I went on the Sip On This podcast last week to talk about how content writing actually has to change now that AI Overviews are the first thing most searchers see. Below is the long version of the conversation. What's changed, what hasn't, and the specific moves that are getting our clients cited.

AIO is now the most-cited AI in the world

Google said it themselves. AI Overviews is the most-cited AI surface on the internet. That makes sense when you remember that the majority of Google searches now show an AIO above the blue links. It's the first thing every searcher sees.

If you're not optimising for AIO specifically, you're invisible to the dominant AI surface, not just one of many. ChatGPT, Gemini, Perplexity and Claude all matter, but they sit on top of a search volume that's still dwarfed by Google. Start with the AIO; the rest follow once your content is structured for it.

Get to the point in the first paragraph

The single biggest writing change is how fast you have to answer the question. The first paragraph is the most important real estate on the page, for the AI and for the human.

If the article is about SEO auditing, open with "SEO auditing is…". Say specifically what it is, in the first sentence. The AI hits the page, finds its answer, and cites you. Done.

Open with "Ten years ago in SEO, we used to…" and the AI bounces. It doesn't care about your warm-up. Neither does the human. Both will leave.

Word counts still matter (long-form authority is still real), but the days of 200-word intros before the answer are over. Lead with the answer. Earn the rest of the read.

Chunking: every paragraph stands alone

The structural rule is chunking. People started saying it a year or two ago at conferences. I've been doing it for years and it's why our clients didn't get hammered when AIOs rolled out.

The shape: open with the direct answer, then H2 / paragraph, H2 / paragraph, all the way down. Each paragraph should be able to live on its own inside an LLM's answer, quotable as a standalone chunk, but seamlessly woven into a coherent piece a human can actually read.

That's the test for every paragraph: could the AI lift just this one chunk into its answer without context? If yes, it's chunked properly. If no, rewrite.

Stop saying "we". Say the brand name

This one is underrated. Most agencies don't do it.

When you write "we're really good at HubSpot services", the LLM reads it and asks, "Who's we?" Then it moves on. There's no entity to anchor the claim to. When you write "Multimediax is really good at HubSpot services", the LLM has an entity and an attribute and stores them together. That's what gets cited.

You can't write the whole article in third person. It'll read like a press release. But every few paragraphs, anchor the claim with the brand name. Same applies to your products and methodologies. If you've coined a framework, call it that, by name, every time. We've watched clients name a service explicitly (e.g. "the MMX Growth Marketing Framework") and start showing up for that exact phrase across both Google and the AI surfaces within weeks.

It's a small writing habit with disproportionate citation impact.

Information gain is what gets you cited

AI knows everything. That's the problem.

If your content is a recombination of what's already in the index, you're feeding the AI back what it already has. There's no reason to cite you over the source. That's why so much "GEO content" is failing right now. It reads well, it ticks the boxes, but it doesn't tell the AI anything new.

What works is information gain: an original angle, original data, a fresh framing, a new piece of evidence. Look at what the AI is already saying about your topic, then say something it hasn't seen yet. That's what it pulls.

It's why one well-researched piece can flip an AI's stance on a brand in six weeks, and why fifteen pieces of "AI-optimised content" can do nothing. The lever is the information, not the structure.

Bottom-of-funnel LLM traffic is the new money

The old bottom-of-funnel signal was a short, specific query like "plumber Sydney". You ranked, you got the lead, the lead might or might not be qualified.

The new bottom-of-funnel signal is the opposite: a long, specific conversation with an LLM. The buyer has told the AI everything about themselves before they ever reach you. What they need, what they've tried, what they're worried about. The conversion rate on that traffic is something else entirely.

The challenge is finding those queries. Google won't show them to you. They're privacy-redacted in Search Console because users give up so much about themselves. So you have to work backwards: who's the buyer, what's the full chain of questions they ask before they buy, and where in that chain is the AI not finding you?

