What I Learned From Publishing Thousands Of AEO Articles

The pipeline I work on has now pushed somewhere north of a few thousand articles out the door, ours and our customers', in more categories than I can keep in my head at once. That's enough volume to be embarrassed by your early opinions. One of mine was that answer engine optimization was mostly about writing more articles, and the version of me who believed it was cheerfully funding a graveyard, a very organized and beautifully formatted one.

So this is my ledger of what held up and what evaporated, plus what I'd tell that earlier me.

Lesson 1: the plan was always the product

Nobody wants to hear this one, because it's boring. Somewhere after the first few hundred articles, writing quality stopped being what separated the winners, since a competent pipeline gives you competent drafts all day long. The libraries that earned citations and the ones that rotted had been sorted before anyone wrote a word.

It came down to whether the topics started from real buyer questions and whether they were grouped into subjects the site could actually own. Publishing order mattered as well, because the right sequence builds a center instead of a sprawl, and so did interlinking by design rather than by memory.

The plan was the asset all along, and the articles were how you cashed it in. That's why, if you hand me a content plan now, I can predict the library's fate with uncomfortable accuracy. The pre-writing checklist we published is basically that prediction, formalized.

Lesson 2: mentions and citations are different animals

We learned this one by almost getting it wrong. During a 90-day file-on-trial experiment, a blended visibility score drifted up and we came close to reporting a win. Splitting brand mentions from actual citations showed what really happened: mentions were up a touch, and citations were dead flat.

The two move independently because they mean different things. When an engine names you, that's reputation. When it cites your URL, your page won a slot. It's the one measurement habit I'd tattoo somewhere visible, so split them and keep them split, because any tool or report that blends them will eventually tell you a flattering lie.

Lesson 3: every engine is its own country

I went in assuming the engines were dialects of one language. They turned out to be separate countries with their own customs. When we tagged 500 Perplexity citations, the domain overlap with ChatGPT's picks for equivalent queries came to around 11%. The big public audits agree: most cited URLs appear in exactly one engine's citations.

Each country has its habits. Perplexity worships freshness and communities, and ChatGPT, since August, trusts first and searches second, with a soft spot for documentation. Google's AI features run on Google's own machinery, while Gemini likes your own site.

Spread one playbook evenly across that map and you under-serve every country at once. That's why the per-engine columns in our reporting stopped being a nice-to-have and became the reporting.

RANKCONTROL

Know exactly what AI says about your competitors.

RankControl's Recon Agent monitors competitor citations across ChatGPT, Perplexity, Claude, Gemini, Grok, Google AI Mode, and Google AI Overviews. See where they show up and you don't.

See how it works →

Lesson 4: August happens, and August will happen again

I watched this one unfold on live tracking. The week GPT-5.6 reshuffled ChatGPT's sources, community citations collapsed something like 86% in days. Documentation jumped from a rounding error to roughly a third of citations, and every strategy built on the old mix aged a year in a week.

The specifics matter less than they look, though, since the real lesson has nothing to do with that particular reshuffle. Citation share is a policy, and policies change with versions, which leaves the instrument as your only durable advantage. Teams with baselines knew what broke by Friday. Teams without them found out a quarter later, in pipeline numbers nobody could explain.

There will be another August. I can't tell you what it will break, only that the tracking will see it first.

Lesson 5: half your citations won't send anyone

Call it the encyclopedia problem. It stung when I first understood it. A large share of hard-won citations work as fact confirmations inside answers, where the user reads the answer and nobody clicks. One team's audit, which matched what we'd observed, found roughly three quarters of cited pages sending zero visits.

Grading every page on the same metric was the only real mistake. I now think of citations the way I think of impressions, with position doing the multiplying, and I build page portfolios to match. Some pages are encyclopedia pages on purpose: they do the fact-delivery and trust work, and I expect no traffic from them. The others are destination pages that an answer can't finish for you, like calculators, living data, templates and decision-stage depth, and those do the converting.

Interlude: the receipts

Opinions are cheap, and I'm asking you to trust ten of them. So these are the numbers the lessons rest on. Each one comes either from our own tracking and experiments or from the public audits our samples kept agreeing with.

None of these came from a lab. They came from watching real answers every week across real categories, which is itself the meta-receipt for lesson four.

Three articles I regret, as archetypes

Every large library carries its embarrassments. Ours cluster into three shapes worth confessing, and you'll likely recognize them, because I now see the same three shapes everywhere.

There's a version of one article in our network that sprinted onto a hot model-release rumor, ranked for a week, and then aged like milk when the rumor resolved differently. That's the trend-chaser. It taught me the difference between news coverage with a durable angle and news coverage that's just early and loud. We still write news, but now we ask what the piece teaches once the news cools, and if the answer is nothing, it doesn't run.

The eleventh listicle hurt in a quieter way. It was a perfectly optimized roundup on a query where ten roundups already existed, saying what they said and ranking behind them forever. Nobody read it and no engine cited it. Its real damage happened upstream, where it occupied a calendar slot a deeper page deserved, and the same-ten-pages test exists because of articles like that one.

Then there's the clever one. It was built around a tactic that toured the industry for a month and written fast to catch the wave. It was technically accurate and still worthless within a quarter, once the tactic died. The boring fundamentals have never made me write a retraction, and that article's ghost is why lesson eight exists.

