the constructing was.
Getting the info alone meant three fully separate paths. I may get the US key phrase tables simply through Letaido because it has all Ahrefs information. However the international ones weren’t out there because it was custom-made by our information scientists beforehand for these posts. So I needed to join it to a separate inner service. Then I needed to seize the AI quotation tables from Model Radar, one platform at a time.
After which there are what appear to be foolish issues. One construct stored throwing a 500 error over a tiny capitalization mismatch: our code despatched the sphere as Cpc, and the API insisted on CPC, all caps. I misplaced a genuinely embarrassing period of time to that one.
Regardless of all of those, I need to say it was genuinely magic. In spite of everything, I didn’t hand-code any of this. I constructed it conversationally in Letaido. Letaido did all of the work. Even the “time misplaced” was Letaido determining the right way to repair it, not me.
There are two jobs I stored intentionally human.
The primary was judging what the agent produces.
Take “most requested questions on Google”. You’d assume pulling the highest questions is simply sorting by search quantity. It isn’t. The uncooked checklist is filled with issues that appear to be questions however aren’t. “Tips on how to practice your dragon” is a film. “Would you somewhat questions” isn’t a query in any respect. Model and product searches sneak in. So do oddly particular queries that learn like a bot wrote them.
An individual spots these in a second. So we run a cleansing layer, together with an LLM move, whose complete job is to make these calls at scale. For the “most searched folks” desk, it really works via as much as 5,000 candidates and decides what’s an actual human title, what’s “[name] web price”, and what’s only a regular phrase that occurs to appear to be a title.
It’s good at this, however not excellent, which is strictly why I take a look at each refresh earlier than it goes wherever.
My colleague Louise ran right into a more durable model of the identical downside. She constructed an agent that ranks the fastest-growing firms utilizing Ahrefs information, and the deceptively arduous half was instructing it what counts as an actual breakout model and what’s simply noise.
Some firm names are additionally atypical phrases. You may’t measure the expansion of “cursor” or “perplexity” from zero, as a result of folks had been looking these lengthy earlier than the businesses existed. So the system estimates what number of searches the phrase was already getting earlier than the model emerged, subtracts that baseline, and counts solely the brand-driven quantity on prime. The corporate stays on the checklist; solely the pre-existing noise comes off.
Then it has to disregard one-month spikes that by no means maintain, and really Google every title to substantiate the corporate itself ranks for it. In any other case “Tropic” the software program vanishes beneath Tropic the skincare model. Each a kind of guidelines is a name Louise made about what “actual” means. The agent simply enforces it.
All human by the approach.
That is additionally why the agent by no means publishes by itself. It creates a draft, and solely goes reside after a human confirms it.
None of this sounds notably spectacular. However I feel that’s the precise magnificence of automation. I’m fully fantastic with an agent that does 90% of a job and leaves me the final 10%. An agent that does 100% and sometimes publishes nonsense to a reside, public weblog received’t be a time-saver. It’s asking for a hearth to place out.
That’s why I nonetheless verify it’s proper and hit publish myself.
I initially constructed the Information Refresh Hub for my very own posts. I didn’t assume it was something particular, however I made a decision to share about it on Slack.
Seems I really underestimated what I did. It impressed my colleagues to start out doing related issues.
Louise constructed an entire household of fastest-growing firm rankings. She didn’t simply replace the info; she additionally used Letaido so as to add judgment, charts, and all kinds of different information.
Our Director of Content material Advertising, Ryan, additionally arrange the identical form of month-to-month automation for his personal information content material. His reaction, close to sufficient phrase for phrase: “This was my dream for AI: precise automation, genuinely saving us hours of drudgery. And it’s lastly right here. SORCERY!!!”
His model now runs on a schedule: pulls recent information, regenerates the charts and tables, builds the WordPress drafts, makes the small date and sample-size edits, and emails him when the article’s able to look at.
No person was informed to do any of this. It unfold as a result of it labored, and the truth that it unfold by itself (with out anybody assigned to make it occur), is a transparent signal that it’s actual and never only a demo. Helpful issues simply get copied, with out anybody needing to name a gathering.
There are three of us working a model of this now.
I can nearly assure that you’ve got a job like this hiding in your individual work. Most content material groups do.
Right here’s how I’d go on the lookout for it.
Begin with a query. Undergo the work you do on repeat and ask two issues of every activity: does it run on a schedule, and will you write down the principles for what “completed proper” appears to be like like?
If each solutions are sure, it’s a candidate. “Pull the identical numbers from the identical place each month and reformat them the identical approach” passes simply. “Write the article” fails the second check, and that’s the half you may need to hold doing your self anyway.
If what you’re working is advertising and marketing work, simply go to Letaido and inform it what you want. It’ll do many of the arduous, tedious give you the results you want. (For those who’re an Ahrefs buyer, you get a free month.)
Then, if I needed to boil down what really made ours work:
- Automate the plumbing, not the considering. Fetching, cleansing, formatting, pasting. These are all mechanical work and it’s precisely what you need to hand off. Preserve the considering half for your self.
- Make the cleansing seen. Don’t let the agent simply hand you a completed checklist. Get it to indicate you what it eliminated, and why, proper subsequent to what it stored.
- Preserve a human on the gate. Drafts solely. Let an individual publish. This buys you many of the security.
- Lock the issues the mannequin shouldn’t contact. Headline stats, verified figures, the opening line. You’d need to pin them down so the agent can’t quietly reword a quantity into one thing that isn’t true anymore.
That’s actually all it’s. It isn’t thrilling, and type of the purpose. The boring, well-defined jobs are those AI handles effectively at the moment, they usually’re sitting in plain sight in just about each content material workflow.
This is likely one of the finest components of AI automation proper now. It may possibly assist with all of the work you quietly dread each single week or month.
Get an agent to do it, however be the editor that claims it really works and pushes reside.
If there’s a lesson in right here, it isn’t a really flashy one. Hand the boring, repetitive stuff to the machine, and hold the components that truly want you.
We’re all managers now.