This is exactly what we built Negative Space for. We run hundreds of buyer-journey queries across the major LLMs, score where you're present vs. absent vs. losing to a competitor, and surface the specific gaps to fill. The work that used to take a senior strategist four or five days to do manually we now do in hours.

Reverse-targeting: write what you don't do

The flip side of bottom-of-funnel is just as powerful. You can disqualify the wrong leads before they reach you.

High-end plumber example: you only dig trenches and do pipe repairs on heritage houses. The cheap-tap-fix calls are killing your team's time and revenue. So you write the content explicitly: "We don't do small jobs. We don't do new builds. We do work on heritage houses in these suburbs, on these streets, with this kind of pipe."

The AI reads that and uses it. When someone asks "I need a quick tap fix in Bondi", you don't get recommended. When someone asks "who can dig out the cast-iron pipe on a Federation house in Mosman", you do. Lead quality goes up because the AI is doing the qualification for you, before the call.

I haven't seen anyone teach this yet. It's one of the biggest practical levers in the space right now.

Hallucinations are a forever problem

Most foundation models are trained on Common Crawl, a public snapshot of the open web that lags by months. When the model trains, it bakes in whatever Common Crawl had at that moment: your old pricing, your old product names, your competitor's three-year-old feature list, a discontinued SKU.

And brands keep changing. Pricing moves, products get renamed, features ship, features get removed. The gap between reality and what the AI "knows" widens every week.

That's why hallucination monitoring isn't a one-off project. It's a continuous discipline. The brands winning at AI visibility have someone (or some software) checking what the AI says about them, every cycle, against what's actually true, and triggering content + outreach to correct the record.

It also explains why being visible in Common Crawl itself is one of the most underrated GEO tactics. If your latest positioning isn't crawled into the next training cycle, the model defaults back to whatever it learned about you last year.

Technical: speed, crawlability, and the new ops loop

The technical fundamentals haven't changed, they've just gotten more brutal:

  • Let the bots crawl. If you've blocked GPTBot, ClaudeBot, PerplexityBot or any other AI crawler in robots.txt, you've opted out of the citation game. Most do this without realising.
  • Speed. The AI is making sub-second decisions about whether to use your page. Slow pages get skipped. I rebuilt the Earned Media site to static HTML for exactly this reason. There's nothing for the crawler to wade through.
  • Version control. If you're shipping fast (and you should be), you need to roll back fast. Don't deploy directly to prod without a staging step and a one-click revert.

The bigger change is the ops loop itself. The barrier between "SEO recommendation" and "code change" has collapsed. The workflow I run on the Earned Media site now: open Search Console, find a Core Web Vitals issue, screenshot it, drop the screenshot into Claude Code, type "fix this." It fixes it. Push to staging, review, ship.

That doesn't replace strategic SEO judgment. It does collapse the implementation cost of acting on it. For small and mid-market businesses, it's a step-change. For enterprise sites with thirty stakeholders on a deployment, it'll take longer to land. But it's coming.

The same goal, different rules

The framing question I keep coming back to: do you write for humans, for Google, or for AI?

The honest answer is they're not separate goals. A well-structured piece works on all three surfaces simultaneously: direct opening, chunked body, anchored to brand entities, packed with information the AI hasn't seen. The reverse is also true. A piece that fails for the AI is almost always failing the human reader too.

What's changed is the discipline. The waffle is gone. The "build trust slowly" intro is gone. The undifferentiated me-too content is gone. What's left is the work that's always actually mattered: original thinking, specific evidence, real authority, distributed across every surface where a buyer might look.

That's the work an AI SEO Agency should be doing. Same outcome scorecard, both engines running in the same engagement, none of the false-choice "GEO vs SEO" pitches that are sloshing around the industry right now.

Want a read on where you stand? Request a free Truth Gap audit. We run your buyers' deal-breaker questions across the major AI models and show you exactly what's missing.