What scale actually bought

Lesson ten tells you to start smaller, so you deserve the other side too: what all that volume actually purchased. Mostly it bought pattern visibility. You can't see an 11% overlap or a 73% zero-click share in a twenty-article library, because the sample is too small and every observation is an anecdote. Thousands of articles across many categories turned anecdotes into distributions, and every lesson above came from those distributions.

Scale also stress-tested the guardrails into existence, since every failure mode you can imagine shows up by article five hundred, along with a few you can't. And it bought speed at completing clusters, with whole subjects owned in months. That's the one advantage of volume that compounds cleanly, provided the planning is right.

My synthesis is that scale makes a wonderful instrument and a mediocre strategy. Use volume to learn faster, never as a substitute for coherence. If you have to pick one, pick coherence, because the public audits now publish the distributions that scale used to be the only way to see.

Lesson 6: freshness is a subscription

I used to think of publishing as shipping, where you ship the article and move on. The citation data beat that out of me, especially on Perplexity, where half of citations point at current-year content and a stale date quietly disqualifies you.

Answer engines treat your library like a subscription they're deciding whether to keep paying attention to. Visible maintenance is part of the product, which means real updates when facts change and dates that tell the truth. At scale, the consequence is operational: refresh capacity has to be budgeted like new-article capacity, permanently, or your back catalog becomes an archive the answers stopped visiting.

Lesson 7: the docs kept beating the blog

This one is painful for someone who runs a content engine. Across the August reshuffle and after it, documentation-shaped pages kept out-earning blog-shaped pages for citation slots, sometimes embarrassingly. Factual, structured, unglamorous pages read as trustworthy to verification-flavored retrieval, while the marketing-voiced article next to them reads as persuasion.

The lesson turned out to be letting the docs energy into the blog, rather than blogging less. In practice you open with the answer and verifiable specifics, and you add tables that state facts and FAQ blocks phrased like real questions. The most-cited articles in our whole network are the ones a stranger might mistake for documentation with a personality.

RANKCONTROL

26 content formats. Published on your domain. Matched to your brand.

Guides, comparisons, listicles, case studies, and more. RankControl generates content that gets cited by ChatGPT, Perplexity, Claude, Gemini, Grok, Google AI Mode, and Google AI Overviews.

Start free →

Lesson 8: the boring fixes beat the clever ones, every quarter

Add up the citation gains across everything we've shipped and audited, and the leaderboard is humiliating for cleverness. The winners were boring. Crawlers got unblocked where a CDN had been silently challenging them, and entity facts got reconciled across profiles nobody had read since 2023. First screens got restructured so pages answer before they warm up, and docs went public.

The clever things produced approximately nothing measurable, approximately every time. I mean the schema exotica and the courtesy files, the submit-your-site directories, and the prompt-shaped hacks that tour LinkedIn monthly. There's a whole essay's worth of this in the strategies guide, but the compressed version fits on an index card: access, structure, truth and mentions, measured weekly.

Lesson 9: guardrails made the volume worth having

Publish thousands of anything machine-assisted without guardrails and you get a very scalable embarrassment. The quality checks we run on every article started as hygiene and turned into strategy, because the patterns they block are exactly what readers and engines learned to discount. Which checks matter most depends on the category, and the list only ever gets longer as new failure patterns show up.

The uncomfortable insight is that at scale, your guardrails are your voice. Whatever you don't check for, you'll eventually publish, and the library converges on whatever the checks tolerate. Does that sound reductive? Watch what an unchecked pipeline converges on by article eighty.

The questions I still can't answer

Ten lessons implies more certainty than I own, so I owe you the open ledger too. I can't properly measure influence without clicks yet. When an answer describes a brand correctly to a buyer who never visits, something valuable happened, and the attribution science for it doesn't exist.

Agent browsing is another blind spot, because agents reading sites on a buyer's behalf look human in every log we have, and I can't say yet how that reshapes all this. Whether citations keep concentrating toward fewer, bigger sources, or the engines broaden out again, is open as well.

And I don't know what the next August breaks. That's the most honest sentence in this essay, and it's the entire reason the weekly instrument exists.

Lesson 10: I'd start smaller and deeper

This is the one I'd actually send back in time. Starting today, I'd publish half as many articles into twice the depth. I'd pick fewer subjects and own them completely, starting with definition pages and adding supporting questions in order, and everything would get refreshed forever. Measurement would run from day zero, so every fortnight teaches you something.

The thousands taught me that the compounding came from coherence rather than count. Coherence is available at any size, which is good news for every team that can't publish thousands and shouldn't try.

That's the ledger. If you want the operational version of these lessons, where the pipeline, the guardrails and the per-engine scoreboard run as one system, that's the product we ended up building, largely because we needed it first. The rest of the lessons will find you on schedule. If you're building your own version by hand, take lesson 10 and lesson 2 and go: stay small and deep, split your mentions from your citations, and let the instrument tell you what your market's engines actually reward.

Your competitors are getting cited by AI. You're not.

Every day without citation tracking is a day your competitors pull ahead in ChatGPT, Perplexity, and Claude.

Show me who's getting cited →2-minute overview · real case-study numbers