We Didn’t Automate Our Content. We Automated the Tedium.
July 3rd, 2026 by
Most “AI writes your blog” systems solve the wrong problem. Writing was never the bottleneck. Research is.
If you know a niche cold, you can draft fast. But most agencies don’t write about one niche. At Search Influence, we write about addiction recovery, senior living, higher ed, legal, plastic surgery, limousines. Nobody on the team is a domain expert in all of those, and pretending otherwise is how you get thin, generic content that reads like everyone else’s. What makes content good, the part that’s slow and genuinely hard, is the research that happens before anyone writes a sentence.
When I sat down to try and automate one of our content editors’ process, I didn’t start by automating the writing. I started by watching how a good editor actually researches, then asked why the research most AI content skips is the research that matters most.
Search Influence is an AI-forward SEO agency, and the project behind this piece was rebuilding our content research around how AI search actually works: query fan-out, information gain, and authoritative sourcing.
Why I Watched a Human Before I Built Anything
There are a lot of well-documented AI content systems out there. I’ve reviewed several. But most of them optimize the draft and treat research as an afterthought, or skip it entirely.
Before I built anything, I watched one of our editors work through a real assignment, start to finish. Every time he searched for something, I asked why. What are you looking for here? What do you do with it once you find it? I wanted his real process in my hands, not the tidy version anyone describes when you ask them to explain how they work.
Two things came out of that.
First, I could coach him. Once I understood his actual steps, I could see where they were strong and where there were opportunities.
Second, I could map his process against the people doing this at the highest level. I leaned on Corey Haines’ marketing skills, Zubair Trabzada’s GEO/SEO skills for Claude, and Nicolas Gorroño (Nico | AI Ranking), whose publication workflows are the clearest I’ve seen. Nico builds a lot of his research on DataForSEO and Make, which is exactly where our own tooling started. Watching our editor and asking “why” let me translate his instincts into the parts of those methods he wasn’t yet using.
The Friction That Killed Our Good Tool
What surprised me: we’d already built the tool for this. I built a Make workflow a while back that pulls Search Console data, Semrush data, and People Also Ask results (via Mark Williams-Cook’s AlsoAsked) and assembles them for a specific client and keyword. It’s a good tool. Our editor told me he liked it.
He also wasn’t using it.
Not because it didn’t work. Because it required pre-planning. To run it, he had to know an assignment was coming far enough ahead to set up the workflow before it landed on his desk. In the real rhythm of his work, that pre-process was enough friction that a genuinely useful tool went unused.
That’s the lesson I keep relearning: the best workflow is the one that doesn’t ask a human to do setup. If using the tool costs more attention than doing it by hand, people do it by hand. The skill we built removes the pre-process entirely. The research is kicked off by the assignment itself.
What We Actually Added
The old way (his manual process, and even our earlier AI copywriting skill) produced fine content. What it didn’t do consistently is the groundwork that makes a page findable in AI search, the discipline now called generative engine optimization (GEO) and answer engine optimization (AEO):
Query fan-out. AI search doesn’t run your query and stop. Systems like Google’s AI Mode, ChatGPT, and Perplexity take one question and fan it out into a batch of related sub-queries behind the scenes, search each one, then assemble an answer from what comes back. Fan-out research maps that same set of sub-queries up front, so a piece answers the full fan the AI is actually running, not just the head term. That’s what gets you pulled into the synthesized answer instead of skipped.
Information gain. Information gain is how much new information a page adds beyond what’s already been said on the topic. It’s a real ranking concept, not a metaphor: Google holds a patent (Contextual estimation of link information gain) that scores a document by the additional information it carries relative to what a reader has already seen, and de-duplicates pages that just repeat each other. Put plainly, it’s the thing you can say that the ten articles already ranking can’t. Usually it comes from pulling in original or third-party data and synthesizing it into something none of the existing pieces bothered to assemble. Say only what everyone else already said and you’re the eleventh redundant result. Add something genuinely new and you become a source worth citing.
Authoritative sourcing, automatically. Every claim traces to a primary or neutral source, screened so a competitor never ends up cited as our authority. This pass produced a learning that turned into a rule.
Never Cite a Competitor
When you send AI to find citations, it doesn’t reliably tell the difference between an authoritative source and a competitor who happens to publish content.
We caught it first on a healthcare client in addiction recovery and saw it again in a legal client’s article about “cheap wills.” The research was pulling in other treatment centers as “sources” for our claims. They rank, they publish, so the model treated them as citable. But citing a competitor as your authority is backwards. You’re handing them credibility and borrowing theirs.
The fix became a hard rule in the skill: source in priority order. Primary sources first (the actual studies, government data, statutes), then neutral authorities (the CDC, NIH, MedlinePlus, the Census), then our own first-party observations. Never a competitor. The skill now enforces the hierarchy instead of trusting the model to sort it out.
What Information Gain Looks Like in Practice
Here’s a concrete one, anonymized. For that same recovery client, the research surfaced a scattered body of clinical literature on which nutrients substance use depletes, why, and what foods restore them. No single competing article had pulled it together. So we synthesized it into one original reference table, exactly the kind of first-party asset Search Influence now builds by default:
| Nutrient | Why it’s depleted | Food sources | What it supports in recovery |
|---|---|---|---|
| Vitamin B1 (Thiamine) | Alcohol blocks absorption and depletes liver stores | Whole grains, legumes, lean pork, sunflower seeds | Neurological function, energy metabolism |
| Vitamin B6 | Alcohol impairs metabolism; opioids disrupt GI absorption | Poultry, fish, potatoes, bananas | Serotonin and dopamine synthesis, mood |
| Magnesium | Alcohol and stimulants increase excretion | Nuts, seeds, dark leafy greens, whole grains | Muscle function, sleep quality, stress response |
| Omega-3 fatty acids | Low dietary intake from substance-related poor diet | Salmon, sardines, walnuts, flaxseed | Brain repair, neuroplasticity |
Every row is backed by a primary or neutral source (Jeynes & Gibson 2017, MedlinePlus, peer-reviewed reviews on PMC). None of it is invented, and none of it is borrowed from a competitor. That table is information gain: the same public research everyone could see, assembled into something more useful than what was out there. It’s what makes a piece worth citing instead of one more entry in the pile.
Try It Yourself: Map the Fan-Out
You don’t need our skill to use the method. The most useful step, and the one most content skips, is mapping the fan-out by hand before you write a word.
Take your topic and sort the real questions people ask into these buckets. Pull them from People Also Ask, the “people also search for” suggestions, and the forum threads that rank. Then look at the pages already winning and mark the buckets they answer poorly or skip. Those gaps are the branches worth owning.
| Category | Questions people actually ask | Do the top results answer it? |
|---|---|---|
| Definition | ||
| Mechanism (how it works) | ||
| Comparison / which is right | ||
| Process | ||
| Cost | ||
| Risks | ||
| Practical how-to | ||
| Timeline |
Citing the definition is common. Very few put cost and timeline together in one clear place. The piece that answers the buckets its competitors skipped is the one an AI engine has a reason to pull from.
The Outcomes
Drafts now come out more thorough and more authoritative, with clearer ties to the entities and primary sources search engines already recognize, because the research is baked in rather than bolted on.
At Search Influence, we had already taken what was a 4-to-5-hour manual process down to about two hours with our earlier ChatGPT based copywriting project. This takes it to roughly 25 to 30 minutes of human time after the machine is done with it. That’s an hour or two saved per piece on top of what we’d already saved, and the quality is better, with cited sources and unique perspectives.
The human doesn’t go away. The human’s job is the part that needs a human: reviewing the draft for the client’s voice, catching their particular sensitivities, making sure the piece represents them professionally. The machine does the heavy lifting. The person does the judgment.
Measure Whether It Worked, or It’s Theater
One part most “AI search optimization” leaves out is the check at the end. Did the page actually get cited? If you can’t answer that, the GEO work is just theater.
Before you publish, write down the five to ten prompts the piece should be cited for, and record which sources get cited today. That’s a good baseline. Re-check at two weeks, then monthly, across Google’s AI Overviews and AI Mode, ChatGPT, and Perplexity. (We run ours through Scrunch to track at scale.)
Citation is non-deterministic. The same prompt returns different sources on different days, so track the trend across several phrasings, not a single pull. All the data is wrong. The only question is how wrong, and whether the trend is moving your way.
The Bigger Bet: SI Studio
This skill is one piece of something larger. I demoed it at SMX Advanced. We’re building SI Studio. The prototype is a slack based system built on Claude and OpenClaw that automates real agency workflows. We’re moving it out of Slack, into a more controlled environment, but the prototype is executing work today.
The idea: instead of a team member logging into ClickUp to figure out what to do that day, they log into SI Studio and find drafts already waiting for the assignments in front of them. A cron job reads their ClickUp tasks, and where a skill exists for that kind of assignment, it runs the skill (SOP) and delivers the output. Nobody pushes a button. The machine pilots it, from “look at this person’s tasks” all the way through to a finished draft for their review.
Three things that solves:
- Smooth start. The assignment shows up as a draft instead of a blank page and a research to-do list.
- Consistency. Every editor has their own way of doing things. When it’s based on a skill (SOP) the output has a consistent starting point no matter who’s at the keyboard.
- Quality of life. People get to spend their attention on the things worthy of a human brain, and less on the mechanical grind.
We’re not trying to replace our writers and other operations teams. We’re giving them a chance to become strategists, and to make the boring, slow, but eminently valuable research and assimilation the thing that gets done for them.
Frequently Asked Questions
What is query fan-out in AI search?
Query fan-out is how AI search engines like Google’s AI Mode, ChatGPT, and Perplexity answer a question. They break it into a batch of related sub-queries, search each one, then assemble a single answer from the results. Content that covers the whole fan, not just the head term, is more likely to be pulled into that answer.
What is information gain in SEO?
Information gain is how much new information a page adds beyond what’s already published on a topic. Google holds a patent on scoring it, rewarding pages that contribute something the reader hasn’t already seen and de-duplicating pages that just repeat each other. In practice, it’s the thing you can say that the articles already ranking can’t.
What is generative engine optimization (GEO)?
Generative engine optimization (GEO) is the practice of structuring content so AI search engines cite it in their generated answers. It overlaps with answer engine optimization (AEO), which focuses on being the extracted answer to a specific question. Both reward authoritative sourcing, clear structure, and information gain over keyword repetition.
How do you get content cited by AI search engines?
Answer the full set of sub-queries behind a topic (query fan-out), add information the ranking pages lack (information gain), cite primary and neutral sources rather than competitors, and structure answers so they can be extracted on their own. Then measure it: track which prompts cite you across AI Overviews, ChatGPT, and Perplexity, and watch the trend.
